University of Alberta

Productivity Studies Using Advanced ANN Models

B Y
Ming Lu @
A thesis submitted to the Faculty of Graduate Studies and Research in p h a l fdultillment
of the requirements for the degree of Doctor of Philosophy
in
Construction Engineering and Management
The Department of Civil and Environmental Engineering
University of Alberta
Edmonton, Alberta, Canada
Spring 2001
National Library Bibliathque nationale
du Canada
Acquisitions and Acquisitions et
Bibliographie Services services bibliographiques
395 Wellington Street 395. rue Wellington
Ottawa ON K1A ON4 Ottawa ON KI A ON4
Canada Canada
Your file votre rfrence
Our fi& Notre rdfdrence
The author has granted a non-
exclusive licence allowing the
National Library of Canada to
reproduce, ioan, distribute or sell
copies of this thesis in rnicroform,
paper or electronic formats.
The author retains ownership of the
copyright in this thesis. Neither the
thesis nor substantial extracts fkom it
may be printed or otherwise
reproduced without the author's
permission.
L'auteur a accord une licence non
exclusive permettant a la
Bibliothque nationale du Canada de
reproduire, prter, distribuer ou
vendre des copies de cette thse sous
la forme de microfiche/fih, de
reproduction sur papier ou sur format
lectronique.
L'auteur conserve la proprit du
droit d'auteur qui protge cette thse.
Ni la thse ni des extraits substantiels
de celle-ci ne doivent tre imprims
ou autrement reproduits sans son
autorisation.
Estirnating labor productivity is one of the most difficult aspects of prepming an
estimate, or a control budget based on the estimate for labor-intensive activities in
construction. The primary objective of research is deveioping artificial neural neavork or
ANN based C$i.Kna~g tools CO offer estimators valuable information about labor
productiviv in bidding nem jobs.
In conjunction with a major Canadian industrial contractor, the thesis research
presents case studies on the theoretical basis and practical considerations for me a s k g
and analyzing labor productivity in industrial construction. Two important activities of
process piping mere investigated: pipe installation in the field and spool fabrication in the
fabrication shop. Emerging cornputer modeling techniques such as data warehouses and
ANN were researched fiom an academic perspective and irnplemented in industry to
meet the challenges in productivity studies. The thesis research has addressed: (1) how to
quannfv labor productivity in indusmal construction Erom a contractor's point of view;
(2) how to measure actual labor productivity in industrial construction based upon on-
site control practices; and (3) how to utilize ANN to analyze the vmiability of actual
labor production rates and the sensitivity of i dens ed influencing factors.
Using actual data, the proposed ANN models were proven to be effective in
both risk analysis and sensitivity analysis of construction labor productivity. The
developed data warehouses and - W- bas ed decision-support tools have been
lmplemented or are in the process ofimplementation at the involved Company. The h a 1
results of the research not only assist estmaton to improve the accuracy of e s t i ma ~ g
labor production rates for studied activities in biddlig new jobs, but also offer the
management a precise and integrated view of corporate productiviq information
spanning across many business divisions. The esperience and lessons levned Gom the
successful, productive and rnumally beneficial collaboration benveen academh and
industry in the diesis research wX potentialiy benefit other universiq-indusq joint
research projects in the hture.
This thesis is organized in a paper format, consishg of five main chapters and
five appendices. Every chapter is an independent paper and c m be read separately.
However, 11 the chapters are logically coherent and pertinen~ to the theme of thesis.
Each ;ippendix is a user manual for one computer program that was developed in house
in the thesis research. Chapter 1 o v e ~ e ws the mhole thesis by introducing background
information, problem statements, research objectives, methodologies used, and
contributions achieved. Chapter 2 discusses a case study of industrial construction Iabor
productivty, which depicts the settings of the research. Chapter 3 presents a
probabilistic neural nenvork classification model along with its application in estimating
the production rates of field pipe installation. Chapter 4 presents a sensitivity analysis
method of back propagation neural networks along with its application in estimating the
production rates of shop spool fabrication. Chapter 5 surnmarizes what has been done
thus far and recomrnends what to do in the future research. Appendk A is for the PINN
trainer program based on the mode1 described in Chapter 3. Appendis B is for the
FabMaster program, which is the data marehouse for the fabrication facilities. Appendis
C is for Fab-OLAP, which is an on-line analytical processing program in cornpanion
wth FabMaster. Appendx D is for the PipinghIaster program, whch is the data
warehouse for the field construction systems. Appendks E is for the SensitiveNN
program based on the model as described in Chapter 3.
First and siacerely, 1 would like to thank my universitg adssor, Dr. S. bI.
AbouRizk, mithout whose visions, guidance and encouragement this academic
achievement would not have become a redty.
Especially, 1 would like to ha& my industry advisor, b k U. H. Hermann of
PCL, whose professionalsm and enthusiasm have set the pace and the standard for the
mhole work. 1 am exaemely gratefd to PCL Industriai Constructors, Inc. for s p o n s o ~ g
the research hnancially and allowing me to use its actual data for developing problems
and validaMg solutions diroughout the thesis.
Findy, 1 wodd like to dia& my wife, Duojia, for her understanding, love and
assistance in making this thesis fiom thoughts to finish. 1 dedicate this work to her and
o u arriving Wua n Hum.
Table of Contents
CHAPTER 1 INTRODUCTION ................................................................................................................. 1
BACKGROUNDS .......................................................................................................................................... 1
Industrial Construction ....................................................................................................................... 1
3
Prodrictivity Strtdies .............................................................................................................................. -
Prodrictiviry Modeis ............................................................................................................................. 3
An@cial Neural Nehvorks .................................................................................................................... 5
.............................................................................................................................. PROBLEM STATEMENTS 8
Procirictivity Stridies .............................................................................................................................. S
........................................................................................................................................ A NN Models 13
RESEARCH OBJECI-IVES ............................................................................................................................ 15
............................................................................................................................ Prodrictivity Strtdies 16
............................................................................................. Probnbilistic Neural Network Modeling 16
.............................................................................................. Sensitivity Analysis of Neural Nehvorks 16
M~ ODOLOGES ..................................................................................................................................... 17
Reviewing Literatrtre to Recognize Issues ........................................................................................... 17
Identihing Factors froni Brairlstorrriing by Donrairz Erperts .............................................................. 17
Using Data Warehorlse tu Qrrantitative Data ..................................................................................... 18
Qriestionnaire Srirvey .......................................................................................................................... 19
Cornputer Progrnmniing ..................................................................................................................... 21
.....................................................................................................................
AcADEMIC CONTR[B UTIONS 21
INDUSTRIAL CONTRBUTIONS ................................................................................................................... 22
CONCLUS~ONS .......................................................................................................................................... 23
REFRENCE~ ............................................................................................................................................... 23
CHAPTER 2 A CASE STUDY OF INDUSTRIAL CONSTRUCTION LABOR
PRODUCTIVITY ..................................................................................................................................... 28
INTRODUCTION ......................................................................................................................................... 28
.................................................................................................................... INDUSTRIAL CONSTRUCTION 30
FIELD PIPE INSTALLATION ........................................................................................................................ 32
................................................................................................................. Prodrictivity Qiianrificcntion 33
.................................................................................................................. Producrivity Measrrrerrtenr 34
....................................................................................................................................... Input Factors 36
...............................................................................................
Probabilistic neriral nenvork trrodeling 40
....................................................................................................................... SHOP SPOOL FABRICATION 42
................................................................................................................. Prodrictivity Quantification 43
.................................................................................................................. Prodrtctivity Measurernent 46
lnprit Factors ....................................................................................................................................... 47
......................................................................................... Sensitivity Analysis of Ittflrienci~zg Factors 51
CONCLUSIONS .......................................................................................................................................... 57
REFRENCES ............................................................................................................................................... 58
CHAPTER 3 ESTIMATING LABOR PRODUCTIVITY USING PROBABILITY INFERENCE
NEURAL NETWORK ........................................................................................................................... 6 0
....................................................................................................................................... INTRODUC~ION 6 0
................................................................................................................................. Probleni Dorrlairz 61
.................................................................................................................. Review of NN Applications 64
..................................................................... P R O B A B ~ ~ INFERENCE NEURAL L'WWORK ( P I W) MODEL 64
Introduction of the PINN Mode1 ......................................................................................................... 64
Overview of the PlNN Topology and Process ........................ ... ...................................................... 67
Data Pre-Processing ........................................................................................................................... 70
Orrtpilt Zone Setzip ............................................................................................................................... 74
Processing Elements ( P E) crt Kotionen Layer ................................................................................... 74
.......................................................................................................................... NN Learning Process 75
List of Tables
Table 2-1: Sample of pipe installation unit labor rates ............................................................. 33
Table 2-2: Input factors to pipe installation productioi~ ...................................................... 39
Table 2-3: Sample of degree-of-difficulty factors for converting welds into units ............... 4 4
Table 2-4: Explanatory factors ro spool fabrication productivity ........................................... 50
Table 3-1: Input Factors and Data Type of PINN Mode1 ...................................................... 72
Table 3-2: Inpur Data Sample of PINN Mode1 ......................................................................... 73
Table 3-3: Scaled Input Vector and Initial Weight Vectors ................................................... 78
Table 3-4: Updating Kreight Vectors in Firs t Leaming Stage .................................................. 79
Table 3-5: Updathg Weight Vectors in Second Leaning Stage ............................................... 81
Table 3-6: Trained PINN Ready to Recd for A Given Input Vecmr ............................... .... 85
....................................................................... Table 3-7: Recall Calculations at Bayesian Layer 86
Table 4-1 : Data Set for Testing BPNN and Regression Analysis ......................................... 117
aN1
Table 4-2: Pan5al Derivative (Slope) (-) at Four Input Points ...................................... 119
NP
aN.
Table 4-3: Statisucs of Partial Derivative (Slope) Values: (-) ........................................ 119
~ S P
Table 4-4: Input Factors of Spool Fabrication Labor Productivity .................................. 128
Table 5-1: PINN vs . BP NN ...................................................................................................... 144
Table B-1: Size Range Codes ...................................................................................................... 159
Table B-2: Material Type Codes ................................................................................................. 160
Table B-3: Item Codes for Spool Level Data Compilation ................................................... 160
Table B-4: Sarnple cf Fabhlaster Outputs ................................................................................ 170
................... Table D-1 : Sample of Quantity Calculation Summary Table in PipinghIaster 190
List of Figures
Figure 1-1: Sample Ques ho~ai r e for Findng Facts about Spool Fabrication .................... 20
Figure 2-1: Output interface of PINN recall program ............................................................. 41
Figure 2-2: Sensitivity Analysis of Spool Fabrication BPNN Mode1 .................................... 51
Figwe 2-3: Tes ting Sensitivity of BPNN to Matenal Type ...................................................... 57
Figure 3-1: Topology of PINN Mode1 ........................................................................................ 66
Figure 3-3: Operations at Bayesian Layer in Recd ................................................................... 80
Figure 3-3: Cornparison of PINN and Back Propagation NN .......................................... 88
Figure 3-4: PINN Output for the Base Case Scenario ............................................................. S9
Figure 3- 5: PINN Output for Scenario 1 .............................................................................. 91
Figure 3-6: P W Output for Scenario 3 ................................................................................... 92
Figure 4-1: St ni cme of Back-Propagation NN Mode1 .......................................................... 103
Figure 4-2: Illustration for Node and Laper Representations ................................................ 108
Figure 4-3: Distributions for Input Sensitivity ..................................................................... 120
Figure 4-4: Sensitivity Analysis of Spool Fabrication BPNN Mode1 .................................... 123
Figure 4-5: T e s ~ g Sensitivty of BPNN to Material Type .................................................... 132
Figure A-1 : Select an identifier key of one previous mal ........................ . . . . . . .................. 147
Figure A-2: User selects data table .... .. ....................................................................................... 149
Figue A-3: Flag status of records .............................................................................................. 150
Figure A-4: Setup structure and leaming parameters for PINN .......................................... 152
Figure A-5: SpecZy training iteraons and train- test PINN .................................................. 153
Figure A-6: Check training results ............................................................................................. 151
. .
Figure A-7: Detected noise m training data .............................................................................. 155
Figure A-8: Global report for a train-test t r i al ........................................................................ 156
Figure A-9: PINN Trainer on-line help .................................................................................... 157
...................................................................... Figure B-1: Program Flow Chart of FabMaster 165
........................................................................ Figure B-2: Main User Interface of FabMaster 128
......................................................................................................... Figure C-1: Select one ratio 181
.......................................................... Figure C-2: T d on "number of pipe pieces per foot" 182
........ ....................................................................................... Figure C-3: View details of data .. 183
Figure D-1: S t r u c ~ e s of Raw Data Tables for A Project ..................................................... 185
Figure D-2: Main user interface of FabMaster ........................................................................ 186
Figure D-3: Handikg: S-Reference Information Inte@o/ Check ....................................... 188
Figure D-4: Welding: X-Re ference Information In t egi t y Check ....................................... 189
.................................................. Figure D-5: Productivity Analvsis Page for Pipe HandlLig 191
Figure D-6: Sample of Pipe Handling Ques tiomaire ............................................................. 192
Figure D-7: Program Flow Chart of Pipinghfas ter ................................................................. 193
............................................................... Figure E-1: Splash Screen of SensitiveNN program 195
............................................................................................. Figure E-2: Program Switchboard 197
Figure E-3: Open FFBPNN-mdb First ..................................................................................... 197
Figure E-4: Select data source table ........................................................................................... 198
Figure E-5: Examine details of data and edit record s tatus ................................................... 199
Figure E-6: Program main interface of SensitiveNN ............................................................. 201
Figure E-7: Check leaming results when NN training temiinates ........................................ 202
Figure E-8: Check s Input Sensitivitp for each input-output pair ........................................ 204
Chapter 1: Introduction
Indus trial Construction
Barrie et al (1992) described industrial construction as:
"Indusmal construction covers a wide range of construction projects that
are essential to our ualities and basic industries, such as petroleum r ehenes and
petrochemical plants, synthetic fuel plants, fossil fuel and nuclear power plants,
off s bore oil/gas production facilities, cryogenic plants etc. Indus trid
construction generalIy eatures large amounts of hghly cornples process piping,
mechanical, electrical, and instrumentation mork; bo th design and consuuction
require the highes t level of engineering expertise Gom multiple disciplines."
In particular, the installation of process piping systems in indusuial construction
is selected for productivty studies because it accounts for the bulk of direct labor hours
of an ind~soial contractor. Process piping is used to transport fltids between storage
tanks and processing units. Instdation of piping systems generally consists of nvo
processes: (1) spool fabrication in a commercial pipe shop; (2) pipe installation in the
field. Although the trvo processes are inseparable and can be integrated to optimize the
econornics of a particular situation, they are treated independent of each otl~er in the
thesis because of the cuxent estimaMg and control practices of the involved company.
The productivity studies described in the thesis are conducted to support the
management's decision-rnaliing in the context of the company's m e n t management
systems, as opposrd to radically changing these systems.
Productivity Studies
In a construction task that is performed by hand labor, productivity is commonly
espressed as the labor production rate (man-hours per installed unit), which measures a
key dimension of performance and is a critical factor to estimating, scheduling and
control of the project (hlfeld, 1988). Little information could be found in literature on
the theoretical basis and practical considerations for measuring and analyzing labor
productivity of indusmial construcuon. In conjunction wth a major indusnial concrac tor
(refened to as "the company" hereafter), productivity smdies were conducted for tsvo
Lnportmt activities in industeid construction: pipe installation in t he field and spool
fabrication in the fabrication shop.
In general, productivity studies encompass three tasks: (1) developing special
methods and techniques to quanti@ labor productivity for e s ma ~g, and to measure
actual labor productivty for on-site control, (2) identifjmg input factors that cause the
vmiability in productiviq, and (3) analy zing the relations hips benveen input factors and
productivitg to enhance the accuracy of productivity estimating or Mprove the on-site
performance dkectly. The focus of investigation is the average labor production rates
(man-houn per unit) of these activities 3t the end of a project, rather than the d d y hbor
production rates, because the primary objective of reseaech is developing ANN-based
estimating tools to offer estimators valuabie information about labor productivity in
bidding new jobs, rather than assessing and improving the crem performance in the field.
Produc tivity Models
Several established models for studying productivity cm be found in the
literature, including work s tudy techniques, expectancy model, action-response model,
regression model, expert systems, and arnhcial neural networks (NN).
Work-study techniques were adopted in a nurnber of productivity models, in
which only a fenr factors related to work method were included (Thomas and Dd y ,
1983). Such work-study models cannot be used to model esternai and management
factors. Thomas et al. (1990) and Thomas et ai. (1991) discussed additional drawbacks of
work-study techniques for construction productivity modeling.
The Espectancy model and action-response rnodel are tsvo alternative techniques
proposed to exphin variations in construcuon productivity. In the elrpectancy model, the
effort that an individual is &g to evert accounts for the ciifferences in job
performance or productivity (Maloney and McFillen 1985). The action-response model
graphically depicts the interaction of a number of factors that lead to the loss of
productiviq- (Halligan et. al. 1994). Both models contribute to understanding the
variations in productivity; however, neither can be used to quanufy the influences of
multiple factors on construction productivity (Sommez and RoMngs, 1998).
Sanders and Thomas (1993) developed an additive linear regression model to
study the effect of sis project-related variables on masonry productivity based on data
obtaioed fiom 11 projects. Eight binary variables mere used in the model to represeat
the variations in productivity due to temperature and hurnidity. The effect of crew size
was also taken into account in the model. The results of this regression model suggested
higher productiviq rates for crew with femer members. Thomas and Sakarcan (1994)
connued the reseaxch of Sandea and Thomas (1993) by developing the additive Linear
regression model for the purpose of forecasckg labor productivity. They only included
job condition variables chat describe the work content and the physical components of
the work. The focus of both studies was to determine the coefficient of condition
variables, or the effect of a present condition on the activity productiviry rate based on
the results of historical study; such coefficients were derived independendy of other
inputs vithout accounting for combined effects. In addition, the determined coefficients
are constants based upon the average values of historical data, and do not reflect the real
situations in wiiich the values of such coefficients may vary with the specific job
conditions.
Esqxrt systems is another technique applied co model labor productivity in trvo
studies found in literature. Hendnckson e t al. (1 987) developed an hvo-stage espert
system named ''MASON" to estimate acuvity durations for ma s o q construction. First,
the maximum espected productivity was estirnated. Nest, this rate was adjusted for
various characteristics of job or site. The masimum productivity estimates and the
followhg adjustments mere based on the knowledge obtained Erom interviews wi t h a
professional mason and a supportkg laborer. Christian and Hachy (1995) developed an
esTert system to estimate the production rates for concrete pouring. The expert system
relied on the knowledge extracted Erom experts and data coLlected rom seven
construction sites. The user simply queried the expert system for an estimate through a
ques tion-and-ansmer routine. In both e-upert sys tems, productiviq mas es timated
through previously dehned decision d e s obtained from domain experts. Because the
nature of forrnulating rules is subjective, the resultant rules may be inconsistent.
h o t h e r disadvantage of analyzing productivirp based on expert systems is expert
systems do not perforrn Functional input-output mapping, i.e. quantitative evaluation of
the impact of job condiuons on productivity.
In the follow-..ng subsection background information about r \ NN models d be
introduced and the technique of modeling productirity using IWN d be discussed.
Artificial Neural Networks
htficial Neural Networks (ANN) research involves multiple disciplines
including biology, xtificial inteiligence, cornputer science, and mathematics and evolves
with the developments in each related discipline. Kohonen (1995) dehned MJN as "
massively pardel htercomected network of simple (usudy adaptive) elements and their
hierarchical organizations, intended to interact with the objects of the real world in the
same way as the biological nervous systems do." Sirnply put, an ANN mode1 is an
analytical mode1 that sirnulates the cognitive learning process of the huma. brain, and is
automaticdy constructed feom leaming esamples or data by ttid and error Gthout
heuiristic design or other human intervention.
ANN deds effkctively mith ill-structured problems, in which the algorithms
required to solve them CaMOt be given in a precise and explicit fashion, or the data for a
partcular problem are either not complete or cannot be s pe de d preasely (Widman et.
al., 1989). ANN has been found to be capable of perfomilig pardel computations on
different tasks, such as pattern recognition, 1inea.r optimization, speech recognition, and
predicuon (Mukhe jee and Deshpande 1995). In short, the s p e d leaming algonthms of
ANN are capable of performing . g h dimensional, non-lineu input-output mapping and
extracMg hidden patterns and predictive information from observing the leuning
esamples.
In recent years, ANN has been rescarched and applied as a convenient dedsion-
support tool in a rarieq of application areas in civil engineering, including modulv
consmiction decision making &Iurt;iza and Fisher, 1993), s t n i c ~ a l analysis (Flood and
Katirn, 1994), e s t i ma ~g construction productivity (Portas and AbouRizk, 1997), mode
choice analysis of beight mansport market (Sayed and Razavi, 1999), construction
m&p estnating (LI et al 1999), measurkg organizauonal effecveness (Sinha and
hfcI<im, 2000), and predicbng setdement durkg tunneling (Shi, 2000).
In o u research, ANN was selected as the main methodology and utilized to
analyze the varGbility of actual labor production rates and the sensitivicy of identified
influencing factors due co two reasons.
Fkst, construction labor productivity is influenced by a variety of factors. bIodel
fit-ting based on construction labor productivity data requires quantification of the effects
of factors on labor productivity and quanacation of the interactions among the factors.
The task of hdi ng a mapping hinction h m the independent variables to the dependent
variable is analogous to that performed by some of the neural netrvork models such as
back propagation (Sonmez and Rowngs, 1998). In statistics, regression analysis is the
most common method to explore this relationship; in particular, the objectives and
operations of nonlinear regression analysis are comparable to back propagation neural
nenvorks. Homever, regression models requke the user to define a priori the parametric
expression for the model (lnear, quadratic, etc.). In the case of modeling productivity,
the user is mainly concemed wth what the productivity dl be for any given set of work
conditions, and may not necessarily be interested in the parametrc expression of the
model, for instance, a highly complex nonlinear hnctional equation. On the other hand,
ANN is capable of nonlinear mapping for most complicated problems such as modeiing
productivity; the modeler does not need to esert much effort to decide on the class of
relationships in a precise and e,uplicit fashion.
Secondly, one of the amactive properties of ANN is th& capacity for tolerating
moderate amouncs of noise in the data. In many real applications, the quantity and
quality of the avdable data for modehg labor productivity rnay not support the fitting
of a regression model. In such cases, ANN may be applied to generalize the knowledge
from incomplete or noisy data and provide good solutions the problem.
Moselhi et al. (1991) pointed out the possible use of ANN for construction labor
pr oducvi ~ modeling. Portas and AbouRizk (1997) developed an M N model to
estimate construction productivity for concrete f or mor k msk. The rnajority of data
used in the study mas collected by questiomaires on a project basis. The prediction of
the ANN model was compared wth that of senior estbators for a single project.
Sonmez and RoMngs (1998) developed ANN models for quantitative evaluation of the
impact of multiple factors on productivity in concrete pouring, fomwodc, and concrete
anishing tasks, using data compiled fiom eight building projects. Their study also
compared regression models induding the pure Linev regression rnodel, the regression
models rvith interaction and nonlinear tenns Gth ANN models, and concluded, " the
use of neural networks helped the over d modeling process. Neural networlcs have
shown potential for quantitative evaluation of the effects of multiple factors on
productivity, especially when interactions and nonlineu relations mere present-"
The problems to be solved in the thesis research were identified dirough
i nvest i ga~g the m e n t eshmating and control practices of the involved Company and
reviewing the established -! I NN models and applications as found in the litetanire. Placed
into civo different perspectives, i.e. productivity studies and ANN models, the dehed
problems can be stated as follows:
Productivity Studies
EstunaMg labor production rates for field pipe instdation commences Mth
establishing base production rates for various work items. Base production rates reflect
the contractor's present labor productivity level under normal rvorlc conditions that are
most ofien encountered in the field. The installation location is one of the major
considerations for an estimator to d e h e a classification of work conditions. For
example, the base produccion rates of pipe installation are valid for the conespondlig
base classification only, in which the installation location is above ground up to 12 fi
hgh. fm estimator determines a degree-of-difficulty factor (often referred to as
c'multiplier" in the company) for each non-base classif5cation to adjust the base rates up
or down in order to reflect the unfavorable or favorable work conditions for the job
bWig estimated. This is a subjective detision process, requiring substantial esperience
and skill on the part of the estimator to determine realistic production rates for the work
conditions to be encountered. Empirical degree-of-difficul~ factors for each
classification of work conditions based on the installation location serve as a guide or
tool to assist in decidng on such factors and can be found in the company's business
manual. For example, the degree-of-difficulty factor for underground pipe installation (3
to 10 ft deep) is about two times the factor for aboveground pipe installation (up to 12 f i
high), while the factor for pipe installation inside building at over 10 fi of height is about
~ v o Mies the factor for underground pipe installation (4 to 10 ft deep).
Historical piping productivity data of 66 projects mas coliected Erom the
company and compiled into numerc format for andysis. The folIowing hvo
observations mith regard to the actual degree-of-difficulty factors can be made Erom the
histoncd dam of the company: (1) the degree-of-ciifficulty factor for one classification of
installation location may reveal a widespread distribution instead of a constant value as in
the company's business manual; and (2) different ciassifications of installation location
may end up with veq- close values of the degree-of-difficulty factor, not as distinguished
as in the cornpany's business manual.
The above obsemations are not iniaally e-xpected and the e-uplanation cm be
attributed to the fact that more factors esist, other than the location of installation,
which contribute to the variability in labor productivity. In practice, an estnator rnay
adjust the value of degree-of-difficulty factor iri the business manual on a job based on
es~erience and specific job conditions, and subjected to the approval of senior
management. Barrie et al (1992) found that construction hbor productivity may fluctuate
d d l y due to nurnerous factors that affect it, and many are highly qualitative in name,
including the effect of location and regiond v ~ u o n s , the learning c u v e , work schedule
and work d e s , environmental effects, crew eqerience and management factors.
Identification of input factors in the study of field pipe installation productivity was
mainly based on Knowles (1 997). A total of 36 input factors are considered relevant and
used to r edehe the classi6cation of pipe installation. Those factors include bath global
project-level information and specific activity-level information.
To estimate a fabrication project, a speual "unitization" scheme is applied to
quanti$ the various work items uniformly into an abstract unit of measure cded
"fabrication unit" or "unit" by weighting them for their degree of difficulty. A degree-of-
difficulty Factor is empirically determined for each weld, taking into account pipe
diameter, mall thickness of pipe, weld type put t weld, socket weld, saddle and lateral
welds) and the time required to lay out and perfom the weld. Quantity of non-welding
work items such as cutting, beveling, handling pipe and fittings, i ns t ahg supports are
also converted into "units" by appIying conesponding degree-of-difficuly factors in the
scheme-
Once the total "units" for a project are detemilied, the focus of productiviq
study in spool fabrication is on the production rate directly (man-hour/unit). Sirnilar to
deciding on the degree-of-dffiultg factor for a classification of work conditions in field
pipe installation, deuding on unit tabor rate for spool fabrication requires the esperience
and judgment of the estnator. The environmental effects and management factors ae
noc considered as sgnficant factors, as in the field productivty studies, because of the
controlled shop environment, consistent policy and management personnel d h g the
period of investigation. h totd of 29 input factors are identified as affecting labor
productivi~ of shop spool fabrication based on consultation -5th expenenced estimators
and shop superintendents in the Company.
It is not straightfomasd to create a conventional andytical model so as to
accommodate the impacts of numerous factors on t he mget xsky variable - degree-ol-
difficdty factor or production rate. It takes yevs of site ex~erience and eshat i ng
practice for an estimator to develop his/her own mental model. The decision process
relies heavily on individuas experences and the results are often inconsistent r e f l e c ~g
the experience and disposition of the estirnator.
ANN has becn proposed by many as an alternative to streamllie the e s t ma ~g
process and reduce the subjective nature of the work.
ANN Models
The classic Back Propagation NN predicts a single value without gi~ing any
bacbvp infornation on the risks of taking ths value as correct. Observing the acmal
values for the degree-of-difficulty factors of field pipe instaliation indicates that the
target nsby variable lies over a relatively wide range. The result from an informal end-
user s w e y showed that estimators are more cornfortable to accept a desion support
model \.th the capability of analyzing the uncenainty of its outpur. Thus, a probabilistic
NN modeling approach that c m predict a distribution or probability densitg h c u o n
over the output range is preferred ar,d has been researched.
Portas & AbouRizk (1997) proposed a feed fonvard baclc propagation neural
network model for estimating construction production rates of formwork. The nenvork
outputs a single point prediction dong wi& a number of output zones, with equal
Iikelihood of the production rate being in any one zone. The output zones are
symrnemc and divided evenly across the range of likely production rate values. D k g
training, the output zone with the output that coincides with the actual production rate is
remarded with a prirnary score of 1.0, representing strong certaliv. A certain degree of
fuzziness is considered by remarding the 2 adjacent output zones with secondary scores
of 0.5, representing weak certainty. iV1 the other output zones are assigned a score of O.
Once the hW is trained and inputs are entered, the NN d predict a point value as weU
as the likelihood of production rates being within the output zones. This model achieved
lmited success due to the fact that the adopted back-propagation NN mode1 is long on
non-linear regression, but short on &ssiication.
Specht (1 99 1) revisited Probabilis tic Neural Network (PNN) and General
Regression Neural Nenvork (0 algorithms wth the objective of i nt e gr a ~g
statistics and neural training. GRNN/PNN is a memory-based feed fomrard neural
network mode4 mhere the training is performed in one pass, thus requing less training
&ne. GRNN/PNN is able to identify a posterior distribution over the NN meiglit
vectors and a point-value prediction is generated based on the predicted distdbution.
However, based on experhentations and observations, GRNN/PNN is not quite
tolerant of noisy data (inaccurate or incomplete records) and imposes a demandlig
standard of data quality that is hard to achieve in reality. The memory demand and
cornpuMg t h e for GRNN/PNN increase very rapidly when the dimension of input
vector and the quantity O E training samples increase.
Kohonen proposed tmo speal NN models, namely Sel f-Orgki ng Map (SOM)
in the late 1980s and Leaming Vector Quantization (LVQ) in the rniddle 1990s. SOM
performs unsupervised classification and clusteng to represent high-dimensional,
nonlinearly related da ta items in an illustra tive, O ften nvo-dirnensiond dis play. LVQ
combines unsupervised and supervised leaning and is recornmended for statistical
pattern recognition problems. i n LVQ, "deusion surfaces, relating to those of the
Bayesian classinec, are defined by neares t-neighbor classification mith respect to sets of
codebook vectors assigned o each class and describing it" @<ohonen, 1995). It is noted
that the predicted result O E LVQ and SOM is detenninis tic, being classiEied into one of n
predehned clusters or classes.
IGiowles and AbouRizk (1997) presented a tmo-stage NN model in p r e d i c ~g
pipe-installation labor productitity. The input factors are used to invoke a LVQ
classification process, follomed by a predictive one. With the classification, the mode1
predicts whether the output is likely in a cypical or non-typical range. The proper feed-
fomard back-propagation network is then esecuted. The drawback of this method is
that a build-up of errors occurs when the classification fails. For instance, if the
classification accuraq is 90% at the hrst stage of NN, and the prediction accuracy at the
second stage of NN is 85'10, the prediction accuracy of the whole 1Wis only 76.5%
(90% &es 85Yo)-
In c ona s t mith a rather wde distribution of the actual production rate in field
pipe installation, the actual labor production rates for shop spool fabrication are
bounded within a relatively narrom range. Thus, the NN modeling of hbor productivity
in the shop puts more emphasis on the sensitivity analysis of influenchg factors based
upon the classic back propagation N-N model, as opposed to the uncertaincy analysis of
cspected production rate.
However, leamuig algorithms such as BPNN do not attempt to infer causality,
hence, classification or prediction is based on b h d coirelation of new examples nrith
previously analyzed examples, mithout giving information on the effect of each input
parameter or influencing variable upon the predicted output variable. In the reported
NN applicatioions, model validation has thus far relied upon measuring accuracy of the
calibrated netrvork to an independent testing data set that are hidden fiom the neural
nenvork in learning. The modelfs sensitivity to changes in its parameters is generally
probed by t e s ~ g the response of a manire neturork on various input scenaros. In short,
a NN model hc t i ons like a "black boxf' package, giving no clue on how the answers or
model outputs are obtained, or how the input parameters affect the output.
Widman et. al (1989) pointed out that the credibility of an AI program
Gequently depends on its abiliq to explain its condusions. Lack of interpretability is a
pi t f d of the neural netmork models recognized by many and has inhibited NN fiom
achieving its fidl potential in real-morld applications. Dhar and Stein (1997) argued that
because NN algorithms such as the back-propagation NN are non-linear, high
dimensional hinctional equations f e a h g paralle1 distebuted data processing, it is liard
to esplicidy hterpret mhich parameters cause what behavior in the NN model. YVhile
mathematical and operational methods do esist for the analysis of neural nenvorks, die
methods are fairly in~olved, and are less than satisfyuig because of their theoretical
assumptions. They stated that "unlike most statistical methods, it can be difficult to say,
even in general, mhich variables are significant in what respect." (Dhar and Stein 1997)
The ultimate goal of the thesis research is to hnd better neural nework modeling
approaches to predicting production rates and productivity indices. When applied in
industrial construction estimahng as decision-support tools, the dereloped ANN-based
models for analyzing productivity should be acceptable and effective to offer estimators
valuable infomiaon about labor productivity in bidding nem jobs. To amin this
ultimate goal, the follonrulg objectives are de he d in regard to three aspects:
Productivity Studies
Investigate the cuirent estimating and ou-site control practices for industrial
construction as applied to the involved Company, in order to advance the theoretical
basis and practical considerations for measuring and analpzing labor productivity in
indusrrial construction.
Probabilistic Neural Network Modeling
Building upon the previous developments achieved b y O thers, es tablis h a more
effective NN approach that suits the needs of estimating indusmal construction projects,
which requires the recreation of a new training acd r ecd algonthm that combines the
hctionality of probabilistic classification and prediction in one integrated neural
netsvork.
Sensitivity Analysis of Neural Nerworks
Define the input sensitivicg. of a NN mode1 in mathematical terms, and establish
a method of i nt er pr e~g the relevance and impact of NN &put parameters on the
predicted output variable so as to gain insighr into the rationale by which NN reason and
make decisions.
Main
folloming.
methodologies udized to fulhll the abore research objectives include the
Reviewing Literature to Recognize the Issues
A compreheasive literature review was conducted in regard to the established
AI N models, productivity studies, ANN applications in the problem domain,
optimization, statistics, and industtial consmction. Literature covers a wide range of
joumals, books, and reports, which document the latest academic developments and
industrial applications in the related areas. Licerature review helps recognize the issues to
be addressed in the thesis, namely, how to get data kom ndusuy in modeling labor
productivity, how to analyze the uncertainq of the output from an ANN-based
productivity model, and how analyze the sensitivity of the input for an ANN-based
productivity model.
Identifying Factors fiom Brainstorming by Domain Experts
The senior management and domain expens at the nvolved Company including
superintendents, production engineers, consmiction engineers, drafking supe~t endent s,
quality control superintendents, and melding foremen were convened for a
bralistomiing exercise to identifV the factors that influence productivity of the studied
activities. It should be noted that those factors as identified to influence labor
productivitg holds only midiLi a specifc setting and over a specific period. The input
factors may need adjusmient by adding relevant ones and deletlig inelevant ones when
the setting of application changes to a different contractor, or a different penod, even if
the consauction process being studied remains the same.
Using Data Warehouses to Gather Quantitative Data
Idenufgrng relevant factors and gatherng )ueh quality data for those factors are
crucial to the success of modehg labor productivity using XNN. Fouonring
identification of factors, data needs to be coilected.
The collaborative company (PCL Industrial) provided us with access to its
business data for validation of ANN models and development of ANN-based decision-
support tools. Although the company has invested resources in management information
systems at various business divisions, those systems mere developed and implemented
separately. Productivity studies using ANN require vast amounts of data fiom different
information management systems. A corporate data warehouse is "a process by which
related data korn many operational systems is merged to a single, integrated business
information view that spans many business divisions" Fang, 1997). With the support of
the company's management, tsvo data marehouses, namely, PipuigbIaster and FabMaster,
were custom-developed for field pipe installation and shop spool fabrication respectively
to integrate the corporate management sys tcms of es timating, production resources
planning, quality control, and labor cost control. Validating and processing of
quantitative data were automated through cornputer programming within data
marehouses. The developed data warehouses provide solid platform of integrated
historical data from which to validate the ANN models and develop ANN-based tools
for productivity analysis.
If data for some factors is not recorded in the elasting management systems,
questionnaire surveys were carefully designed and personnel at the cornpnny were
i n t e ~ e we d to collect the needed data.
Questionnaire Survey
With the heIp of domain experts, questions and descriptive idormaiion of
choices for r at i ne on a 6ve-point scale mere f o d a t e d into a questionnaire format
with the objective of reducing ambiguities and confusions. It is worth mentioning that
such questionnaire smrey for modehng productivity using ANN is intended to fnd facts
of the past projects only. BasicaUy, in conducting the quest i o~ai r e survey, no persond
judgment or opinion about the relationships between the facts and the results is
involved. Questions of "What" type mere asked about the factors affecting productivity
only, and no questions of "Why" and "How" types were asked about the relationships
between the factors and the productivity. In k t , i\NN would sort out the relationslups
between the facts and the results on its own through an iterative leaming process based
on e x p l o ~ g sample data. Intelligence emerges mhen ANN hnds the input-output
patterns or relationships hidden Ln the data. This feanire draws a distinct line betsveen
A m approach and other intelligent modeling approach such as expert systems: ANN
relies on facts and data, but requires less direct input fiom domain experts (Dhar and
Stein, 1997). In short, modeling productivity using ANN is relatively an objective
approach compared to expert systems. Figure 1-1 shows the questioanaire designed for
fnding additional facts about spool fabrication.
PCL INDUSTRIAL CONSTRUCTORS INC:
Fabrication Facility Productivity Questionnaire
General:
Reported By
Report Date
r
Bob Smith
FabMaster
Processed Flag
Project #
Project Name
Schedule:
1700204
Gas Plant & Piperack Process Modu
How busy was the shop? (Based on shop workload in ternis of units and concurrent jobs processed)
- -- --
-
1
Very Slow Relatively Slow 9 Normal Relatively Busy Very Busy
I
Were there rnany rushed spools?
.-
-
None a Relatively Few 12 Normal Relatively Many Li Many (30% plus)
(5% less) (1 0%) (20%)
.----
Engineering:
What was the rework percentage due to drawing changes?
7
None fl Relatively Few Normal Relatively Many Many (30% plus) 1
(5% less) (1 0%) (20%)
1
Were there any late drawing issues?
g Relatively Few i-~ Normal Relatively Many 7 Many (30% plus)
(5% less) (1 0%) (20%)
i
I
What was the drawing revision rate?
Materials:
Were there rnany material shortage problems that impacted production?
1
None Relatively Few Ci Normal Relatively Many Many
1
Figure 1-1: Sample Questionnaire for Finding Facts about Spool Fabrication
Foilowhg the formulation of a questionnaire, superintendents, project managers
and estimators who were involved in the past projects were interviemed to compile facts
and gather the needed information. The interview process was straightfonvard; the
domain e-xperts had no difficulty hnding the records or recalling the facts on the projecrs
that they maaaged.
Cornputer Programming
Mcrosofi Visual Basic, Visual Basic for Application, and Access svere found to
be flesible and powerful in handling large amounts of data and comples programming
logic, hence, were selected as cornputer programmhg tools to develop both the data
warehouses and ANN models in the thesis research. hli the programs in di s thesis
research including data ~varehouses, ANN trainers, and ANN recall programs were
developed in house without third-party sohvare and hac-e been utilized in the involved
Company.
The solutions to the identified problems, which are provided through the thesis
research, d contribute to the general knowledge of productivity srudies and ANN
modeling in regard to:
Advancing the theoretical basis and practical considerations for measuring and
analyzing labor productivity in industriai construction, which has been documented
in a paper entitled "A case study of industrial construction Iabor productivity" and
has been submitted for publication in the Journal of Consmiction Engineering and
Management, ASCE;
Devising a new neural netrvork scheme to meet the requirements in modeling Iabor
productiviq of indusnial construction, and is termed the Probability Inference
Neural Network (PNN). P m T is a dassihcation-prediction combined neural
nemork mode1 based on I<ohonenYs LVQ concept (Kohonen, 1995), but integrated
\.th a probabilistic approach, which has been documented in a paper entitled
" Es t k n a ~ g labor productiviv using probability inference neural nenvork" and is
published in the October/2000 issue of the Joumal of Cornputhg in C i d
Engi ne e ~g, Vol 14(3), pp 341 -338, ASCE;
EstabEshing a simulation-based method of nterpreting the relevance and impact of
back propagation NN input parameters on the predicted output variable so as to
gain insight into the rationale by which back propagation NN reason and make
decisions, mhich has been documented in a paper entided "Sensitivity anaiysis of
neural nenvorks in spool fabrication productivity studies" and has been submitted
for publication in the Journal of Computing in Civil Engi nee~g, ASCE.
The developed data warehouses and ANN-based decision-support tools have
been irnplemented or ate in the process of implementation at the involved Company.
The hnal results of the research not only assist estimators in irnproving the accuracy of
e s t i ma ~ g labor production rates for studied activities in biddlig new jobs, but dso offer
the management a precke and integrated view of corporate producavity infomiation
spanning across rnany business divisions. The experience and Iessons learned fiom the
successful productive and m u d y beneficial collaboration betsveen academia and
industry in the thesis research d potentially benefit other university-industry joint
research projects in the future.
The problems addressed in the thesis research were idenufled through
i nvest i ga~g the curent estimating practices in i ndus q and understanding the real
concems of industry prokssionals. Emerging compter-modeling techniques such as
data warehouses and ANN were researched Gom an academic perspective in order to
meet with the challenges in industry. The proposed novel ANN models and developed
decision suppoa tools were proven to be effective in both uncertainty mdysis and
sensitivty arialysis of construction labor productiviq; they were validated using real data
Gom industry and successfully applied to assist esmators in deciding on labor
production rates for new jobs.
Alfeld, L. E. (1 9 88). Constndon prodi~~~ivity - on -site mea~~zfnmetzt and nzmzcrgenzeizt.,
McGraw-Hill, New York, NY.
Bmi e, D. S., and Paulson, Ir., B. C. (1 9%). Pmj~siona/ com-hzccion management,
incl~ding CJW., deng~i-conrtnict, aodgenerd contrachg. 3" ed, McGram- W, New York, NY.
HaLligan, D- W., Demsets, L- A., Brown, J. D., and Pace, C. B. (1994). "Action-
response models and loss of productivity in construction." joz~rnuf of Com-tr/~~-tion
Engizeeniig andManagement, ASC E, 1 20(1), 47-64.
Li, Y., Shen, L. Y., and Love, P.E.D. (1999). "MN-based mark-up estimation
s ys tem with self-e-uplanatory capabilities ." Joz~n~cd Comtni~.tion Engieenig md
&lanagcme)zt, ASCE, 125(3), 185-189.
Ma l o q , W. F-, and McFiUen, J. (1985). "Valence of and satisfaction with job
outcomes ." ]ozmai of Corn-tntction Etzgineerirrg and Management, ASCE, 1 1 1 (1) , 5 3-73.
Mukherjee, A., and Deshpande, J.M. (1995), 'SvIodeling initial design process
using d c i d neural networks", Joumaf Coqzhzg i11 Ci d Engineering, ASCE, 9 (3), 1 9 4
200.
Murtaza, MB., and Fisher, D.J. (1993), 'Tu-euromex: Neural Network System for
Modular Cons tmction D ecision Making'', Jorrmal Compti~zg ~ I I Ch7 Enginecnk& ASCE,
8(2), 221 -333.
Sander, S. R., and Thomas, H. R. (1993). "blasonry productivity forecasting
model." Jounmi of Constnicr'on Engineering ami Ma~zagemelr, AS CE, 1 1 9 (1 ) , 1 63- 1 79.
Saped, T., and Razmi, A. (1999). "Cornparison of neural and conventional
approaches to mode choice analysis" Jozmzai of Compziing itz Civil Engi~teebng ASCE, 14(1),
23-30.
Shi, J. (2000). "Redung prediction enor by trans forming input data for neural
networks", ]ozcnzuL Coqzhng in Cid Engineenhg, ASCE, 1 4(2), 109-2 15.
Sinha, S. K. and Md(im, RA. (2000). c c 4 Mc i a l neural netmork for measung
organiza tio na1 e ffec tivenes s .", Jotcri~nI Compting in Ci d Engzheenhg, AS CE, 1 4(l), 9 - 1 4.
Sonmez, R and Ronings, J. E. (1998). "Consmiction labor productivity
rno deling mith neural netmorks ." JozmaI of Con~tmctioiz Engineering mzd ~tlanngeme~zt, AS C E,
l23(6), 498-504.
Thomas, H. R. and Sakarcan, AS. (1994). "For ecas ~g labor productivity using
factor model." JotimaLof Con~~n~ciio~i Eirgineenhg and~u~z age ~e / ~t , ASCE, 120(1), 228-239.
Thomas, H. R. (1991). "Labor productiviy and work sampling: the bottom line."
JozmiaI of Co~zsfmction Engince* and Marzugement, ASCE, 1 17 (3), 423-444.
Thomas, M. R., bfaloney, M.F., Horner, R.M., Smith, G.R., Handa, V.K., and
Sanders, S.R. (1990). "Modehg construction labor productivity." ] o z d of Consindoil
Engineering und lUziiagemerz~, ASCE, 2 16 (4), 703-725.
Thomas, H. R., and Daily, J. (1983). "Crew performance measurement via
activity sampling." J o j m d of Constmction E ~zgineeritg min Mmzagen~e~zi, AS CE, 1 09 (3), 3 09 -
320.
Wang, C., B. (1997). Techno Viaoiz II, McGraw-Hill, New York, N.Y.
Widman, L-E., and Loparo, KA (1989). " Ar t i Ed intelligence, Simulation, and
modeling: a critical s wey", Artff;n'aL inteIhgence, dation, and modebrzg- L.E.Widman, K.A.
Loparo, and N.R. Nielsen, eds., John Wiley & Sons Ltd, New York, NY, 1-45.
Chapter 2: A Case Study of Industrial
Construction Labor ~roductivity'
In a construction task that is performed by hand labor, the labor production rate
(man-hours per uistalled unit) measures a key dimension of performance and is a cntical
factor to estimating, scheduling and control of the project (Alfeld, 1985).
Thomas et d (1999) identi6ed the
management as nvo factors that affect
complexity of the design and the project
labor productivty and invesgated the
measurements of daily labor productivity in building conspuction including masonr)-
cons tniction, concrete fomwork construction, md s t r u c ~ a l s tee1 erection. Thep found
that good project management and consis tency
constant d d y labor production rates.
in design complexity result in relatively
1 A version of this chapter has been submitted for publication. ASCE, Journd of Consuuction
Engineering and Management.
A good conelauon was also found benveen the final curmhtive production rate
(an index of the average iabor performance over the entire project penod) and the
variance of daily production rates. For instance, theit study of ma s o q construction
observed "high vuiability in daily production rates on the poor perfomiing projects due
to disruptions in the work resulting from congestion, sequencing, lack of materiais, etc"
(Thomas et al, 1999).
Little information could be found in literature on the theoretical basis and
practical considerations for measwng and analyzing labor productivity of industriai
consrmction. In conjunction with a major industrial contractor (refened to as "the
company" hereafter), we conducted a productivity case study for nvo important activities
in indusnial consmiction: pipe installation in the field and spool fabrication in the
fabrication shop. The focus of investigation is the average labor production rates (man-
hours per unit) of these acbvities at the end of a project, rather than the daily labor
production rates as in Thomas 1999, because die prirnary objective of research is
developing ANN-based eStiIIIa~g tools to offer estimators valuable information about
labor productivity in bidding new jobs rather than assessing and improving the crew
performance in the field- This paper intends to address: (1) hom to quana$ labor
productivity in indusmal construction fiom a contractor's point of view; (2) hom to
measure actual labor productivity in indusmal construction based upon on-site control
practices; and (3) how to ualize Artifcial Neural Nenvorks (MN) to analyze the
variability of actual labor production rates and the sensitivity of identXed influenclig
actors.
The paper is
construction pertinent
Constructionyy section.
organized as folloms: important characteristics of industrial
to productivity studies are hrst discussed in the "Indusmal
Next, the "Field Pipe Xnstallatioa" section reviews the cur ent
e s t i r na ~g method, the present reportng and accounthg systems for field pipe
insrdation in the Company, and summarkes the techniques for quantification and
measurement of field pipe installation productivity. Further, the input factors that cause
the variability in the productiviv of field pipe installation are discussed, and a
probabilistic neural network approach to modeling pipe uistdation productivity is
o v e ~e we d . The subsequent section "Shop Spool Fabrication" shifts the focus of
productivity studies to the fabrication faalities of the Company, and summarizes the
techniques for quantification and measurement of spool fabrication productivty. The
input factors that affect the production rate of spool fabrication are identified, and an
NN-based sensitivity andysis appraach to modeling spool fabrication productivity is
presented.
Barrie et al (1992) described industrial construction as:
"Indusmal construction covers a wide range of construction projects di at
are essential to o u utilities and basic industries, such as petroleum rehneries and
petrochernical plants, synthetic fuel plants, fossil fuel and nuclear power plants,
off shore oil/gas production fadties, cryogenic plants etc. Industeal
construction generdy features large amounts of highly complex process piping,
mechanical, electrcal, and instrumentation work, both design and construction
require the highes t Ievel of engineering e&xpemse kom multiple disciplines."
In particular, the installation of process piping systems in indusEal construction
is selected for productivq studies because it accounts for the buLk of direct labor hours
of an industrial contractor. Process piping is used to transport fluds benveen storage
tanks and processing units. Installation of piping systenis generally consists of tnro
processes: (1) spool fabrication in a c omme r d pipe shop; (2) pipe installation in the
field (Germin, 1996). hlthough the nvo processes are Liseparable and can be integrated
to optimize the econornics of a partcular situation, they are treated independent of each
other in the paper because of the current estmating and control practices of the
involved Company. The productiviq studies described in this paper are conducted to
support the management's decision-making in the contest of the company's curent
management sys tems, as opposed to radically changing these sys tems.
Parker et al (1984) di s~gui shed industrial construction from heavy constniction
in that indusmal construction does not require fleets of construction equipment and
plant (such as scrapers, loaders, cranes and trucks etc) to handle basic materials (such as
e d , rock, concrete and asphdt etc). They M e r pointed out that industrial
constniction "tends to be much more labor-intensive, though some of the largesr
hoisting and materials-handling equipment is also required" (Parker et al, 1984). An
industrial contractor usudy owns the equipment or rents it fiom a long-terrn supplier,
thus, the technology and machinery adopted in consuuction can be considered invariable
for a relatively long period of time. This feature lends the producnvity studies of
indusmal construction to the unit-cost estmating method, mhich is cornmonly applicabIe
to labor intensive work where 'labor production rates must be independent of
equipment use and vary among projects only because of differences in labor
productivitf (Parker et. al., 1984). For instance, considering the bid item "Pour
Concrete Floor" in building constniction, to estimate the total cost in terms of Iabor
hours, work quantities are taken off in square meters of floor, then multiplied by a labor
production rate, i.e., the labor hours required on one square meter of floor. Analogously,
for field pipe installation in indusmd construction, the amount of work-in-place is
usually counted in pipe footage; field productivity for pipe installation is measured in the
form of unit rate, i.e. manhours per foot of installed pipe.
Pipe installation in the field involves "the physical placement of pipe / pipe
subassemblies, valves, and other specialty items in their required final location relative to
pumps, heat exchangers, turbines, boilers, and other processing units" (Genvin, 1996)
Productivity Quantification
In practice, pipe is customarily identified by diameter of pipe (dehned by
nominal pipe ske) dong mith \vaU thickness of pipe (dehed by schedule nurnber).
Hence, the production rate of pipe installation can be detemiined by the diameter and
mall thickness of pipe; the Iarger the diameter and the thicker the pipe, the more Iabor
hours is required 10 install one foot of pipe. Table 2- 1 shows samples of the labor rates
for handling and erecting stright run pipe (mm-hours/ft) as found in the public source
(Page and Nation, 1982).
Table 2-1: Sample of pipe installation unit labor rates (Source: Page
and Nation, 1982)
Nominal Pipe Size
(Diameter)
Schedde
Number (Wall
Thickness)
Base Labor Rate
(MWFt)
Estimating labor production rates for field pipe instdlztion starts with
establishing base production rates for various work items. Base production rates reflect
the contractor's present labor productivicy level under normal work conditions that are
most oftea encountered in the field. The installation location is one of the major
considerations for an estimator to d e h e a classification of work conditions. For
esample, the base production rates of pipe installation are valid for the conesponding
base classihcation only, in which the installation location is above ground up to 12 ft
high. An estimator detemiines a degree-of-difficulty factor (&en referred to as
c'multiplier" in the company) for each non-base classification to adjust the base rates up
or down in order to reflect the unfavorable or favorable work conditions for the job
being estnated. This is a subjective decision process, requiring substantial expenence
and skill on the part of the estimator to determine realistic production rates for the work
conditions to be encountered. Empical degt-ee-of-difhculty factors for each
classficaon of work conditions based on ihe installation location serve as a guide or
tool to assist in deuding on such factors and can be found in the company's business
manual. For esample, the degree-of-difficultp factor for underground pipe installation (4
to 10 ft deep) is about two t he s the factor for aboveground pipe installation (up to 12 ft
high), wMe the factor for pipe installation inside building at over 10 fi of height is about
two &es the factor for underground pipe installation (4 to 10 ft deep).
Productivity Measurement
In the contest of pipe installation, keeping uack of piping labor by individual
fittings and pipe sections is economically impractical, if not impossible, to implement in
the curent field reporting system of the company. Alfeld (1 9 88) argued that measuring
labor producvty requkes grouping similar accomplishrnents and separating dissimilm
accomplishment on the job site. The cost control practice of the company for field pipe
installation is descnbed next.
At the end of a day, the foremen tum in time cards for th& crews, charging the
number of labor hours to a series of cost codes. The cost codes of field pipe installation
for a particdzu project separate pipe fitters' hours by classihcations of installation
location. Thus, the total labor houts of pipe installation at vmious locations for one
project can be readily remieved fi-om the field labor cost control system of the company.
Homever, this is not the case for the amount of work accomplished. Large amounts of
various work items dong -sith variations in size and wall thickness of pipe cause the
inclusion of details of work accomplished in the foreman's t h e cards to be impractical,
34
such as the amount of work-in-place counted in footage by diameter and mail thickaess
of pipe, the s aew joint or bolt-up connections and the valves and supports installations
associated mith the installed pipe. Fortunatelp, the detailed records about the amount of
work accomplished can be obtained indirectly h-orn the company's quality control system
and estimating system. Thus, we can match the actual manhours mith the work
accomplished for one classifxation of installation location, in order to compute the
actuai degree-O f-dificuls. factor (@) as gken in Equation (1):
Where H is the actual labor hours charged to pipe installation in one
class~cation of installation location,
N stands for the total number of work items contained in one classification of
installation location,
P, is the base labor rate for the iCh work item,
And Qi is the actual quantity accomplished for the ch work item
Note that the estrnabng process desuibed in the preceding subsection is actualiy
to transform Equation (1) to compute the labor hours 0, simply by plugglig the
quantity take-off as read from construction &anrings into the quantity terni (QJ in (1).
Hence, the task of estimakg labor productivity boils down to detennining the degree-
of-difficulty factor (@) accutatelv for a future project scenario. It is e-xpected that a
constant value of the degree-of-difficul~ factor (or at least a nanom range) could be
found for each classi&ation From the company's bistoncal records and shodd be close
to the empirical value in the business manual.
Input Factors
Kistorical piping producticity data of 66 projects was collected Erom the
Company and compiled into numeric format for malysis. Because data is not well
formatted or readily accessible, a data marehouse was developed first to integrate the
conuactor's e s t i r na ~g system, quality control system and hbor-cost control system in
order to ease the burden of data collection and ensure the high quality of collected data.
The follonring tmo observations wth regard to the actud degree-of-difficulty
factors can be made from the histoecal data of the Company:
The degree-of-difficulty factor for one classification of installation location rnay
reveal a widespread dismbution instead of a constant value as in the company's
business manud;
Different classifications of installation location may end up with very close values of
the degree-of-difficulq factor, nor: as distinguished as in the company's business
manual.
The above observations are not initidly expected and the explmation can be
attributed to the fact that more factors esist, other than the location of instailauon,
mhich contribute to the variability in labor productivity. In practice, an estmator rnay
adjust the value of degree-of-difficulty fzctor in the business manual on a job based on
experience and job conditions, and subjected to the approval of senior management.
Barrie et al (1992) found that construction hbor productivicg may fluctuate d d l y due to
numerous factors that affect it, and many are highly qualitative in nature, induding the
effect of location and regional variations, the learning cuve, work schedule and work
d e s , environmental effects, crem esperience and management factors. Portas and
AbouRizk (1997) determined seven categories of activity factors and five categones of
project performance factors to be relevant to the labor production rate of concrete
formwork constxuction. Thomas et al (1999) identihed the complesiq of the design and
the project management as nvo major categories that affect labor productivity of
masonry construction. In regard to industrial construction, Knomles (1997) invescipted
a specmim of e-xplanatory factors to idenufy those that affect the productivity of pipe
installation and pipe welding in the field.
Identification of input factors in this study was based on Knowles 1997, with die
addition of t h e more factors, i.e. the contract type (lump sum or reimbursable),
installation of miscellaneous fittings (flanges, specals, elbows etc.), and the on-site labor
charging errors between the cost code of pipe installation and that of pipe welding (since
pipe fitters and welders mostly work side by side). A t o d of 36 input factors are
considered relevant and used to redehe the classification of pipe installation. Those
factors include both global project-level information and specific acctivity-level
information, as shown in Table 2- 2. Aside hom location of instdation, more activity-
s p e d c factors are considered such as material type of pipe, the installation of non-pipe
cornponents (valves, supports, and rnisceUaneous items), non-weld joints in ins tailation
(screw joints, bolt-ups), the quantities of installed pipe at different size ranges ( s md
bore, medium bore, and large bore), the leaming cuve factor (total quantity of installed
pipe in footage), the crew elrperience etc. Factors pertinent to project are also included,
such as the effect of location and regional variations (project location, province/state),
project type variations @roject dehnition, contract type, and prefabrication percentage),
mark schedule and mork rules (overtime and unionized), environmental effects
(seasonal), management factors (superintendent and project manager) etc.
Table 2-2: Input factors to pipe installation productivity
Projea Location
Administration
Y- of Construction
Province/State
Contract Type
Client
Engineering Fimi
Project Manager
Superintendent
Project Definition
\Vork Scope
Project Type
Prefab/Field Work
,iverage Crew Size
Peak Crew Size
Uninized
Equipment & Material
Estra Work
Change Order
Drawing & Specs Qualiqr
Location Classification
Total Quantity (Learnuig)
Installation Quanti ties
&Taterial Type
hIethod Of Installation
Pipe Supports
Boltups
Valves
Screwed Joints
hfisc. Components
Welding Impact
Season
Crew Ability
Site Working Conditions
Inspection, Safety & Quality
Overd Degree of Difficulty
Urban, Rural, Camp Job
General E-xpense
89-93,93-93,95-96,97-99
M, SI<
Reimbersable, Lump Sum
an indes derived &oui historical data
an indes derived Gom his torical data
an index derived fi-om historicd data
an indes derived from histoicicd data
Chernical, Cryogenic, Gas, Refining
Confiaed / Scattered
Upgrade Shutdown, Grass Root etc.
Percentages for Prefabrication
<25,25-50,50-100, >IO0
<25,25-50,50-100,100-150, >150
Yes, No
Equip.& Mat1 Cost/ Direct MH
Original Project Cost/Final Projeject Cost
No. of Change Orders/Total Direct blH)
1 Poor 3 Average 5 Escelient
U/ G on Site, Fab Shop, A/G on Site etc-
Total Quanuty In DiaInFt
Qty for Size Ranges, <2", 2"-IG", >TG"
Moy,Carbon Steel, FRP/PVC,etc.
Percentages of Hand Rigging
No. of Pipe Supports/Foot of Pipe
No. of Boltups/Foot of Pipe
No. of Valves/Foot of Pipe
No. of Screwed Joints/Foot of Pipe
Instail i1lsc.Components hfH/Foot of Pipe
LVelding Multiplier (hliscoding on Site)
Percentages of Winter & Summer Work
1 Very Low, 3 Average 5 Vesy High
1 Esmeme Problems - 5 No Problem
1 Estremely Detailed - 5 Highiy Tolerant
1 Very Lom 3 Average 5 Very High
It should be mentioned that a questionnake survey was careWy designed and
conducted to collect some qualitative information that is not obtainable Erom the
company's reporting and accounting systems. Such information mas converted into
numenc fonriats for the foUoMng NN analysis (See Lu et al, 2000 for derails).
Probabilistic Neural Network Modehg
ANN has been proposed by many as an alternative to streamline the eshmating
process and reduce the subjective nature of the work The dassic Back Propagation NN
predicts a single value without giving any backup information on the rislcs of taking this
value as correct. Observing the actual values for the degree-of-difhculty factors of field
pipe installation indicates chat the target r i sky variable Lies over a relatively nride range.
The resulr from an informal end-user survey showed that eshat or s are more
cornfortable to accept a decision support model with the capability of analyzing the
uncertainty of its ourput. Thus, a probabilistic NN modehg approach that can predict a
distribution or probability density hnction over the output range is preferred and has
been researched.
A nem neural network scheme was devised to meet the requirements in modehg
labor productivity of industrial constniction, and is termed the Probnbility Inference
Neural Network (PINN). PINN is a chssihcation-prediction combined neural nenvork
model based on I<ohonenYs LVQ concept (Kohonen, 1995), but integrated with a
probabilistic approach. Because the response of PINN is in the form of a probability
density funciion (distribution) at the output range, an estimator be able to decide on
the degree-of-difficulty factor for a future scenario by cornbining the PINN7s
recommendation with personal judgment.
In the PlNN model, the actual output range of the target nsky variable is divided
into a number of output zones or sub-ranges wth an equal width. Output zones are
actuaUy some discrete dusters wth c o n ~ u o u s boundaries. For field pipe installation,
the hrgher the value of the degree-of-difficulty factor, the higher the value of labor
production rate, hence, the more difficult and more demanding the job is. Thus, each
output zone gives an indication of the relative wotk cbfficulty and productivity levei; for
instance, output zone (0-0.71 stands for easier mork and higher productivity level
NN Recall Probability Density Graph
1.0
compzuing with output zone (0.7-1-41. The median of each sub-range can be used to
represent the typical value for each output zone and to derive a predicted vahe in
addition to the predicted dismbution, such as mode and meighted average value.
Portable computer software was developed to implement the training and testing
of the PINN model on real historical productivity data of field pipe installation at the
company. The model was validated based on an independent data set resemed for
testing. Sensitivq malysis of the model was perfonned bjr obserping the PINN's output
in response to controlled changes in inputs and comparing PINN's output a@st that
of a n es~erenced estimator. FoUowing satisfactory t e s ~ g and sensitivity analysis, a
r ecd program based on the traned PINN model was irnplemented as a deasion support
tool for estirnating the degree-of-difficulty factors of held pipe installation at the
company. Figure 2- 1 shows the output interface of the recd program, indicating the
predicted probabilit)' density function over the output range, and the likelihood of the
degree-of-difficulty factor f d h g into each sub-range. Those mho are interested in the
topology and algorithm of the PINN model, and the effectiveness of applying PINN to
e s h a t e labor productivity in the contest of field pipe installation may refer to Lu et al
2000.
Spool fabrication in a commercial pipe shop involves "the cutting, bending,
tacking, and welding of individual pipe components to each other and their subsequent
heat treatment and nondestructive esamination to fonn a pipe subassembly or spool for
installation" (Gervin, 1996). A pipe spool is a portion of piping system consisting of
various piping components, such as h g e s , elboms, reducers, tees, supports, and pipe.
These components are prefabricated into distinct assemblies that are later assembled as
part of an industeal plant or production skid/module. Such prefabrication is usudy
performed under coatrolled shop envixonment located away from the job site, which
allows for bettes productivity and quality control, and heace cuts the field labor costs.
Major spool fabrication processes, such as cut, bevel, fit, weld, and handle
sections of pipe and firtings, also tends to be labor-intensive. Productiiig- data is
coIlected fiom the fabrication shop of the Company for 63 projects completed fiom
1995 to 1999, duriag which period the technologies and machines for welding and
cutchg in the shop eemain relatively stable. Like field pipe installation discussed
previously, the productbity study of spool fabrication is suitable to the unit-cost
estnating method-
Productivity Quantification
Alfeld (1988) pointed out the labor production rate in the shop could not be
quantified with the same units as in the field - man-hours per foot of Listded pipe,
because the shop does not install the pipe but cuts, fits and welds spools; other units of
rneasure such as spool counts and pipe sections do not satisfy the needs of quantifjing
the work accomplished in the shop either, because (1) each spool varies so much in
components, size and configuration that a simple count ofspools would be misleading;
and (2) large-size pipe requixes far more manhours to cut and weld than do the smaLl-size
pipe. Weld-inch was ulized as a unit of rneasure to quannfy the accomplishrnent in a
fabrication shop and Table 2- 3 shows saniples of the degree-of-difhcultp factors for
convemng various butt welds into weld inches as found in Mfeld, 1988.
Table 2-3: Sample of degree-of-dititiculty factors for converthg welds
into uni t s (Source: Alfeld, 1988)
Nominal Pipe
Size
(Diame ter)
Circumference Fab. Units Weld
Type
(3)
Butt
Butt
Butt
Butt
Weighting
Factors
Similar to the concept in hlfeld 1985, in the fabrication shop of the Company, a
special "unitization" scheme is applied to quanrifv the various work items uni f ody into
an abstract unit of measure c de d "Fabrication Unit" or "Unit" by we i g h ~ g them for
their degree of difficulty. The "unitkation" is a conversion based on a standard diameter
inch dong the circurnference of a weld. A degree-of-difficulty factor is empirically
decemned for each weld, t z h g into account pipe diameter, wall thickness of pipe, weld
type @utt weld, socket weld, saddle and laterd welds) and the time required to lay out
and perforrn the weld. Quantity of non-melding work items such as cutting, bevelulg,
handling pipe and fitnngs, lastalling supports are also converted into "Units" by appljing
corresponding degree-of-difficulty factors in the scheme.
A commercial fabrication shop usually handes several jobs simultaneously so
that it is efficient for the crew to set up and do ail the sarne size pipe fiom dfferent jobs
at the same tirne. In the fabrication shop of the Company, it is difficult enough keeping
nack of the manhours charged to each individ~d job in the shop floor control systems.
Charging labor hours to each individual pipe section or fitng is considered impractical
and ineffiaent in hght of the curent control technologies and management systems in
the fabrication shop.
The basic formula for spool fabrication eshating is shown in Equation (2):
VVhere H is the total manhours charged to one job,
P is the production rate (?vIf-I/Unit) for the job,
N stands for the total number of mork items (\veld or non-weld) contained in the
Subscript i stands for the P work item in the job,
@; is the degree-of-difficulty factor for the ih work item in the job,
Qi is the quanticg for the ih work item in the job in its onginal unit of measure
such as the meld counts for an weld work item, e.g. the weId count for "G Nominal Pipe
Size (Diameter), 40 Schedule Number F a l l Thickness), Butt-\Veld Type" weld is 20.
Prodiictivity Measurement
g(@i * ~ i )
in Equation (2) is actually the total quantity of
i=l
fabrication mork in Units for the job. The hrst step in esthnating a spool fabrication job
is a process c de d "uni&ation" for computing the total units of one job. The es t-cor
reads the quantity takeoff fiom spool drawulgs and l ook up the degree-of-difficulty
factor for each work item. This task is straightfomard but tedious because the amount
of work items in a job is usudy large; for esample, several jobs the Company completed
contain over 1,000 spools, over 10,000 welds and over 10,000 pipe sections and fittings.
The difference in the degree-of-difficultg factor benveen field and shop should
be noted:
First, the degree-of-difficuiq~. factor in ihe case of shop spool fabrication
corresponds to each work item rather than a classification of grouped work items as
in field pipe installation.
Second, the degree-of-difficulty factors in the case of shop spool fabrication are held
constant in the "unitization" scheme rather than variables as in field pipe installation.
Hence, the focus of productivity study in spool fabrication is on the production
rate directly, i-e. the P term in (2) or man-hour/unit. Dedding on P requites the
expexience and judgment of the estimator. Similar to the productivity study of field pipe
installation, a data warehouse was builr up to integrate the reporting and a c c o u n ~ g
systems in the fabrication shop in order to obtain labor hours and quantity of fabricauon
mork on each job. The data warehouse also contains a built-in computer progam,
developed to automate the tedious task of quantifping about 63 fabrication jobs into
"units" in a prese and consistent may. Actual production rates over the penod of
investigation mere observed for W e r analysis.
Input Factors
Xfter consdting with e-xperienced estimators and shop superintendents in the
Company, a number of quantitative and qualitative factors are considered relevant to the
shop labor productivity, such as:
The ma t e d components in fabrication, Le. the percentage of non-carbon steel
(stainless, aluminurn, d o y steel etc.) units over the total units, because non-carbon
steel spools require extra care and more tirne in storage, handling and welding
cornparhg Mth carbon steel spools;
The average length of pipe sections in a spool, indicated by in-Line fittings (pieces)
per foot of pipe in spool. I n- Le fitangs, such as unions, couplings, swages, reducer
etc are used to connect pipe sections in a saaight line without tums or branches.
The complexity of spool configuration, indicated by non in-line fittngs (pieces) per
foot of pipe in a spool, val ves/ suppor t s/ ~ges (pieces) per foot of pipe in a spool;
The stengency of quality control, indicated by the non-destructive test requirement,
which is a percentage mith respect to weld couats according to the client's specs.
The quality of spool drawing indicated by the drawing revision rate.
The shop workload, indicating shop's state of being busy or slow, and number of
concurrent jobs handled at one time;
The effect of double handliog spools between weld stations, iciicated by the
percentage of multi-station roll meld inches over total roll mell inches. A meld may be
done on more than one station by different welders in the shop, depending on the
welding process and the welder's quaL6cation. It requires estra time to move spools
benveen melding stations and lay out a weld nt different stations.
The effects of rushed spools due to client's priority, late drawing issues fiom the
client, and material supply problems
The amounts of night s hift and overtime, and estra mork in tems of labor hours;
The es~erience and profiuency of crem, indicated by apprentice ratio, repair rate and
rem-orked spools.
The environmental effects are not considered as signihcant factors, as in the field
productivity studies, because of the conaolled shop enwonment. A couple of
management factors that mere initially included n-ere dropped out of analysis after
esarnining the collected data, in mhich slighr variations were observed due to the
consistent management policy and management personnel dueing the 5-year period of
investigation. It should also be mentioned that another factor describing the complexiq
of spool configuration was idenafied by domain er;perts, i.e. the number of pipe pieces
per foot of pipe in spool. The sensitivity analysis results based on coilected data reveal
that the effect of the number of pipe pieces is very similar to that of the number of in-
h e fitnngs. Such strong conehtion generalized by ANN model from the actual data is
presented to dumain experts and hnds esplanations from domain experts: pipe sections
in a spool are mostly c o ~ e c t e d by in-iine fittings such as unions, couplkgs, mages,
reducer etc; both ratios, narnelp, in-line fittings (pieces) per foot and pipe pieces per foot,
indicate the average length of pipe sections in a spool. To simpliQ the inputs of model,
the ratio of pipe pieces per foot mas dropped out of analysis, as agreed by domah
exTerts. Eventually, nineteen input factors that affect labor productivity of shop spool
fabrication are identified as listed in Table 2- 4.
Table 2-4: Explanatory factors to spool fabrication productivity
NN Input Factor
(2)
:n Line Fitting @CS) per Foot of
?ipe in Spool
\Ton In Line Fitting @CS) per Foo
>EPipe in Spool
Jdve @CS) per Foot of Pipe in
;pool
Support @CS) pet Foot of Pipe in
;pool
?lange @CS) per Foot of Pipe in
;pool
liIulti-Station Roll CVeld Inches /
rotal Roll Weld Inches
Lepair Rate
Ldbgraphy Test Requirement
'lon CS Units / Total Units
;hop Work Load
Drawng Revision Rate
Prionty Rushed Spools
Rework Spools
Material Shortage Problems
Late Drasving Issues
Night Shih MHs / Total MHs
Over Time hMls / Total hEIs
Extra Work hHs / Total MHs
Apprenticeship bMs / Total h Hs
Remarks
(3)
.i ratio iadicating the average length of pipe sections
n spool
.\ ratio indicating complexiv of spool confguation
i ratio indicating comple'ty of spool configuration
i ratio hdicating cornplexity of spool configuration
i ratio i n d i c a ~ g compleldty of spool configuration
Multi-S tation Roll Weld requires extra handling
Demeen weld stations
in indes of crew's pro ficiency
ln indes of quality control strngency by specs.
\Jan CS component in fabrication requires estra care
n storage, handling and wvelding
i 5-point rathg based on shop worliload in units
ind no. of concurrent jobs indicating honr busy the
;hop was.
A 5-point rating based on percent of revised spool
drawings indicating dranring quality
A 5-point rating based on percent of mshed spool
due to client priority indicating shop work schedules.
&
A 5-point raMg based on percent of reworked
spools due to drawing enor s and quality defects
L
A 5-point ratng on efficiency of material supply
A 5-point raMg based on percent of late spool
draning issuance by client that impacts fabrication
Night Shift affects labor productivity
Or er T h e affects labor productivity
Estra Work affects labor productivity
Welder a u ~ c a t i o n system affects labor
I
productivity: Apprentice os. Joumeyman
Data for the idenuhed factors is collected fiom the company's various
management s ys tems including labor cos t tracking svs tem, wveld tracking sys tem, payroU
system, matenal tracking system. Because data is unavailable in cunent systems of the
Company for such factors as the material shortage problems, quantity of reworked
spools, quantity of rushed spools due to pnoriq, shop workload etc., a questionnaire
survey was csrefdly designed and conducted wth the suppott of the compmy
management The key personnel involved in the projects including shop
supe~t endent s, project managers and coordinators, QC staff, and welding foremen
were interviewed to help recall sorne facts and gather the needed information.
Sensitivity Analysis of Influencing Factors
In contrast with a rather wide distribution of the actual production rate in field
pipe installation, the actuai labor production rates for shop spool fabrication are
bounded Nithin a relatively narrow range. Thus, the NN modehg of labor productivity
in the shop puts more emphasis on the sensitivity anaiysis of i n f l u e n ~ g factors based
upon the classic back propagation NN mode4 as opposed to the uncertainty analysis of
espected production rate based on the PZNN model.
L e h g algorithms such as back-propagation NN do not gve information on
the effect of each input parameter or influencing variable upon the predicted output
variable. The NN model's sensitivity to changes in its input factor is generdy probed by
t e s ~ g the response of a mature nenvork on various input scenaros. The relationships
between an output variable and an input parameter were sorted out based on the NN
algonthm so as to define the input sensitivity of a back-propagation NN model in esact
mathematical ternis in light of both normalized data and raw data (Lu et al, 2000). For a
three-layer BPNN using Siaomoid transfer hcr i ons and linear normdkation procedures,
' N n
R
the input sensitivity with respect to the change of 10/o input relevant ranges (- ) is
as,
R
aNn -
- -
MAX, -MIN,
- 2 ~ W - Nc t ( l - Nc l ) *N, ( l - N, )
3% 10
FI =ln
i =I
Where, subscript p stands for a node in die input lager of the network;
Subscript c stands for a node in the middle layer of the network; C stands for the
total nurnber of nodes in the middle layer;
Subscript n stands for a node in the output layer of the nenvork.;
Wii stands for the weight of connection benveen node i and node j;
S stands for the input signal to a node;
N stands for the output signal fkom a node;
hW& is the maximum value in the data set corresponding to output node n;
MIN, is the minimum value in the data set corresponding to output node n.
From Equation (3), for a mature network, the sensitivity of an input parameter
over an output variable is dependent on the curent input values. A Monte Carlo
simulation can be performed at the NN input space in order to observe the statistics of
input sensitivity. In our research, s taus tical analysis of simulation resulrs involves
calculating 5 percentiles of the slope variable for each input parameter, i.e. the loch, 25",
52
50", 75", and 9 0 ~ . The input sensitivitg of all input parameten is summarized and
presented in a tornado-like graph as illustrated in Figure 2- 2 for the piping fabrication
labor productivity NN model. The horizontal a'ris represents the relative input sensiwty
as detennined by (3), i.e. output response (negative or positive) nrith a change of 10/o
relevant range in an input parameter. The vertical avis is the baseiine conespondkg to
no output response or zero change in output. Five short vertical bars correspond to each
input parameter, representing respective. the five percendes fiom left to right in an
ascending order and reflecting the central trend, the spread, and the shape of the
observed slope data distribution fiom simulation. In short, statisticd analysis of input
sensitivity based on Monte C d o simulation enables the modeler to understand the
rationale of NNYs reasoning and have pre-knomledge about the effecveness of model
implementation in a probabilistic fashion, as illustrated by the spool fabrication
productivity model next.
A total number of 70 records mere compiled and used to train a NN model with
19 input nodes at the input layer correspondlig to 19 input parameters, 19 hidden nodes
ac the middle layer, and 1 output node at the output layer that is the unit labor hours.
The number of hidden nodes can be determined based on mals; NN learning is found to
be wusceptible mhen to the number of hidden nodes is close to the number of input
nodes. The learning rate is 0.4, the momentan is 0.1, and sigmoid transfer hrnctions are
used in hidden and output nodes. After satisfactory training (standard enor of the output
is 0.00143), the Monte Carlo based sensitivity analysis is performed on the rnamed
nenvork for 10000 simulation runs.
I In Line Fitting per Ft
!, Non In Line Fitting per Ft
i. Vaive per Ft
C . Support per Ft
i- Flange per Ft
i. Mlt Stn RW %
'. Repair Rate
I. RT Rate
). Non CS % (Un)
IO. How Busy
1, Drawing Revision
12. Priority Rushed Spools
3. Reworked SpooIs
14. Material Problems
5. Drawings Late
6. % night shift
'7. % overtime
8. % extra
9. % apprentices
Figure 2-2: Sensitivity Analysis of Spool Fabrication BPNN Mode1
Several independent Mals from NN training to the sensitivity andysis mere
conducted on the sarne data set. The best mal, in which the input sensitivity of most
hctors followed the same trends, as detennined by esperienced domain e'lperts, is
shovm in Figure 2- 2. An esamination of Figure 2- 2 reveals the relationships berneen
the influencing factors and the fabrication productivity, which me generalized by NN
through obserrring historical project data in the past 5 years. For example, factor 1 is
about in line fitang pieces per foot of pipe in spool, st-hicb indicates the average length
of pipe sections in spool. According to our domain experts, in h e fittings, such as
unions, couplings, smages, reducer etc are used to connect pipe sections in a straight line
without nirns or branches. Thus, the more in line fitting pieces in spools, the more s md
sections of pipe in spools, and the easier to handle the work. From Figure 2- 2, BPhW
determines the chances to decrease hbor hours per unit wirh the increase of this ratio are
about 78% and agrees with the &end identified by domain experts. Factors 2 to 5 are
four ratios i n d i c a ~ g the complesity of spool configuration. By our domain esTerts, the
higher such ratios, the more comples the spools' configuration, and the tougher to
hbncate the spools. From Figure 2- 2, the dominant trends of the four ratios are all on
the plus side, which matches the judgment of our domain experts. It is also observed
from Figure 2- 2 that factor 18 (extra work percentage) is relatively tighdy enveloped
around the baseline, which indicates that extra work is not as dominant as other factors
in c ont r i bu~g to the variance in unit Iabor rates. The e-xplanation c m be partly
atmbuted to the fact that the amount of extra work impacts the efficiency of
administration or management more directly chan the productiviq of crew on the shop
floor. Other input factors can be interpreted and validated in a similv muiner, and are
not elaborated further due to space limit.
In particular, the effect of ma t e d type of spool fabrication on the labor
productivity was tested based on the BPNN model, because ma t e d type (carbon steel,
stainless steel, aluminum etc.) is a major consideration of an lndustnal estimator in
adjuscing unit labor hours of spool fabrication. The labor rate of non-carbon steel
fabrication is empiticaUy 1.5 Urnes the rate of carbon steel in the company's business
guideline. 24 records in the data set wth OO/o non-catbon steel component (100% carbon
steel fabrication) were selected as t e s kg records. Nest, for each t es hg record, the input
parameter of non-carbon steel component mas changed fiom 0% to 100%, with odier
parameters intact. Those testing records were fed to the netsvork and let NN r ecd the
output, i.e. the unit labor rates for non-cubon steel fabrication. ni e output fkom NN
was compared agauist the onginal output of each record, i.e. the unit kbor rate for
carbon steel fabrication. Based on the test results in Figure 2-3, NN increases the unit
labor hours on 75% of the records; the amount of decrease for 5 records, i.e. No. 1, 2, 5,
6, 9, is rehtively smalI c o mp a ~ g mith the amount of increase for others. If the sample
size is large enough, the percentage should corne close to about 90%, as observed fkom
Figure 2- 2 for factor 9. On average, the ratio of non-carbon steel Iiibor rate over carbon
steel labor rate is 1.4, which is dose to 1.5 as in the guideline.
Test NN Sensitivity By Changing Material Component from 100% CS to 100% Non-CS: 75%
1
Records increase, Avg. Ratio 1.38
1 +Actual (100% CS, 0% Non-CS) 0 NN Output (0% CS, 100% Non-CS) )
O 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25
Rec. No.
Figure 2-3: Testing Sensitivity of BPNN to Material Type
that the guideline gives
only, while NN is able to
an average nurnber (1.5) in consideration of
figure different numbers for different scenarios
taking into account 19 relevant factors. In short, such a NN-based decision support tool
d be more sophisticated and intelligent than the traditional business guideline to assist
estirnators in deding
CONCLUSIONS
on the labor production rate of spool fabrication.
Special methods are utilized in practice for the quantihcation and measurement
of labor productivity in i ndust d construction. Estirnating labor productivity is one of
the most difficult aspects of preparhg an esmate, or a control budget based on the
estimate for labor-intensive activities in indusnid consrnction. ihtifcial neural nenvoks
are capabIe of sorng out hidden patterns and estracting predictive information from
comples data sets, and are proven to be effective in both uncertainty analysis and
sensitivity analysis of construction labor productivity. The NN-based decision support
tooIs are developed to assist estimators in deciding on labor production rates for new
jobs; such tools can be more sophisticated and intelligent than traditional business
manuals or guidelines.
Meld, L. E. (1 9 8 8). Corn-tn~ction prodz~cfi~ity - on - d e rnensireme~zt md rnnn~zgemerrt.,
PvlcGraw-HiU , New York, NY.
Gerwin, E. (1992). "Fabrication and installation of piping systems". P@hg
Hdbook, uth ed., Nayyar, M L. eds., McGraw-W , New York, NY, 297-361.
Knowles, P. (1 9 9 7). Predicting L b o r Prodz~ctiuip U~it'g Arez~rid Netl vo r k ~ hLas ter of
Sciences Thesis, University of Alberta, Edmonton, AB.
Lu, M., AboURizk, S.M. and Hermann, U.H. (2000). "Estimating labor
productivity usuig probability inference neural nebvorks.", J. oJ Corzpthng i f 1 Cid
Engineenkg, ASCE, 14(4), 241-248.
Page, J.S. and Nation, J.G. (1 982). Edinoloor'r p@ing man hout- manrral, 3d ed., Gulf
Publishing Company Book Division, Houston,.
Puker, AD., Barrie, D. S., and Snyder, R. M. (1984), Pfamrig rzm? Estimuti~rg
Hemy Cot;l~'tna*tr;?n, McGraw-Hill, Inc., New York, NY.
Portas, J., and AbouRizk, S.M. (1997). N e d Netmork Mode1 For Estimating
Construction Pr oduct i ~v. 1- of Con.r/r. Engrg. e9iVlgmt-, ASCE, 123(4), 399-410.
Thomas, H. R., and Zaorsia, 1. (1999)- "Consmiction baseline productivity:
theory and practice." J. Consk Engg. Ami hlgmt., ASCE, 125 (S), 295-303.
Chapter 3: Estimating Labor Productivity Using
Probability Inference Neural Networkl
Estirnating labor production rates (m-hr/unit) is both an m and a science. In
generd, the estimator develops the rate for a given project by s t k g wth a "base rate"
and modifying it to reflect the spefic conditions he/she expects to encounter in the
project being estimated. The base-rate is often detemiined statisticdy from past
historical data, or fiom industry standards. In the context of indusmd productivty
e s t b a ~ g , the estimator accordingly adjusted the rate up or down by applying a
d i f f i c d ~ multiplier to reflect overall favorable or unfavorable conditions. In determinhg
the difhculty multiplier, consideration is only given to a couple of major factors that are
thought to affect job productivity, such as installation location of pipe (inside a
fabrication shop or on the job-site), and material type of melduig (e.g. carbon steel or
s tainless steel).
1 ,A version of this chapcer has been published. ASCE, Journai of C o m p u ~ g in Civil
Engineering, October/2000, Vol 14(4), pp 241 -218.
The chalienges of this approach include the fact that it is not straightformard to
create a conventional mathematical model so as to accommodate the impacts of
numerous factors on the target rishy variable. The deasion process relies heavilv on
individual's e'rperiences and the results are often inconsistent refleckg the esperience
and disposition of the es hat or .
Llrti&d neural nenvorks have been proposed by many as an alternative for
sneamiining the process and reducing the subjective nature of the work. Most models,
however, were based on point predictions of production rates wth which estimators
were uncornfortable. The point prediction by NN can be de he d as a single value
predicted by neural nenvork models wthout any bachp information on t he nsks of
taking chis value as correct. The new NN model presented in this paper aises out of the
need for accurate prediction in the form of a distribution at the output range. The
estimators dl be able to make a decision for a h t u r e scenario based on the results
recded by the NN model and personal preferences and esi-eriences.
In the folloming section, previous NN applications in the problem domain are
k t reviewed.
Review of NN Applications
Mosehi, Hegazy, and Fazio (1990) cite the prediction of a realistic productivity
ievel for a certain trade as an aspect of constniction that can be modeled +th neural
nenvorks. Factors such as job size, building type, overtime work and management
conditions are typically considered by an estimator and can easily be manipulated for use
as neural network inputs.
k s h e n a s and Feng (2 992) analyzed earth-movng equipment produccivity with
a neural network application. ,A modular neural nent-ork structure was used to m&e it
possible to add specikations of new equipmeat mith only a brief training session. Each
module represents a distinct type of equipment that w-as trained with trvo inputs, four
hidden nodes, and one output within a back propagation training algorithm.
Wales and AbouRizk (1993) used neural networks as a rneans of applying the
effects of environmental site conditions to the labor production rate on an activity.
Daily average ternpeIatue, precipitation, and cumulative precipitation over the previous
seven days were identified as three key environmental site conditions and used as inputs
uito a feed forward back propagation neural nemork training algorithm. The output was
a productivity factor such thac a value larger than 1.0 indicates that environmental site
conditions produce a greater than average productivity. On the other hand, a
productivity Factor of less than 1.0 indicates that the environmentai site conditions resdt
in below average productivity.
Chao and Skibniewski (1994) performed a case smdy in which a neural network
was used to predict the productivity of an escavator. They idenafied IWO main factors
that affect an escavator's productivity: job conditions and operation elements. Job
conditions include the characterisucs of the environment such as soil conditions, and
speufic characteristics of the excavator and excavation such as the vertical position of
the cutting edge. Operational elements, in contrast, include characteristics not directly
related to the escavating operation; for esample, the effect of wait time for trucks and
e'rtra tasks other than excavating. Two neural networks mere used for the purpose of
this case study. The hrst was used to estimate the excavator cyde time. Four key factors
were identihed as having an influence: cycle time (including swing angle), horizontal
reach, vertical position, and soi1 type (job conditions). The output of the hrst nenvork
was then incorporated into the second nenvork, which esamined the effect of the
operational elements on the productiviq.
Portas & AbouRizk (1993 proposed a feed fonvard back propagation neural
nenvork model for estimating construction production rates of formork. The network
outputs a single point prediction dong Mth a nurnber of output zones, with equd
likelihood of the production rate being in an); one zone. The output zones are
symmetric and divided evenly across the range of likely production rate values. During
training, the output zone whose output coincides mith the actual production rate is
remarded with a primary score of 1.0, representing strong certainty. A certain degree of
fuzziness is considered by remarding the 2 adjacent output zones mith s econdq scores
of 0.5, representing weak certainty. LU the other output zones are assigned a score of O.
Once the NN is trained and inputs are entered and the NN ~vi predict a point value as
well as the likelihood of production rates being within the output zones. This mode1
achieved limited success and its limitation was overcome by the work discussed next.
Knowles (1 997) ~resented a wo-s tage NN model in predicting pipe-ins dat i on
labor productivity. The input factors are used to invoke a LVQ classification process
and then a predictive one. With the classification, the model predicts whether the output
is likely in a cppical or non-typical range. The proper feed-formard back-propagation
netmork is then esecuted- The drawback of this method is that a budd-up of errors
occurs when the dassification fails. For instance, if the classScation accuracy is 90% at
the Fust stage of NN, and the prediction accuracy at the second stage of NN is 8S0/o, the
prediction accuracy of the whole NN is only 76.5% (90% times 85%)- This problem
motivated the development of the model descsbed in chis paper by t;iliiog a different
probabilistic approach, mhich is more direct and more meanligfd in texns of giving a
point prediction and quantifymg its assouated probability.
Introduction of the PINN Model
Specht (1991) revisited Probabilistic Neural Network (PNN) and Generd
Regession Neural Netwok (GRNN) dgonthms with the objective of integrating
statistics and neural training. GRNN/PNN is a memory-based feed fonvard neural
nenvork model, where the training is performed in one pass, thus requiring less training
tirne. GRNN/PNN is able to identify a posterior distebution over the NN weight
vectors and a point-value prediction is generated based on the predicted dkmbution.
However, based on es,-perimentaons and observations, GRNN/PNN is not quite
tolerant of noisy data (inaccurate or incomplete records) and imposes a demanding
standard of data qu;ility that is hard to achieve in realiry. The memory demand and
cornputing time for GRNN/PNN increase very rapidly when the dimension of input
vector and the quantitg of training samples increase.
The PlBN model uses similm topology as the GR.NN/PNN model but is a
rehement of i t This is because PINN generalizes the underlying statistical patterns
\vivithin training data and codes those pattems into a lLnited number of weight vectors
through iterative l e h g . As a result, the number of weight vecors d not
propomondly increase with the increase of dimensionaiity and quantity of ualliLig data.
Weighred
Average
Pro bability
Density
Graph
Figure 3-1: Topology of PINN Mode1
The PINN mode1 imbeds the output zone concepts described in Portas &
AbouRizk (1997). In the application domain of industrial labor productiviry estimating,
the profile of acnial historical productivity data reveals a spread range. The range of NN
output value, i.e. the production rate multiplier, is evenly divided into a number of sub
ranges, or output zones, which are actually some discrete clusters with continuous
boundares. The higher the multiplier value, the more diEhcult and more demanding the
job is, and the Iowa the productivity for the job. Thus, each output zone gives an
indication of the relative mork difficulty and productivity level, for instance, output zone
[O-0.q stands for easier mork and higher productiviq level cornparhg with output zone
p.7-1.41. The median of each sub range can be used to represent the spical value for
each output zone and to derive a NN predicted value.
The PLNN uses the same strategy as the model desaibed in -\bouRizk e t al.
(1999) by incorporating 1<ohonenys LTTQ in NN learning. The main difference is that the
classificarion and prediction networks are combined in an integrated netsvork, which
required the development of a different aalung and recall algorithm. Mmz a and Fisher
(1993) unlized Kohonen's unsupervised learning algorithm called self-organiung map or
SOM for modular construction dedsion making. Kohonen's L e d g Vector
Quantization (LVQ) combines unsupervised and supervisecl leaning and is
recomrnended for statistical pattern recognition problems (Kohonen, 1995). Three
options for the LVQ-algorithms (LVQ1, LVQ?, and LVQ3) mere proposed. ICo honen's
research shows that each of the three LVQ variations yields simiTar accuracy in most
stastical pattern recognition tasks, although a different philosophy underlies each
algorithm. LVQl was utilized in the l e d g process of the PINN model.
O v e ~ e w of the PINN Topology and Process
The topology of PINN model is &en in Figure 3- 1. It is composed of four
iayers. The middle layers are a Kohonen classifier and a Bayesien Iayer. The outcome of
the PINN model at the output layer is a probability density function or a distribution
r e f l e c ~g the Likelihood of the target variable occurring in a given zone- The mode of
the distnbution and its mean can serve as point predictions.
The PINN process consists of four stages as follow:
(1) Preparation, mfiich deals with
Scaling data at the input layer, which d be discussed in the subsection tided
"Data Pre-Processing";
Setting up output zones at the Kohonen layer, which d be addressed in the
subsection titled "Output Zone Setup"; and
What are Processing Elements and how they h c t i o n at the Kohonen Iayer,
mhich d be discussed in the subsection titled 'Trocessing Element at the
Kohonen Layer".
(2) LeamLig, which takes place between the Input layer and the Kohonen layer uslng the
LVQ algorithm. This does not involve the Bayesian layer or the Output layer. This
di be discussed in the subsection "NN Learning Process".
Once leaming is achieved, the input-output patterns are coded into the weight
vectors of the processing elements at the Kohonen layer.
(3) Investigating whether the neural network has been successhlly uained. This is
accomplished through the followng steps:
Feed the input vectors of the training and t e s ~ g records into the input layer of
the PINN model.
Project the input vector of one record onto the Kohonen layer by using the
results of stage (3). The Eudidean distances between each processing element's
weight cector and the scaled input vector are caiculated.
-ln in-zone cornpetition occurs wthn every output zone at the Kohonen layer,
which is detailed in the subsection titled "In-Zone Cornpetition Stcitegy ar
Kohonen Layer". The processing element wth the minimum Euclidean distance
value nruis.
Project the minner PE at the Kohonen layer onto the Bayesian layer. The
Bayesian layer holds a probabiliy deasity hncuon (PDF) approsimator. The
Euclidean Distance d u e s of the winner PEs are the inputs to the PDF
approlrimator. The folloMng subsection of 'TDF .ippro'dmator at Bayesian
Layer" discusses the components and operations in more details.
The output is mapped fiom the Bayesian layer and presented in the f o m of a
probability density function at the output layer. Two point predictions are
calculated in addition to the probability density hinction namely, the mode and
the weighted average. The subsection "Outputs at the Output Layery' Licludes
details about die NN outputs.
Check the NN outpua against the achial outputs of the training and teshng
records. If the results are satisfactoq-, then the neural network is dzclated to have
been trained; othernrise, repeat stage (1) using different panmeters at each layer.
(4) Recd. Once the model calibration is done, the nemal nework can be used to r ecd
the output value for any given input vector, mhich is similm to stage (3) usiag the
h a l results detemiined in stage (2) and (3). A sample cdcuiation is given in the
subsection "Sample of Re c d Process".
Data Pre-Processing
At the input layer of the PINN model (shown by Figure 3- l), the number of
input nodes corresponds to the dimension of the input ~~ect or . The dimension of the
input vector depends on the number of input factors and the input data types. Three
input data q.pes are used to d e h e NN input factors, i.e. "Raw", "Rank", and "Bina.$"'
"Rad' is used sirnplv for quantitative input factors, like general espense ratios, winter
construcon percentages, or quantities of work. " Rd " is used to conven subjective
factors, like crem abiliq rabngs, into numeric format. And "Binary" is used to group
testual factors into numeric formats, lilie material type and project dehnition. It shodd
be noted an input hctor of the "Raw" or "Rank" type corresponds to one input node at
the input layer, while an input factor of the "Binary" type corresponds to a number of
input nodes depending on the number of groups for the factor. For illustration, input
factors and data types for the "Pipe Installation" neural neworlc mode1 are listed in
Table 3-1. A sample record for the "Pipe Instdation" NN training is also listed in Table
3-2 showing both the ram data and conrerted NN input data. The NN input data is
norrnalized and scaled betsveen O and 1 at the input layes. These scded inputs dl be
passed fornard Fur NN training. At the Kohonen layer all weight vectors are rmdornly
initiillized betnreen O and 1.
Table 3-1: Input Factors and Data Tne of PINN Mode1
NN Input Factor
Project Location
Adminis mauon
Year of Construction
Proxrince/S tate
Contract Type
Client
Engineering Firm
Pro j ect Manager
Superintendent
Project Definition
Work Scope
Project Type
Prefab/Field Work
Average Crem Size
Peak Crew Size
Uninized
Equipment & Ma tend
Estra Work
Change Order
Drmring & Specs Quality
Location Classification
Total Quantiy (ZearnLlg)
Installation Quantities
Matenal Type
Method O f Installation
Pipe Supports
Bolmps
Valves
Screm-ed Joints
bfisc. Components
Welding Impact
Season
Crew Ability
Site Working Conditions
Inspection, Safety & Quality
Overall Degree of Difaculty
Data Type
(2)
Binary
Raw
Binary
Binary
Binary
Raw
Raw
Raw
Raw
Binary
B inary
Binary
Raw
Binary
Binary
Bi n q
Raw
Raw
Raw
Rank
Binary
Raw
Raw
Binary
Raw
Raw
Raw
Raw
Raw
Raw
Raw
Raw
Rank
Rank
Rank
Rank
Options & Remarks
(3)
Urban, Rural, Camp Job
General Ex~ense
89-92,73-94,75-96,97-79
M3, SK
Reirnbersable, Lump Sum
an index derived fiom historical data
an index derived kom historical data
an index derived from historical data
an index derived from histoncal data
Chernical, Cryogenic, Gas, Refining
Confincd / Scattered
Upgrade Shutdown, Grass Root etc.
Percentages for Prefabrication
<25,25-50,50-100, >IO0
<25,25-50,50-100,100-150, >250
Yes, Ko
Equip.& Mat1 Cost/ Direct MH
Original Project Cost/Final Projeject Cost
No. of Change Orders/Total Direct MH)
1 Poor 3 Average 5 Excellent
U/G on Site, Fab Shop, .i/G on Site etc.
Total Quantity In DiaInFt
Qty for Size Ranges, <Y, 2"-16", >16"
.ilioy,Carbon Steel, FRP/PVC,etc.
Percentages of Hand Rigging
No. of Pipe Supports/Foot of Pipe
No. of Bolmps/Foot of Pipe
go. of Valves/Foot of Pipe
90. of Screwed Joints/Foot of Pipe
[nstall hfisc.Components MH/Foot of Pipe
Welding Multiplier (bfiscoding on Site)
Percentages of Winter & Summer LVork
I Very Low, 3 Average 5 Trery High
L Extreme Problems - 5 No Problem
L Extremely Detailed - 5 Highly Toleranr
1 Very Low 3 Average 5 Very High
Table 3-2: Innut Data SamtAe of PINN Mode1
NN Input Factor
(1)
Project Location
Admi& tra tion Requiremen t
Year of Construction
Province/S tate
Contract Type
Client
Engineering Firm
Pro ject Manager
Supe ~t e nde nt
Project Dehnition
Work Scope
Project Type
Prefab /Field Work
Average Crew Size
Peak Crew Size
Uninized
Equipment & Ma t c d
Extra Work
Change Order
Dra~ving & Specs Quality
Ac tivity Location Classification
Total Quantity (Learniflg)
Installation Quantities
Material Type
Method Of Installation
Pipe Supports
Boltups
Valves
Screwed Joints
h/lisc. Components
LVelding Impact
Season
Crew Abiliq
Site Working Conditions
Inspection, Safety & Quality
Over d Demee of DXficdtv
Raw Data
Rual
0.235
91-95
Alberta
Reirnbersable
S hell
Colt
John Doer
Bob Smith
Chernical
Confhed to Speciic Area
PIant Upgrade No Shutdown
10/o, 90/o, 0%
25-50
50-100
Yes
9 -4
0.661
0.029
Excellent
Inside <loft E-Figh
6055
210,905,4940
Carbon Steel
Hand Rigging0/0, Machine Rigging ' / O
0.45
4.77
1.59
O
3.18
1.25
WinterO/o, SumrnerO/o
Average
Many Problems
De tailed
Low
NN Input Data
(3)
Output Zone Setup
-1s discussed in the section "Introduction of the PINN rnodel", the likely range
of output values is evedy divided into a number of output zones. The output zone
boundary semp is important for PINN l e d g and recall. \Vide zones are generally not
helpfd to the decision-maker and hence should be aroided. Zones that are unacceptably
tight may prevent PINN h m l e d g . It requires some aials to obtain reasonable
output zone boundaries and the following two aspects should be considered:
1. Presion requirement of the user, i.e. the zone width or sub-range that suffices for
the user to make dedsions.
2. Dismbution of actual output data over the output zones. h uniform distribution of
actual output data over aU zones generaily yields better results.
Processing Elements (PE) at Kohonen Layer
At the Kohonen layer, each output zone contains an equal nurnber of processing
elernents. Each processhg element is associated with a weight vector (also referred to as
a codebook vector (Kohonen, 1995)).
Visually, a weight vector is a set of links that emanate Lom one processing
element and end at each input node as illustrated in Figure 3- 1. Th s the dimension of a
weight vector is equal to that of the input vector (the number of input nodes). An output
zone at Kohonen layer can be visualized as a chip containing a number of pins (PEs).
DuPng NN trnining the orientation of those pins is gradually he-tuned to capture the
underlying s taustical patterns Mthin the training data (Kohonen, 1995).
Our expenence uidicates that the nurnber of processing elements nssigned to
each class should be close to the average frequency in the histogram of training data
output values, i.e. the average number of naiaing samples in one output zone.
NN Leaming Process
Data of al l the training records is scaled at the input hyer. The scaled input data
is fed into the mode1 to calibrate the weight vectors of the Processing Elements (PE) at
the Kohonen layer, using the LVQ aIgorithrn suggested by Kohonen (2995).
The leaniing process involves a number of iterations, each of which is
comprised of the following:
1. The Euclidean distances between the input vector of a training record and each PE's
weight vectors are calculated. The PE that has the smallest Euclidean distance vdue
is declared to be a global winner. If the global \vinner PE does not belong to the
same output zone that the actual output value of this training record Falls into, the
weight vector of the global Ninner PE is pendized according to the Foliowing
equation (1):
RR stands for "Repulsion Rate", which is a Ieaming rate to penalize the global
winner PE;.
X, is the input vector of the training record, and
Wii is the m e n t weight vector of the global mnaing PE,.
Wii' is the updated weight vector of the global whi ng PE,.
The repulsion rate is iniaally set berneen O and 1, and is reduced gradually und it
approaches O at the end of learning.
2. Followhg the global competition, an in-zone competition among processug
elements occurs only at the output zone into whkh the acnial output value of the
training record fds. Prior to the Li-zone competition, a "conscience" value is added
to each PE's Euclidean distance and adjusted over the learning iterations, so as to
effecuvely prevent one PE nithin a specific output zone from whi ng all the cime
and activate as many PEs as possible in the leaming process. The formulas to
calculate the conscience Euclidean distance c m be found throughout the pertinent
literature. The interested readers can refer to Appendk II for details.
The
The
method we adopted is as Follosvs:
cc
conscience" Euclidean distance between each PE's weight vector and the
input vector is calculated. The PE wth the shortest "conscience" Euclidean distance
value is declared to be an in-zone m e r . Onlp the in-zone minner PE is remarded u s h g
equation (2):
Where:
AR stands for "-Attraction Rate", which is a leaming rate to remard the in-zone
+er PEi-
X, is the input rector of a training record, and
W, is the current weight vector of the in-zone %&mer PE,.
W,' is the updated weight vector of the in-zone whi ng PE,.
Wie the repulsion rate, the attraction rate is initially set benieen O and 2, and is
reduced gradua* u n d it approaches O at the end of l e k g iterations.
-1 sample calculation of one learning iteration is presented nest to illustrate the
learning process.
As shomn in Figure 3- 2, the dimension of input vector is 12, and the output
range ([O*]) is divided into 1 output zones, Le. [O-11, [l-21, [2-31, [3-4-1. Each output zone
contains 3 processing elements. Note that this sample is a simple mode1 and serves for
illustration. Problems encountered in a r ed situation, wliich are suitable for the PINN
mode1 to solve, are mostly high dimensional; the number of input nodes mav esceed
100.
The input vector of a training record is scaled benveen O and 1, and the weight
vector of a processing element nt the Kohonen layer is randomly initialized between O
and 1. Table 3-3 shows the input vector XI and the meight vectors of the 3 processing
elements in zone 1.
Table 3-3: Scded Input Vector and Initial Weight Vectors
Input Vector
The Euclidean distance (ED) is calculated berneen the input vector X, and each
weight vector Wi, as ED,; = 1.4163, ED2, = 1.5746, and ED,, = 1.2963. Suppose that
PE3 in zone 1 is the global e e r PE among d the processing elements, svhich gives a
minimum ED value of 1.2963. If the actual output of this training record does no t Ed
into zone 1, i.e. outside the sub-range [O-11, then the meight vector of PE, (W,,) is
updated by equation (1) as shown in Table 3-4. Notice that the Repdsion Rate is set to
be 0.8 at the staa of learning in the sample calculation. Kohonen (1995) recommends
smaller initial value such as 0.06 for the Repulsion Rate and -Attraction Rate for obtaining
better resdts.
Table 3-4: Updating Weight Vectors in First Learning Stage
Input Vector
If the actual output of this training record does fd into zone 1, i.e. nrithin the
sub-range [O-21, then no weight vector is updated in the global cornpetition. The learning
process steps into the second phase.
In the hrst training iteraon, the conscience value for every processing element
in zone 1 is detemiined to be equal to O (see hppendk I). So the "consuence" Euclidean
Distance value is equal to the onginal Euclidean Distance value. PE3 is the in-zone
Figure 3-2: Operations at Bayesian Layer in Re c d
competition b e r , which gives the minimum "conscience" ED value of 1.2963, and its
weight vector W3, is updated by equation (2) as shown in Table 3-5.
Table 3-5: Updating Weight Vectors n Second Leaning Stage
Input Vector
The above learning process d iterate through al1 the training records for a
suffiuent number of m s . Notice that during the leaming process the Repulsion Rate,
Attraction Rate, the conscience value are dynamically updated to calibrate the weight
vectors.
In-Zone Cornpetition Strategy a Kohonen Layer
The in-zone cornpetition at the Kohonen layer occurring in the r ecd stage
differs from that o c d g in the leaming stage. Once adequate trainhg is complete, the
PINN is capable of mapping the input ont0 the output. At the Kohonen layer, for one
output zone, the PE that has the shortest global Euclidean distance (no conscience
value) between its meight vector and the input vector is declared to be an in-zone wnner
PE. Only the in-zone ber PE advances to the Bayesian Iayer.
UnWre the GRNN/PNN, which takes the average of PE's Euclidean distance
within one output zone as the panmeter to pass fornard (Specht 1988), PINN takes the
minimum of the PE's Eudidean distance wtbin one output zone as the parameter to
pass onvard. The reason for the difierence is that in GRNN/PNN, each PE
corresponds to one training record, and differenc numbers of PEs Lie in different output
zones. In the proposed PINN, the PE does not match the training record, and an equd
nurnber of PEs dwell in each output zones and mode together in the Kohonen layer of
PINN to generalize the underlying patterns within the 6 n i n g records by mplementing
LVQ.
PDF Approximator at Bayesian Layer
As illustrated in Figure 3- 2, at the Bayesian layer each output zone only contains
the minner PE feom the in-zone competition at the Kohonen Layer in the recall stage.
The main components at the Bayesian layer are a kernel function and a Sotmas
Activation fnction that are used co approximate the probability density of one input
vector being wvithin each output zone in the steps as oIlow:
1. The square of Euclidean distance value of the minner PE from each zone is passed
into the kemel function, which is the Gaussian hinction of Bayesian methods in
sratistics as described in Specht (1988) and shomn in equation (3). If the number of
output zones is N, then for each input vector Xj, the kernel function is evduated for
N times and output one "q" value for each zone.
where: i =1,2,3,. . .N, N is the nurnber of output zones;
Xj is one input vector fed into PINN at the input layer;
(QUii-XJT(~,,-X,) is the square of the Euclidean distance value between the input
vector X, and the winner PEys meight vector Wii in the output zone i, i =1, 2,3,. . . >N.
(T is a smoothing factor, and is the only adjustable parameter of the Gaussian
hc t i on (3) and controls the shape of the probability density hct i on. The grearer a,
the more dispersed the probability density graph. CT is critical to PINN1s predicting
capability and can be detemiined through iterative adjustments. In regard to industrial
productivity application, 0 should fd in the range between 0.8 and 1.2.
The Sofmia-u Activation function (Sarle, 1997) as shomn in equation (4) makes
the surn of the calculation results (q values) from (3) equal to one, so that the final
output Erom the Bayesian layer can be inrerpreted as posteior probabilities ("p" values).
where: qi is the output value 6om Gaussian h c t i o n (3) for output zone output
zone i, i =l, 2,3,. . .,Ne
N is the nurnber of output zooes.
Outputs at Output Layer
At the output layer, the probabiliq distribution predicted by the PINN is
presented in the form of a Probabiliy Density Function graph, which portrays the
uncertainq of the output value.
In addition to the predicted distribution, PINN calculates two point prediction
values:
1) Mode value: the median of the output zone or sub-range that has the greatest
probability.
2) Weighted .Average Value: the sum product of the median and the probability of each
output zone. The user should neat ths point-value prediction carefully by checking
the probability density hinction grap h.
Sample of Re c d Processing
h sample PINN recall calculation is given in the folloming sections for
iuus tration:
Suppose that the NN is mained and ready to recall the output for an input vector.
A s shomn in Figure 3- 1, the dimension of input vector is 12, and the output range ([O-
41) is divided into 4 zones, ie. [O-11, [l-21, [2-31, (3-41. Each zone contains 3 processing
elements.
Table 3-6 lists the scaled input vector X, and the weight vectors of the 3
processing elements in zone 1.
Table 3-6: Trained PI NN Ready to Recail for A Given Input Vector
Input Vector
The Euclidean distance (ED) values between the weight vectors and the input
vector are calculated as ED,, ~0. 9513, ED2, = 1.1670, ED,, = 1.3249.
At output zone 1, processing element PE, with a minimum ED value (0.9513) is
the in-zone cornpetition whner, and proceeds to the Bayesian layer.
Suppose the b e r PEs fiom the other 3 zones are also detemiined in the
similar manner and proceed to the Bayesian layer. Table 3-7 lists their Euclidean distance
values and outputs fiom the Gaussian function and the Sofma-.: function.
Table 3-7: Recail Calculations at Bayesian Layer
From Table 3-7, the probability of output being within zone 1 is 0.731, hence the
mode output value is found to be the median of zone 1, i.e. 0.5. The weighted average
output value is obtained by calculating the sum-product of the Sofmiax output @ values)
and median of each output zone, i.e.
Values for Each Zone
(1)
Median
Winner PE's ED
Gaussian Output q:
Sofmna,~ Output p:
IMPLEMENTATION OF THE PIPIN MODEL
Cornputer s o b a r e based on the PINN mode1 is developed for learning and
testing in the environment of MS Access 97 and Visual Basic for .ipplications. Historical
Zone 1
(2)
0.500
0.951
0.636
0.731
Zone 4
( 5)
3.500
3.922
0.014
0.016
Zone 2
(3)
1.500
2.567
0.037
0.043
Zone 3
(4)
3.500
1.843
0.183
0.210
piping productivity data of 66 projects resulciog in 119 records of a construction
Company mas collected and compiled into NN input data for three Iabor-intensive
activities, i.e. pipe installation, pipe welding and pipe hydro-testing. In the following
sections, pipe installation is used to illustrate the testing and validation of the PINN
model.
The PINN model for "pipe installation" has a total number of 81 input nodes.
(The input factors and a sample data are shomn in Table 3-1 and 2). The output range is
divided into 20 output zones with an equal width of 0.72. 10 processing elements are
assigned to cach output zone. The attraction rate and repulsion rate are both equal to
0.06. The srnoothing factor of the kemel function is equd to 0.8.
One hundred one records are used for PTNN leaming, while 18 records are
reserved to test the calibrated model. The learning process takes 300 iterations.
Validation of the PINN Mode1 on Testing Data
The testiog of the calibrated nemork on the 18 unseen records was surnmarked
in Figure 3- 3. Measured agahst the actual output values of the test data, for the mode
value, the average absolute error is 0.57, and the rn~uimum absolute error is 2.02; for ehe
weighted average value, the average absolute enor is 0.75, and the maximum absolute
error is 2.23. Considering a wide output range of about 15, the error is reasonable and
acceptable.
To compare the PINN model Mth a back propagation neural network, the same
training records were used to train a t hee layer feed fonvard back propagation neurai
87
network, mhich has 81 input nodes at the input layer, 40 hidden nodes at the middle
layer and 1 output node at the output layer. The ~aining parameters are l e d g rate
equal to 0.8, momentum rate equal to O.+ the transfer hc t i on is symmemc logisac
nction. Afier training was completed, the tesMg set of 18 records preriously used to
test PINN mas fed into the model. The tesng resdts of the back propagation NN
model comparing with that of the PINN are shown in Figure 3- 3. From Figure
observed that the PINN rnodel outperforms the back propagation NN model
of point prediction accuracy, coming closer to the actual output values.
PINN VS FF BP NN
+PINN (Mode)
7
1 - -A - FFBP NN
in terms
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18
Test Record #
Figure 3-3: Comparison of PINN and Back Propagation NN
Figure 3-4: PINN Output for the Base Case Scenario
Sensitivity Analysis of the PINN Model
A r ecd program based on the trained
developed so as to di da t e its effectiveness
PENN model for pipe
and accuracy in the
installation was
contest of the
application domain. "What if' scenarios are tested on the NN model
input factors in order to understand the impact of such changes on
by changuig some
the output values.
The response of the NN model is compared against chat of an expenenced estimator at
the participakg consmction Company for the purpose of model validation-
The base case scenatio is taken from one testing record. The acnial difficulty
multiplier for this scenario is 1.24. The mode value predicted by the PINN model is
1.181, giving an absolute enor of ( 1 1.181 - 1.24 1 = 0.059); and the weighted average
value is 1.313, giving an absolute error of ( 1 1.313 - 1 .X 1 = 0.073)- Figure 3- 4 shows
the predicted probability h c t i o n or distribution, the chance of output f&g into zone
2 (10.8-1-51, median = 1.181) is 69O/0.
In the following validation tests, the acmai values remain unknown, so the
responses of the estimator based on personal experiences and common senses serve as a
benchmark to measure the performance of the PINN model. The esmator responds
widi a trend or direction insread of a precise number because there are so other input
factors to take into account.
The location of pipe installation is a major consideration when the esperienced
estimator detemiines the pipe installation productivity. The instahtion location for die
base case scenario is "Piping nrithin a fabrication shop", what if the location is changed
to "Operating plant installation on the site"? The esperienccd estknator responds by
increasing the difficulcy multiplier to a certain estent to reflect the unfavorable job
conditions. Response of the PINN model is shomn in Figure 3- 5. It is observed the
mode value increases to 1.903 and the weighted average value increases to 1.983; the
chance of the output value falling into zone 3 (l1.5-2.31, median = 1.903) increases fiom
13% in the base case scenario to 78%. The PINN
estunator in the decision process for this scenmio.
has taken the same direction as the
In the base case scenario, the job is done 100% in the winter in Alberta. m a t if
NU RecaIl Probability Density Graph
1 .O
Figure 3- 5: PINN Output for Scenario 1
the job is done 100% in the summer? The estbator anticipates a reduction for the
diffidty multiplier, mhich rneans an increase of productivity level. Response of the
PINN mode1 is shown ui Figure 3- 6. It is observed the mode value rernains 1.181,
homever, the weighted average value decreases to 1.151. The chance of the output value
falling into zone 2 ([0.8-1.51, median = 1.181j decreases significantly from 69% to 44Y0,
while the chance of the output value falling into zone 1 ([0.2-0.81, median = 0.5)
increases si@cantly from 15% in the base case scenaro to 40/o. Again the PINN
chooses the sirriilar course of action as the estimator in the decision process for this
s cenario.
NN Rewl ! Probabifity Demity Graph
i .O
Figure 3-6: PINN Output for Scenario 2
The PINN model creates a meaningful representation of a comples, real-Me
siniaaon in the problem domain and is effective in dealing with high dimensional input-
output rnapping mith multiple influentiai factors in a probabilistic approach. The
application of the PTNN model in indusmial labor production rate e s t i ma ~ g helps the
estimator choose a course of action by gving a better understanding of the project
information available and the possible outcomes that could occur. Bccause the
probability densicy of each output zone is provided, the predicted dismbution and point-
prediction values give the estimator much more confidence in the predicted result. In
combination of the personal esperiences and preferences labor production rate for a new
project can be deterrnined.
Chao L.C., and Skibniewski, M.J. (1 993). "EstirnaMg Construction Productivity:
Neural-Network-Based Approach." Joumal of C O ~ ~ U M ~ in Civil Engineering, ,lSCE,
8(2), 234315
Hermann, R. and Lu, hl. (1997) "-ipplication of Neural Networks in Industrial
Estimating", Proceedings of the 27& Canadian Society of Civil Engineers -innual
Conference, Edmonton, AB, 15-35.
IKnowles, P. (1797). " Pr edi c~g Labor Productivity Using Neural Nenvorh."
Mas tcrs of Science Thesis, University of Alberta, Edmonton, AB.
Karshenas, S. and Feng, X.( 1992) "Application of Neural Networks in
Earthmoving Equipment Production EstimaMg." Cornputhg in Civil Engineering,
Proceeding of Eghth Conference, Dallas, Texas, 841-847
Kohonen, T. (1 995). "Self-Organiziag Maps", Springer Series in Information
Sciences, S p ~ g e r , London, U.K.
Moselhi, O., Hegazy, T., and Fazio, P. (1990 ) 'Teural Newodc as Tools in
Consmichon." Journal of Construction Engineering and Management, -ASCE, 117(4),
606-623
Murtaza, M.B., and Fisher, DJ- (1993), "Neuromes: Neural Nesvork System for
Modular Consmction Decision Making", Journal ComputGig in C i d EngineeSng,
,%SCE, 8(2), 221-333.
Portas, J., and AbouRitk, S.M. (1997), c%Jeucal Nenvork Model For Estimating
Construction Productivity." J. of Constr. Engrg. & hlgmt., ASCE, 123(4), 399-410.
Sarle, W.S., ed. (1997), Neural Network F-IQ, part 1 of 7: Introduction, penodic
posting to the Usenet newsgroup comp.ai.neur-al-ne ts, URL:
fip://ftp.sas.com/pub/neural/FAQQhtrriI
Specht, D. F. (1 988). Trobabilis tic Neural Networks for Classihcation,
Mappings, or =lssociative Memory." IEEE International Conference on Neural
Nenvorks, 1988,1,525-532.
Specht, D. F. (1991) "Generd Regression Neural Nenvorks.'' IEEE T.cansactions
on Neural Netsvorks, Nov. 1991,2, 535-5176.
"Conscience" Euclidean Distance is defined as (5):
Di' = Di + Ci
Where:
Di is Euclidean distance,
and Ci is conscience value (6):
Ci =cf x ~ x ( n x w f -1)
Where:
D is the r n~&um Euclidean distance out of the global cornpetition in the
previous s u p e ~ s e d leamhg stage,
cf is a Conscience Factor, mhich is initially set berneen O and 1 by the user,
and n is the number of PEs per output zone,
wf is dehned as "Win Frequency", and the initial estimate of the Win Frequency
value (do) is set to the reciprocal of PE Number per Output Zone for dl the PEs, i.e.
1 /ne
With NN learning ongoing, both Conscience Factor (cf) and Frequency
Estimate (fe) are reduced graduaiiy und it rpproaches O at the end of leaming.
During the uns upe~s ed l e h g stage, for the in-zone whner PE, its mf value is
updated as (7):
For the in-zone loser PEs, th& mf values are updated as (8):
Basically, the above formulas are intended to increase the wf values for the
wimer PE and hence increase its conscience value so that the whmer PE d l have less
chances to again than the other loser PEs in the following learning iterations.
Ar t i f i d neural netrvorks (NN) mimic the cogniuve learning process in the
human brain, and deal effectively wdi dl-stmctured problems, in which the algorithms
required to solve them cannot be given in a precise and e~;pliut fashion, or the datx for a
particular problem are either not complete or cannot be specified precisely (Widman et.
al., 1989). NN has been found to be capable of performing parallel computations on
different tasks, such as pattern recognition, Lnear optimization, speech recognition, and
prediction @Iukherjee and Deshpande 1995).
A version of this chapter has been subtnitted for publication. ASCE, Journal of Computiag in
Civil Engineering.
In recent years, Back Propagation NN (BPNN) has been researched and applied
as a convenient decision-support tool in a variety of application areas in civil engi nee~g,
induding moduIar construction decision making (Murtaza and Fisher, 1993), structural
analysis (Flood and Kat h, 1994), e s t k a ~ g construction productivity (Portas and
AbouRizk, 19977, mode choice analysis of fieight transport market (Sayed and Ra z a i
1999), construction md x p estmating (Li et al 1999), measuring organizational
effectiveness (Sinha and blcI(im, 2000), and predicting settlement during runneling (Shi,
2000). The speud l e d g aigorithms of BPNN are capable of perfonriing high
dimensional, non-linear input-output mapping and e s t r a c ~ g hidden patterns and
predictive information from observing the l e d g esamples.
However, Leamkg algonthms such as BPNN do not attempt to infer causality,
hence, classification or prediction is based on blind correlation of new esamples with
preriously analyzed esamples, nrithout giving information on the effecr oE each input
parameter or inauencing variable upon the predicted output variable- In the repoaed
NN applications, model validation has thus far relied upon measuring accuraq of the
calibrated network to an independent testing data set that are hidden from the neural
nenvork in learning. The model's sensitivity to changes in its paramerers is generally
probed by t e s ~ g the response of a mature nenvork on various input scenarios. In short,
a NN mode1 funcrions like a "black box" package, giving no clue on (1) how the ansmers
or model outputs are obtaked; (2) how the input parameters affect the output.
Widrnan et. al (1989) pooulted out that the credibiliry of an AI program
frequently depends on its ability to explain its conclusions. Lack of interpretability is a
pi t f d of the neural netmork models recognized by many and has inhibited NN from
achieving its full potential in real-world applications. Dhar and Stein (1997) argued that
because NN algonthms such as the back-propagation NN are non-linear, high
dimensional fimctional equations feaniring pardel distributed data processing, it is hard
to e-xplicitly interpret which parameters cause what behavior in the NN model. i'vhile
mathematical and operational methods do esist for the analysi of neural networks, the
methods are fairly involved, and are Iess than satisfping because of th& theoretical
assumptions. They stated that "unlike most statistical methods, it can be difficult to say,
even in general, which variables are signrficant in what respect." (Dhar, et.al. 1997)
Our research intends to address the idenfied issue by concent r a~g on
sensitiviy analysis of BPNN. Similar to regression analysis, the sensitivity of an NN
input pxrameter could be expressed as the hrst-order partial derivative benveen an NN
output variable and the input paiameter. In the "Literanire RevieW" section, we bnefly
introduce several related methods for knodedge eirplanaaon and factor analysis of NN
found in iiterature. In the section entitled "BPNN Algorithm and Input Sensitivity", the
back-propagation NN algorithm is described kst , followed by the derivation of
mathematical relationships between an hTN output variable and NN input space in lighr:
of both normalized data and original data. The follonring section "BPNN vs. Regression
Analysis" discusses the difference between BPNN and regression analysis of statistics
and demonstrates the sophistication and superi~rity of BPNN over regression analysis in
a case study based on a s md data set. Nest, statiscd analysis of input sensitivity based
on Monte C d o simulation is described in the section entitled "Statistical Analysis of
Input Sensivity" to understand the rationale of BPNN's reasoning and the effectiveness
of model implementation in a probabilisuc fashion. In the "Industrial Application"
section, the new approach is applied to estimate the labor productivity of spool
fabrication in an industnal setting, and important aspects of the application including
problem definition, factor identification, data collection, and the tesbng resuln bascd on
real data set are discussed and presented.
Li et al (1999) realized the inability of BPNN to provide e.xplanations on its
output negatively affects the user-acceptance of BPNN. key investigated the use of
KT-1 method for automatically estlacting d e s from a mature BPNN in an atternpt to
explain why and how BPNN makes a p a r t i d u recommendation in developing a
decision support tool to e s h a t e the consmiction markup. KT-1 method is a heuristic
approach to generating conhmng/disconfimiing d e s fiom each hidden or output
node based on the weighted connections and the threshold value of each node, and is
constrained by the complesiv of network structure. As pointed out by the authors, such
automatic de-extraction systems as KT-1 cannot marrant a N y informative
euplanation faality because KT4 lacks the associative knowledge (Le. cornmon sense,
professional knomledge etc.) in its de-esnacting process (Li, et. al., 1999).
Sinha and bIcIGn (2000) utilized BPNN to measure organizational effectiveness
of construction h s . They have applied statistic analysis methods, such as Principal
Component Analysis (PCA), stepwise regression and conelation analysis on n BPNN
model in an attempt to identify the dominant factors that influence the target output
variable, and further to reduce the dimension of input space. Homever, the theoretical
underpinning of such statistical techniques requires carehil study of rheir applicability in
sotring real problems. For instance, lacking n amareness of the assumptions of least
squares regression (nomality, homoscedasticityy independence of errors, and linearity)
may cause the rnisuse of regression and correlation analysis (Levine et. al, 1998); use of
PCA, which assumes linear relationships betsveen variables, Wht bias the selection of
deteminant factors bp escludng those that have non-linear relationships with the target
output variable @Zefenes, et. a1.,1995).
Alternative approaches to factor analysis include using auto-associative back-
propagation neural nenvorks to perfonn non-linear dimension reduction and sensitivity
analysis (Cano1 and Ruppert, 1988). The neural nemork has one hidden layer mith k
hidden processing elements, where k is less than the dimension of input space IZ. The
output space is a repiica of the input space. Analogous to the rationale of PCA, this
method is to compress data by representiog many variables by a few components: if we
can reproduce the input space using k (k < n) hidden processing elements nrithout loss
of information, then the activation values of the 6 processing elements in the hidden
layer d compute the 6rst k principal components at the input space, under appropriate
conditions. One pidall of this method obseved by Refenes et. al (1995) is that the
stochastic nature of the data-generating process at the neural nenvork input space may
cause high variance in the analysis results. It is also noted that the output variable is
excluded from the auto-associative neural network analysis, hence, such analysis of input
parameters or influenchg factors does not take into account the relationships behveen
the input parameters and the output variable.
Explanations on the importance of input parameters can also be obtained by
examinkg the weights of a mature network so as to chwacterize the strengths of the
relationship betmeen inputs and outputs. Knowles and AbouRizk (1997) added up the
absolute value of weights from one input node to every hidden processing element in a
trained BPNN model wth only one hidden layer for esUmating pipe installation
producuvity. The total weight vdue of an input node may iadicate the intensity of the
connection from the input node to the hidden layer of the nemork; the higher the sum
value, the more signihcant the input parameter is. rilthough this heuristic approach is
straightfonvard to understand and easy to use, the resdts may bc unstable or inaccurate
due to the fact that it fails to take into account the co~ect i ons between the hidden layer
and output layer. In modeling the behavioral mode choice of the US. Geight transport
market, Sayed and Razavi (1999) combined the leaniiag abilitp of BPNN and the
transparent nature of hzzy logic in order to explain the knowledge contained in a
BPNN model, which is stored in the form of a weight ma& that is hard to interpret.
The neuroftxzzy rnodel Ocilitates the selection of si pfi cant variables that affect the
output and displays the stored knowledge in terms of hzzy linguistic rules (Sayed and
Razati, 1999).
Based on the above methods found in literanire, the effect of each input
parameter on the output variable in t e r ms of magnitude and direction s di remans
unknown, i.e. the input sensitivi~. In the follonring section the algoPthmic aspecrs of the
BPNN model are studied in order to de he the input sensitivq in an exact mathematical
term.
Figure 4-1: Structure of Back-Propagation NN Model
BPNN ALGORTTHM AND ~[NPuT SENSITMTY
Frorn the biological perspective, BPNN is origindy proposed as an AI model to
simulate the cognitive Ieataiag process in human brain, in which millions of murons are
ktercomected and interact with one and another through cornplex electrochemical
reactions and signal processing. An d c i a l NN model such as BPNN is merelv an
over-simpEed representation of the real NN in tems of mechanism and structure. h
typicai BPNN has a multi-layer structure. Each layer contains a number of processing
elements (PE) or nodes, mhich are M y interconnected bemieen layen (Figure 4-1). The
ntensity of connection betmeen trvo processing elements is represented using a weight.
Put into the perspective of mathernatics, BPNN is essentially a gradient-decent
optimization algorithm to search for the optima in a high-dimensional weight space with
the objective of minimizing the global eirror benveen NN output values and acmd
output values. An iterative weight-adjus~g scheme is used to modiq the weights of alI
the connections in the NN structure in a stepwise fashion.
BPNN Aigorithm
The basic formulae to describe signal processing of a PE in BPNN are simply as:
Where:
Subsapt c stands for a processing element in the hidden laper or output iayer of
BPNN;
Subscnpt i is the node index at the previous Iayer of BPNN;
W, stands for a weight value betnreen node i and node c;
S stands for the input signai to a node;
N stands for the output signal from a node.
Equation (1) shows that a processing element receives a weighted linear
combination of input signals rom the previous layer. Equation (2) is Sigmoid (iogistic)
h c t i o n and is the most comrnonly used transfer (squashg) h c t i o n in BPNN,
through which a processing element transforms the input signal into an output signal.
Note that a bias node Mth constant input value -1 Erom the previous laper is aiso
connected to a processing element and involved in the calculation, representing the
activation threshold of a processing element (Figure 4-1).
The digital s b a l s flom through the BPNN foilowing (1) and (2) from layer to
layer und the output layer is reached.
The global error (E) of the BPNN optimization search is expressed in (3):
Where:
N stands for the output signal Gom BPNN;
D stands for the target d u e ;
Subsaipt i is the index of records in the training data set and T stands for the
size of training data set.
The weight of the BPNN is adjusted using the delta d e to move to the opposite
aE
direction of - as (4):
awpc
aE
=- h. N .-
as,
h is a gain ratio in (0,1), also called learnulg rate, which sets the pace of BPNN
Iearning;
Subscrip t p stands for a processing element in the previous layer of the nenvork;
Subscript c stands for a processing element in the m e n t layer of the network.
For a processing element
For a processing element
at the output Iayer of BPNN,
at the hidden layer of BPNN,
T --
In (6), subscript n is the index of processing elements in the nest layer, and J is
the to tai number of processing elements in the nest layer.
Usudy, a momentum tenn is added to the weight adjusting scheme to take into
account the meight change in the previous step as shomn in (7).
dE ?
AW,, =-A- N O -
P as,
+CL * q c
9
AWpc s the weight change in the previous step, and pi s the momennun ratio
vhich is usudy less than h .
BPNN adjusts meights following (7) by observing the training data set repeatedly,
until the global error E is reduced to an accepted l m1 to declare the BPh' N to be
crained.
The BPNN shodd have at least three layers: the input layer, one hidden Inyer,
and the output layer. The three-layer-strucnired BPNN ~vith Sigmoid transfer functions
has been found by many to be adequate in solving non-linear op timization problems. In
the following sections, we use the three-layer Sigrnoidal BPNN to illustrate the
mathematical inferences of input sensitiviq for simpliciq- of representation. However,
interested readers can readily estend the derived input sensitiviq to BPNN with more
comples structures and other trans fer hmctions.
Input Sensitivity Based on Normalized Data
Based on the BPNN algorithm presented in previous sections, me can soa out
the relationships between an output variable and an input parameter to de hne the input
sensitivity of BPNN in an esact mathematical tenn.
Notations used in the followkg mathematical formulae are lis ted as belom:
Subscript p stands for a node in the Previous layer of the netntork-
Subscnpt c stands for a node in the Current layer of the nemork; C stands for the
total number of nodes in the cunenc layer.
Subscript n stands for a node in the Nest layer of the network.
LVii stands for the weight of connection between node i and node j.
S stands for the input signal to a node.
N stands for the output signal from a node.
If the curent node is an input node (in the hrst layer of the nenvork), S is the
normalized input data in range (0,l).
Previous Layer Curren t Layer Next Layer
Figure 4-2: Illustration for Node and Layer Representations
If the curent node is not an input node (in a hidden layer or an output layer),
Note thac equation (9) is the same as (1) escept thac (9) di s~gui shes one node or
processing element p fiom others n t the previous layer. The relationship betrveen the
output signal N, and the input signal Sc is defined in (2), fiom which, we have:
As shown in Figure 4-2, in the three-Iayer BPNN, node p is an input node in the
input layer, node c is a hidden node in the middle layer, and node n is an output node at
the output layer. The focus of BPNN sensitive analysis is on investigating the &SE-order
partial derivative of the output signal Erom node n (Nd over the input signal to node p
(SJ. By (8), we have S, = Np.
fiom (IO), we know,
From (1 l), we know,
So, (12) cari be expressed as:
Because more than one processing element exists in t he current Iayer (hdden
Iayer), assume, the number of processing elements at layer c is C. A general form of
input sensiuviry for BPNN is then expressed as (1 4):
Input Sensitivity based on Original Data
In daivuig (13), me assume aIl data including inputs and outputs has already
been normalized in the range (0,l). From the perspective of real applications, usually it is
convenient and straightfomard to probe the sensitivity of BPNN based on the original
or raw data instead of scaled data.
Various linear or non-linear normahation methods can be used to uansfonn
raw data- Shi (2000) retiewed the established data transformation methods for BPNN
and proposed a new one calIed ccdisabution transformation", which fits statistical
distributions to raw input data and ualizes the resdtant Cumulative Densiq Functions
(CDF) to scale inputs to [0,1]. The theoretical underpinning of such transformation is
relativdy weak as pointed out in Shi, 2000, but such transformation does complicate the
application of NN. The conclusion about the superiority of ccdistribution
transformation" scenario over the traditional linear transfomation scenmio is b v e d at
ernpiricdy based on independent esperiments on each method. Due to a number of
variable factors (such as learning rates, momennuri) and stochastic phenornena (such as
the initialization of network weights and the esistence of multiple local optima in the
searching space), the improvement of netnrork performance may not be attributable or
only p d y attributable to the input transformation methods.
The non-hear mapping capability of BPNN is mainly owhg to the non-linear
transfer functions in hidden and output PE S. Xccording to our experiments on BPNN,
a good selection of hidden layer smctures and transfer functions based on mals
generally results in improvement of BPNN's performance. Thus, we recornmend using
such robust, undistorted and simple data normalizabon methods as he a r transformation
to normalize both inputs and outputs in BPNN and satisfy the neural cornputauon
requirements. The simpliaty of BPNN will be maintained wthout sacrihcing its
fwictionality, mhich can be M e r demonstrated rom sensi t i vi ~ analysis of BPNN in
the following sections.
If we take into consideration the data normalizauon procedure, the simplest and
most comrnonly used one is a linear process as folloms:
(UB - LB)
( Sp- MINp) +LB
Np = (MAX, -MINp)
Where, I23 is the upper bound of the nomaled interval (LB,UB), and LB is the
lower bound, for sigrnoid transfer function, usudy LB = 0, and UB = i;
bLL2 is the maximum value in the data set corresponding to input node p or
parameter p;
MIN, is the minimum value in the data set conesponding to input node p or
parame ter p.
A formula similar to (15) is applied to normalize the output data, in order to
match the output range of the trmsfer hct i ons in BPNN, i.e. (0,l) for Sigrnoid uansfer
functions. If we take N, as the raw output data, there is a scale-back process involved at
the output layer, which w d cancel out the UB (1) and LB (O) in combination with the
scaling process at the input layer. So w e can arrive at a more general form of (14) based
on linear norrnalization procedures as:
JN, - MAX,-MIN,
-- . ~ W, , W, , - N, ( l - Nc , ) - Nn( l - Nn) (W
as, MAX, -MIN, i=i
This slope or partial derivative is dehned as absolute input sensitivity and
represents the expected change in output variable N,, per unit (1) change in input
parameter S,, holding the other input parameters constant. In a red-wodd problem, each
input parameter may have different unit of measure, and hence various relevant range,
which encompasses dl values from the s nde s t to the kgest used in training the model.
Simily to regression analysis, it is important for BPNN to interpolate nrithin ie range
rather than estrapolate beyond the mage in order to make sensible predictions. For one
input parameter ranging from 1 to 20000, one unit change is too s m d to be considered
mhile for another input parameter ranging from 0.1 to 0.6, one change is too big to
occur. Thus, it is more appropriate to use a relative one-unit (such aslOO/o of relevant
ranges) as the basic unit change in input parameters instead of an absolute one-unit (1).
Through such transformation, the input sensitivity is undistorted and more meaningful
in terms of comparing the effect of different input parameters upon the output variable.
The relative input sensitivity is dehed as (1 7):
relevant ranges i.e. 10% Bmes ~ ~ Y , - MINJ. Note that die input sensitiviq is
independent of the relevant ranges of input parameters and represents the amount that
output variable N, changes (either positive or negative) for a particular unit change in the
input parameter S, , i-e. the 10% of its relevant range.
BPNN vs. REGRESSION ANALYSIS
The above sensitivity analysis of BPNN is analogous to the dassic multiple
regression analysis in statistics, mhich predicts the values of response or dependent
variable based on the values of multiple e.<planatory or independent variable and c m be
dehned as (1 8):
Where B,, is an intercept representing the average value of N,, when ail the
es~lanacory variables S , are equal to zero, i =l to M, M is the total number of
a N n
explanatory variables. - is a slope for the i" esplmatory variable and its dehiniaon
as pi
is identical to the input sensitivity of BPNN as above. However, ody by esamliing the
difference between BPNN and regression analysis c m the sophisucation and supedoety
of BPNN over regression analysis be demonsaated, as discussed neat.
aNn in BPNN is An examination of Equation (16) indicates that the value of -
dependent on severd factors:
1. The interna1 structure of BPNN, i-e. the number of hidden nodes and number of
hidden layers.
2. The BPNN data set, i-e. the relevant range of each input parameter and output
variable,
3. The weight values of BPNN, i.e. the intensitg of c o ~ e c u o n among processing
elements Erom the input layer to the output layer. This is actually the result of
BPNN maining, and hence dependent on the training data set.
4. The cunent input values loaded at the input nodes. From (l), (2), and (1), it is
evident N, and N, are hincuons of the cunent input values at the input layer and
the weight values of BPNN.
We can also observe that once a BPNN is trained on a data set, the &sr three
factors (BPNN smicture, meights and training data set) are &ed, so the sensicivis. of an
input parameter over an output variable is totally determined by the fou& factor, i.e. die
cunent input values. If we treat the current input values as the coordinate values of an
input point at the BPNN input space, the dimension of nihich is equal to the number of
input parameters, me can conclude that, for a trained BPNN,
--
- F(1nput - Po int)
as,
Here, F stands for a fuaction.
Indeed, BPNN perfoms a multiple Iinea regression analysis at each individual
data point to fit a non-lineu high dimension hyperplane to the training data set. The
slope value dong each dimension, dong mirh the intercept value P,,, varies Erom data
point to data point, in contrast with being constant in regression analysis. Simply put in
tnro dimension space, BPNN is capable oE fitting a flexible cuve and ai l the observed
data points fidi on the line; while regression analysis can only approiamate a straight h e
chat strings up the data points with the mi a i mu amount of deviation based on least
square method.
Aside from above discussions, three other advantages of BPNN over regression
analysis are worrh mentioning:
(1) BPNN poses no theoretical constraints on data in contrast .Nith the assumptions of
least square regression and the required residual analysis in regression analysis
(Levine et al, 1997).
(2) BPNN supports more than one output in input-output mapping in conttast with
only one output in regression anaiysis.
(3) BPNN reLz~es the requirements of data in terms of both quantity and quality in
contrast with regression analysis. That means BPNN is capable of non-linear
mapping with only a very limited quantitg of obserc-ed data points and is tolerant of
noisy data (inaccurate or incornplete data).
Table 4-1: Data Set for Testing BPNN and Regression Analysis
Output
In order to illustrate the cornparison of BPNN and regression analysis, we
studied the input sensitivity of BPNN trained on an &cial data set with 4 inputs, 1
output and only 10 records as shomn in Table 41. The BPNN mode1 has four input
parameters, 1 output variable, and one hidden layer with three hidden nodes, wliich is
detemiined based on trials. The leaming rate is 0.8 and the momenturn is 0.3. Standard
Error of the estimate in regression analysis is a measure of variation around the fitted
line of regression and is cdculated as a measurement of accuracy to compare the
performances of tmo techniques. Standard Error is actudy a slight variant of the global
enor tenn E in BPNN as (3). After achieving satisfactory training (standard error of the
NN output is reduced to 0.00158), we calculated the partial derivative values of the
output variable over each input parameter using (16) at various input points. n i e results,
as shown in Table 4-2, indicate that for a specific input parameter, the slopc value over
the output variable varies svith the input points. In order to analyze such variation, a
Monte Carlo stmulation is performed at the BPNN input space to observe the statistics
aNn value for each input parameter. In each simulatioa w, an input point is of -
randomly generated in the BPNN input space and triggers a BPNN r ecd process. A
slope value of each input parameter over the output variable is calculated. If the number
of simulation runs is large enough, we c m assume we d traverse the entire BPNN
input space by i nt er pol a~g. A program in MS VB and Access is developed to perform
the Monte C d o simulation exi~eiiments for 1000 iterations. The resultant Probability
Density Functions (PDF) of slope values for the four input parameters are shown in
Figure 4-3 and the statis tics ;ire sumrnarized in Table 4-3.
aN,
Table 4-3: Statistics of Partial Denvative (Slope) Values: (-)
~ S P
3Nl
Table 4-2: Partial Derivative (Slope) (-) at Four Input Points at
NP
BPNN Input Space
Input Factor
Index @)
(9
1
7
-
3
4
Input Factor
Index @)
(1)
1
7
-
3
4
Point
(0.5,0.5,0.5,0.5)
(2)
-0.1594
0.8915
-0.1685
0.9974
Point
(0.9,0.9,0.9,0.9)
(4)
0.22 69
0.3267
-0.31 69
0.2878
Point
( 01, 010101)
(3)
-0.3057
0.26 1 3
-0.0398
0.4799
Maximum
(2)
0.9797
1.9502
0.4833
2.2255
Point
(0.2,0.4,0.6,0.8)
( 5)
0.1115
0.3917
0.3302
O.1G11
Average
(4)
-0.01 51
0.5769
-0.2678
0.6365
Minimum
(3)
-1 .O366
0.0042
-2.1053
0.003 1
Std. Dev.
(5)
0.3863
0.4038
0.5489
0.5064
95% Confidence
Interval
(6)
-0.0390 - 0.0089
0.5518 - 0.6019
-0.3018 - -0.2338
0.6051 - 0.6679
Figure 4-3: Distributions for Input Sensitivity
A regression analysis is conducted on the same data set of 10 observations in MS
Excel. The results are that the slope of the tirst input parameter is 0.1217, the slope of
the second input parameter is 1.0141, the dope of the third input parameter is minus
0.5925, and the slope of the fourth input parameter is 0.5509; the intercept is minus
0.03054. Note that those slope values are constants in contrast with distributions as
obtained &om BPNN. The standard error based on the outputs of regression analysis is
as high as 0.1285 compared Mth merely 0.00158 of BPNN.
In short, BPNN outperforms regression analysis by a significant mugin in our
experiments, which agrees with the previous analysis and comparisons.
The simulation results reveal disuibutions of slope data for BPNN, which take
various shapes (Fig. 3). If the actual distribution of input sensitivty to be encountered in
operations is available, cornparison of the actual distribution nrith the conesponding
Monte C d o distribution obtained from BPNN can serve as an effective means for
mode1 validation. However, in most real BPNN applications quantitative information is
unavailable to fit such actual distributions of input sensitivty due to the complesity of
the engineering or management problems being solved. This is also the reason of
choosing BPNN insiead of other conventional mathematical models in the k t place.
An experienced domah expert may also have difhculty figuring out such distributions of
input sensitiviv on a subjective basis, because the deusion process generally relies on
assessrnent of the entire input scenario and there are so many interacting factors. The
domain experts may share some cornmon hunches about the probability of increasing or
decreasing the output variable mith a certain adjusanent of an influencing factor. But the
amount of adjusment is generally very subjective depending on the lapur scenario and
persond e-xperience and temperarnent.
Therefore, instead of fitang distributions, statistical analysis of simulation resdts
involves calculatlig 5 percendes of the slope variable for each input parameter, i.e. the
lochh, 2Sh, 5oLh, 75th, and 90". The input sensitivity of di input parameten is summarized
and presented in a tornado-fie graph as ilIustrated in Figure 4-4 for the piping
fabrication labor productiviq BPNN model. The horizontal asi s represents the relative
input sensitiviy as d e t e k e d by (17), i.e. output response (negative or positive) nith a
change of 10% relevant range in an input parameter. The vertical mis is the basellie
coreesponding to no output response or zero change in output. Five short vertical bars
correspond to each input parameter representing respecuvely the five percentiles f?om
left to right, reflecting the central trend, the spread, and the shape of the obsenred slope
data dismbution from simulation. The guidelines for i nt er pr e~g the graph and
sim-dation resulcs are listed as below:
The leftmost bar ( l Oh percentile) being to the left of baseline represents that the
chance of the slope value for the corresponding inpur parameter being posiave is
above 90%, or with increase of the input value the probability for the output value to
increase is 90%.
10. How Busy
1 1- Drawing Revision
12. Priority Rushed Spools
13. Reworked Spools
14. Material Problems
15. Drawings Late
16. % night shift
17. % overtime
18. % extra
19. % apprentices
1
- - - - - - - - - - - - - - - - - - -l--.l--l - #- - - - - - - - - - - - - - - - - - - - - - - - -
1 - In Line Fitting per Ft
Non In Line per Ft
- - - - - - - - - - - - - - - - - - - - - - - - - I-i--t . . . . . . . . . . . . . . . . . . . .
- - - - - - - - - - - - - - - - - - - - - - - - - q( - i -1- - - - - - - - - - - - . - - - - - - - . -
Figure 4-4: Sensitivity Analysis of Spool Fabrication BPNN Mode1
l - per
- - - - - - - - - - - - - - - - - - - - - - - , -
4. s ~ p p o n per ~t
- - - - - - - - - - - - - - - - - - - - - - - - -
5. Flange per Ft
1 1 - 1 - -1 - - - - - - - - . - - . . . . - -
-11 -1- -1- -1- - - - - - - - - - - . - - - - - - -
- - - - - - - - - - - - - - - - - - - - - l - - 1- .l- -1- - - - - - - - - - . . - . - - - - - - - - -
6. Mlt Stn RW %
The rghmiost bar ( 9 0 ~ percentile) being to the rght of baseline represents that the
chance of the slope value for the conesponding input parameter being negative is
above 90/o, or mith increase of the input value the probability for the output value to
decrease is 90%.
The 25& and ~5~ percentiles can be explained in a simiL21: manne= as the 1 OI h and 9oKh
percentiles according to the relative positions of the conesponding bars to die
baseline in the graph.
The middle bar (50th percende) riding on the baseline represents chat the chance for
the output variable to increase or decrease is 50%.
An input parameter with a dope dismbution clustering around the basellie has less
effect on the output variable than that with a slope disuibution distant &om the
baseluie. Thus, the magnitude of input sensivty can be inferred Erom obsenvlg the
absolute values of percentiles as well.
Note that the statistical descriptors (percentiles) are based on simulation samples
rather than the entice population. However, the sample size is assumed to be large
enough (10000 runs to draw Figure 4-4) to traverse the input space of BPNN, and
the confidence interval estimates are rather tight, hence the statistical descriptors
based on the samples can represent those for the population.
The proposed sensitivity analysis method is of stochastic nature because of
independent mals for BPNN aai nl i g (such as initialization of network parameters,
hidden layer structure, and local optima) and Monte Cado process. If BPNN training
is achieved and the simulation iteration is large enough, the results for most input
parameters are stable in temis of direction and magnitude of input sensitivity, escept
for a couple of input parameters smapping sides nrith respect to the baseline fiom
trial to trial. 4 semi-optimal BPNN mode1 can be determined by selecbng the nial in
mhich input sensitivity of major input panmeters makes sense or is agreed upon by
the domain e-xpert.
In case b a t the sensitivity of one input parameter dways takes the opposite direction
in the tomado graph comparing against domain es~ert ' s esperience or common
sense, the dehi t i on and data collection procedures for the input parameter dong
mith the data itseli shouid be carefdy reexamined for s honf ds before the input
parameter is dropped out of BPNN analysis.
The sensitivty analysis of BPNN as described in the previous sections is applied
to andyze a BPNN mode1 for est hat i ng labor production rate of pipe spool fabrication
in the fabrication facility of PCL Indusuial Constructors Inc, mhich is one of the largest
and most modem pipe fabrication and module facilities in Western Canada.
Spool Fabrication Basics
A pipe spool is a portion of piping system consiseng of various piping
components, such as flanges, elboms, reducers, tees, supports, and pipe. These items are
prefabricated into d i s ~ c t assemblies thar are later assembled together as part of an
industnal plant or production skid/module. Such prefabrication is usuallp performed
under controlled shop environment iocated away fiom the actual project site, which
doms for better productivity and quality control, and hence cuts the field labor costs.
Major spool fabrication processes, such as cut, bevel, fit, weld, and handle
sections of pipe and fitthgs, tends to be labor-intensive. Productivitg dam is coUected for
63 projects completed 6-om 1995 to 1999, d k g which period the technologies and
machines for welding and cutang in the shop remain relativelv stable. The productivky
studies of spool fabrication is suitable to the unit-cost estimating method, in which labor
production rates must be independent of equipment use and v q among projects ody
because of differences in labor productivity (Parker et. al., 1984).
Due to the variation in size, wall thickness and coniguration of each individual
spool, a special unitkation scheme is utilized in the Company to quanti9 the various
work items uniformly into an abstract unit of measure called "Fabrication Unit" or
"Unit" on the basis of weld inches of standard wall thichess pipe. Quantity of non-
welding work items such as cutting, bevehg, handling pipe and fittings, installing
supports are also converted into "Units" by applying corresponding empirical factors Ui
the scheme.
Factor Identification and Data Collection
The labor hours per fabrication unit become the focus of investigation, wtJch
ranges L-orn 0.1 MH/Un.it to 0.5 h~fH/Unit in the collected historical data. The unit labor
hous fluctuate from job to job due to a number of quantitative and qualitative factors,
indudiig the complelacg of spool confguration, the material components in fabrication,
the stringency of quality control, spool dra~ving qualiq~, the amounts of night shift yid
ovenime, extra work, crem esperience etc. The environmental efects and management
factors are not considered as significant factors because of the connolled shop
environment and consistent policy and personnel of management during the period of
inves~ation. 19 input parameters are i denaed as listed in Table 4-4.
Table 4-4: Input Factors of Spool Fabrication Labor Productivity
NN Input Factor
(2)
In Line Fitting @CS) pe
Foot of Pipe in Spool
Non In Line Fitting @CS
per Foot of Pipe in Spool
Valve @CS) per Foot O
Pipe in Spool
Support @CS) per Foot O
Pipe in Spool
Flange (pcs) per Foot O
Pipe in Spool
CIIulti-S tauon Roll Welc
inches / Total Roll Welc
Inches
Repair Rate
Radiograp hy Tes,
Requirement
Non CS Units / Tota
Units
<hop Work Load
Drawing Revision Rate
Prioritg Rushed Spools
Rework Spools
lulaterial Shortage Problems
Late Drawing Issues
Night Skift hWs / Total
blHs
3ver Time bfHs / Total
bfHs
Zxtra Work MHs / Total
rvms
Qpprenticeship MHs /
rotal MHs
Data Source
(3)
Ma t e d Track.
sys.
Material Track.
S YS.
Material Track.
sys.
Material Track.
sys.
Material Track.
sys.
Weld Track. Sys.
Weld Track. Sys.
Weld Track. S y s ,
Weld Track. &
Matenal Track.
S YS.
Questionnaire
Questionnaire
Ques tiomake
Questionnaire
Quesuonnaire
Payroll Sys.
Payroll Sys.
Payroll Sys.
Payroll Sys.
Remarks
(4)
h ratio indicatkg the average length of pipe
sections in spool
h rauo b d i c a ~ g comple'ty of spooi
c o n w a tion
h ratio indicating complesity of spool
configuration
h ratio indicating complexity of spool
configuration
A ratio indicating comple-uty of spool
configuration
Multi-Station Roll Weld requires extra handling
between weld stations
An index of crew's proficiency
An indes of qualis. control sbi ngenq by specs.
Non CS component in fabrication requires estra
:are in storage, hancilhg and weldng
A 5-point rathg based on shop workload in
lnits and no. of concurrent jobs indicating how
3usy the shop was.
A 5-point ratbg based on percent of revised
4 5-point ratkg based on percent of rushed
ipooi due to client pnoety ~ i d i C a ~ g shop work
jchedules.
1 5-point rating based on percent of reworked
;pools due to drawing errors and quality defects
i 5-point 1 a ~ g on efficient); of material supply
i 5-point rating based on percent of late spool
irawhg issuance by client that impacts
kbnction
Gght S M affects labor productivity
3ver Time affects labor productivity
3 m a Work affects labor productivity
Velder qualification sys tem affects labor
xoductiviq: Apprentice vs. Journeyman
Data is collected fiom the company's various transaction systems induding labor
cost tracking systern, weld aacking system, pagroU system, material tracking systern. In
order to ease the burden of data gathering and ensure high quality of data, a histoncal
project data marehouse is custom-developed using Mcrosofi hccess and VBA to
integrate ram data Erom different transaction systems and automate the validation of raw
data and the calculauon of productivity inormauon. Because data is unavailable in
cunent transaction systems of the Company for such factors as the draming revision rate,
late drawing issuance, materid shortage problems, quanticy of remorked spools, quantity
of nished spools due to prioriy, shop work load data, a questionnaLe survey is carefully
designed and conducted nrith the support of the Company management. The key
personnel involved in the projects UicludLig shop supe~t endent s, project managers and
coordinators, QC staff, and welding foremen are interviewed to help recall some facts
and gather the needed information.
BPNN Training and Sensitivity Analysis
A total number of 70 records are compiled and used to train a BPNN mode1
nrith 19 input nodes at the input layer correspondng to 19 input parameters, 19 hidden
nodes at the middle layer, and 1 output node nt the output layer that is the unit labor
hours. The number of hidden nodes can be detemJned based on trials; BPNN learning
is found to be unsusceptible mhen the number of hidden nodes is close to the number of
input nodes. The leamng rate is 0.4, the momenturn is 0.1, and sigmoid transfer
hct i ons are used in hidden and output nodes. After satisfactory training (standard enor
of the output is 0.00143), the Monte C d o based sensitivitg anaiysis is performed on the
matured network for 10000 simulation runs. Note that Equation (17) is used to
determine the input sensitivity, which is based on the change of 10% of relevant input
range.
Several independent tnals fiom BPNN training to the sensiuvtp analysis are
conducted on the same data set. The best trial, in which the input sensitivity of most
factors folloms the same trends, as determined by ex~erienced domain e-xperts, is shomn
in Figure 4-4. An examination of Figure 4-4 reveals the relationships berneen the
influencing factors and the fabrication productivity, mhich are generahed by BPNN
through obserring histoncal project data in the past 5 vears. For exarnple, factor 1 is
about in line f i b g pieces per foot of pipe in spool, which indicates the average length
of pipe sections in spool. According to our domain esperts, in line f i dgs , such as
unions, couplings, swages, reducer etc sre used to connect pipe sections in a straight line
without nims or branches. Thus, the more in line fitting pieces in spools, the more smali
sections of pipe in spools, and the easier to handle the work. From Figure 4-4, BPNN
detemiines the chances to decrease labor hours per unit mith the increase of this ratio are
about 78% and agrees with the trend identiiied by domain esperts. Factors 2 to 5 are
four ratios indicating the complesity of spool configuration. By our domain esTerts, the
higher such ratios, the more comples the spools' configuration, and the tougher to
fabricate the spools. From Figure 4-4, the dominant trends of the four ratios are all on
the plus side, *ch matches the elrperience of our domain experts. It is also observed
fiom Figure 4-4 that factor 18 (extra morli percentage) is relatively tightly enveloped
around the baseline, which indicates that estra work is aot as dominant as other factors
130
in conmbuting to the variance in unit labor rates. The explanation cm be pardy
attributed to the fact that the arnount of estta work more directly impacts the efhciency
of administration or management than the producvity of crem on the shop Boor. Other
input factors can be interpreted and validated in a similar marner, and are not elaborated
further due to space limit.
Model Testing and Validation
In particular, the effect of matend srpe of spool Fabrication on the labor
productiv~ is tested based on the BPNN model, because material s.pe (carbon steel,
stainless steel, alumnum etc.) is a major consideration of an industtial estimator in
adjusting unit labor hours of spool fabrication. The labor production rate of non-carbon
steel fabrication is ernpirically 1.5 mes the rate of carbon steel in company's business
guidelule. 24 records La the data set wth 0% non-carbon steel component (100% carbon
steel fabrication) are selected as testkg records. In the nest step, for each testing record,
only the input parameter of non-carbon steel component is changed Erom 0% to 100%
mith other parameters intact. Those testing records are fed to the nemrork and let BPNN
recall the output, i-e. the unit labor rates for non-carbon steel fabrication. The output
Erom BPNN is compared against the original output of each record, i.e. the unit labor
rate for carbon steel fabrication. Based on the test results in Figure 4-5, BPNN increases
the unit labor hours on 75% of the records; the amount of decrease for 5 records, i.e.
No. 1, 2, 5, 6, 9, is rektively small compaEng with the amount of increase for othee. If
the sample size is luge enough, the percentage should corne close to about 9O0/0, as
obsemed bom Figure 4-4 for Factor 9. On average, the ratio of non-carbon steel labor
Test NN Sensitiviy By Changing Material Component from 100% CS to 100% Non-CS: 75%
Records increase, Avg. Ratio 1.38
1 + Actual(100% CS. 0% Non-CS) a NN Output (0% CS. 100% Non-CS) 1
O 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25
Rec. No-
Figure 4-5: Testing Sensitivity of BPNN to Material Type
rate over carbon steel labor rate is 1.3, which is close to 1.5 irr the gudeline.
Note that the guideline gives an average number (1.5) in consideration of
matenal type only, mhi Ie BPNN is able to figure different numbers for different
scenarios taking into account 19 relevant factors. In short, a BPNN-based decision
suppoa tool d be more sophisticated and intelligent than the uaditiond business
guideline.
The model validation approach of BPNN based on the proposed sensitivity
analysis is superior to the conventional validation approach of testing the mature
nenvork with an independent data set, in that such sensitivity analysis enables the
modeler to understand the rationale of BPNNYs reasoning and have a pre-knowledge
about the effectiveness of model implementation in a probabilistic fashion.
The insight into the BPNN model gained h-om the proposed sensitivity anaiysis
method gives the user more confidence in the BPNN's prediction, hence faciLitates the
implementation of BPNN-based decision support tools. The success of our indusmd
application in estmating labor productivity of spool fabrication esceeded our initial
expectations. Not only does this new method prove to be effective in addressing
problem domains in which BPNN has been applied, but also it potentially malces BPNN
app ealing to new engineering or business applications.
Carrol, R.J. and Ruppert, D. (1 988). TranJfomntion alrd EVe&hting in Regssion,
Chapman & Hd , New York, NY.
D har, V. and Stein, R. (2 997). Inte/hgent De~ikon SIlpport Sy~tems: The S~i ei m of
Kplowlec&e IVork. Upper Saddle River, Prentice-Hall, Inc., New Jersey.
Flood, I., and Kartarn, N. (1994). "Neural networks in civil engineering: systems
and applications ." 3. Const~. Engrg. And~\/rgm&., ASCE, 124(1), 18-33.
Knomles, P. (1 997). Pr*icting L b o r Pmdztctiui. Using Nermd Ne~works. Mas ters of
Science Thesis, University of Alberta, Edmonton, AB.
Levine, D. My Berenson, M. L., and S tephan, D. (1998). Statihir )r Ah~zugcr~
zmkgL V I ~ ~ J - O ~ EXCEL, Prentice-Hall, Inc., Upper Saddle Rive, New Jersey.
Li, Y-, Shen, L. Y., and Love, P.E.D. (1999). ''ANN-based mark-up estimation
syscem nrith self-esqhnatory capabilines." JomvaI of Corirtn~ctian Engizeeting U I I ~
fVImagenzent, ASCE, 125(3), 155-189.
bfukherjee, A., and Deshpande, J.M. (1995), "hlodeling initial design process
using yftcial neural nenvorks", joz~maI Conputing in CIMI Engimenig, ASCE, 9(3), 1 9 4
200.
Murtaza, M.B., and Fisher, D.J. (1993), 'Weuromes: Neural Nemork Systern for
EvIodular Construction Decision Making'', j ot ~r~zd Co7rpzlcing NI Civil Eemering, AS CE,
8(2), 221-233.
Shi, J. (2000). "Redung prediction error by transfomiing input data for neural
networlis", Jouniai C o m p u ~ g in Civil Engineering, ASCE, 14(2), 109-1 15.
Sinha, S. K. and Mc I b , RA. (2000). "Artifidai neural network for measuPng
organizational effectiveness.", Journal C o m p u ~ g in Civit Engineering, ASCE, 11(1), 9-
14.
Parker, AD., Barrie, D. S., and Snyder, R. M. (1984), Planning and Es t i r n a ~g
Heavy Construction, McGraw-Hdl, Inc., New York, NY-Portas, J., and AbouRizk, S.M.
(1997). Neural Netsvork Modd For EstunaMg Construction Productivity. J of Corn-t~
Engrg. dN Mgmt., ASCE, 123(4), 399-410.
Refenes, AN., Zaprnis, AD., Connor, J.T. and Bunn, D.W. (1995). "Neural
Netsvorks in Invesment bIanagement". IfzteZhgerzt Systestns for Finance md Bz~s~Iz~J'J', S.
Goonadake, and P. Treleaven, eds., John Wiley & Sons Ltd, , Chichester, EngIand,
177-209.
Sayed, T., and Razavi, A. (1999). "Cornparison of Neurd and Conventional
hpproaches to Mode Choice Analysis" Joiird oJ Compz~ling in Ci d Enginehg, ASCE,
14(1), 23-30.
Widrnan, L.E., and Loparo, KA (1989). "i Mal intelligence, Simulation, and
modeling: a critical survey", Art@&/ inteihgence, nimztiatio~ and modeLing. L.E.Widman, I<.A
Loparo, and N. R Nielsen, eds., John Wilev & Sons Ltd, New York, NY, 1-45.
Chapter 5: Conclusion and Recomrnendation
In conjunction with a major indusuial contractor of Canada, the thesis research
conducted case studies on the theoretical basis and practical considerations for
measuring and analyzing labor productivity in i n d u s d consmction. Two important
activities of process piping were investigated: pipe installation in the field and spooI
fabrication in the fabrication shop. The p r i mq objective of research i developing
ANN-based eshat i ng tools to offer estirnators valuable information about labor
productivity in bidding new jobs, because estimating labor productivity is one of the
most difficult aspects of preparing an estimate, or a control budget based on the estimate
for labor-intensive activities in industrial construction. ArtXcial neural netsvorks are
capable of sort kg out hidden patterns and e s l x a c ~ g predictive information from
cornplex data sets, and nrere proven to be effective in both uncertanty analysis and
sensitivity analysis of coosttuction labor productivity in the research. The thesis research
has addressed: (2) how to quanti* labor productivity in industrial construction fiom a
contractor's point of view; (2) how to measure acmd labor productivity in industrial
construction based upon on-site control practices; and (3) how to ualize M c i d Neural
analyze the labor production rates and the
sensitivity of identifed influenckg factors.
Productivity Studies and Data Collection
The thesis research reviewed current estimating practices as applied to the
involved Company and generalized special methods utiIized in practice for the
quantification and measmement of labor productiviq in industrial construction. The
input factors that cause the varkbility in the productivity for studied activities were
identified through literature review and consultation with es~enenced domain experts at
the involved Company. With the support of the company's management, two data
nrarehouses were custom-developed for field pipe installation and shop spool fabrication
respectkely to iritegrate the corporate management systems of es t i ma~g, production
resources pIanning, quality control, and labor cost control. It should be mentioned that
questionnaire surveys were carefdy designed and conducted to collect some qualitative
and desuptive information that is not obtainable Ecom the company's reportkg and
a c c o u n ~ g systems. Eqerienced supe~t endent s, project managers and eschators of
the involved Company mere interviewed to help recall some facts and gather the needcd
information. The data warehouses provide solid platforni of integrated h i ~ t ~ a d data
from which to validate novel ANN models and develop ANN-based tools for
productivity analysis.
Probabilistic Neural Network Modeling
The thesis research derived a probabilistic neural network classil5cation model
cded the Probability Inference Neural Network (PINN), mhich is based on the same
concepts as those of the Learning Vector Quantkation (LVQ) method combined with a
probabilistic approach. The PINN model was intended to overcome limitations of other
137
neural netmork models and mas developed for predicting labor production rates for
indusmal consmcon. The thesis presented and explained the topology and algorithm
of the PINN rnodel in details. Portable computer sofnvare was developed to impiement
the aWiing, t e s ~ g and recall for P m . PINN was tested on real historicd productivity
data at the involved Company to analyze the degree-of-difficuly factor of field pipe
installation productivity and compared to the classic feed fonvard back propagation
neural netmork model; ths showed marked improvement in performance and accuracy.
The PLNN model creates a meaningful representation of a comples, real-life situation in
the problem domain and in general is effective in dealing with high dimensional input-
output mapping with multiple influential factors in a probabilistic approach. The
application of the PINN model Ln industrial labor production rate e s t k a ~ g gives an
estirnator a better understanding of the project information available and the possible
outcomes that could occur. Because the response of PINN is in the form of a
probability density hc t i on (dismbuuon) at the output range, an estirnator WU be able
to dede on the degree-of-difficulty factor for a future scenaro by combining the
PINN's recommendation with personal judgment.
Sensitivity Analysis of Back Propagation Neural Networks
Validation of a NN model has thus far relied upon measuring accuracy of the
calibrated netsvork to an independent testing data set that are hidden Erom the neural
nenvork in leaming. A NN model's sensitivity to changes in its parameters is generally
probed by t e s ~ g the response of a mature necwork on various input scenarios. The
thesis research also investigated the classic back propagation NN algorithm to study the
effect of each input parameter or influencing variable upon the predicted output variable.
The input sensitiviv of back propagation NN is defned in esact mathematical terms in
Light of both normalized data and ram data. The diffe~ence betmeen back propagation
NN and regression analysis of statistics is discujsed and the sophistication and
superioriv of back propagation NN over regression analysis is h t h e r demonstrated in a
case study based on a smail data set In addition, statisticai analysis of input sensitivity
based on Monte Carlo sirnulauon enables the modeler to understand the rationale of
back propagation NN reasoning and have pre-knomledge about the effectiveness of
model implernentation in a probabilistic hshion. The sensitivity analysis of back
propagation NN was successfully applied to analyze the labor production rate of pipe
spool fabrication at the involved Company. Important aspects of the application
lnduding problem defkition, factor identification, data collection, and model testing
based on real data were discussed and presented in the thesis. The model validation
approach of back propagation NN based on the proposed sensitivity andysis is superior
to the conventional validation approach, in which the mature nemrork is tested with an
independent data set and the modei's sensitivity is probed through obserwig the output
with respect to changes in input based on a lirnited nurnber of scenzuios. The insight into
the back propagation NN model galied from the proposed sensitivicy analysis method
gives the user more confidence in the back propagation NN's prediction, hence
facilitates the implementation of back propagation NN -based desion support tools.
Not only does this nem method prove to be effective in addressing problem domains in
which back propagation NN has been applied, but also it potentidy makes back
propagation NN appealing to new engineering or business applications.
Conclusion
The problems addressed in the thesis research were idendied through
i nvest i ga~g the m e n t estimating pracuces in industly and understanding the real
concems of indusw professionals. Emerging computer modeling techniques such as
data marehouses and ANN mere researched Erom an academic perspective and
implemented in industry to meet mith the challenges. The proposed novel ANN models
and developed decision support tools nrere validated using real data from indus- and
s uc c e s s ~y applied to assist estimators in deading on labor production rates for new
jobs. The esperiences and lessons learned h m the successful, productive and mutudy
beneficial collaboration betnreen academia and industry throughout the thesis research
dl potentidy serve as a mode1 to guide other university-industry joint research projects
in the future.
There are a number of issues that need to be addressed in greater detail in the
fu tue,
Quantification of Textual / Descriptive data
Three input data types are used to define NN input factors in the thesis, i.e.
"Rad' , "Rank", and "Binary". "Raw" is used smply for quantitative input factors, like
general expense ratios, winter construction percentages, or quantities of work. "Rank" is
used to conven subjective factors, like crew ability rangs, into numeric format. And
"Binary" is used co group texmal or descriptive factors into numenc fonnats like matenal
1.10
type and project defuition. It should be noted XI input factor of the "Ram" or "Rank"
tppe corresponds to one input node at the input layer, while an input factor of the
"Binary" type corresponds to a number of input nodes depending on the number of
groups for the factor. %Binary" data - pe satisfies the cornputing requirements of neural
netmorks for converting textual or descriptive data, however, some disadi-antages
associated mith "Binaxy" data may affect the performance and sensitivity analysis of
neural networks.
First, increased dimension of NN input space caused by "Binary" data type
increases the complesity of network smicture and the quantity of netrvork parameters.
Based on experimentations and observations, the PINN mode1 is not very susceptible to
the increase of the NN input space dimension, homever back propagation NN does
suffer in terms of leaming t h e and generalization ability Nith the increase of input
dmensionality. The generalization ability is not guatanteed to irnprove, but chances are
very high that the learning me wilI increase considerably.
Secondly, the input sensitivity of back propagation neural nenvotlcs is de6ned for
each NN input node. A change for an input factor of '%inary" data type entails changes
in more than one input nodes of NN. Thus the input sensiuvity for an input Factor of
"Binary" data type must take into account the combination effect of involved input
nodes. What input nodes are involved depends on hom the change is made. For
esample, suppose four different material w e s are considered, conespondmg to four
NN input nodes, a change kom type 1 (1000) to type 2 (0100) trggers changes in the
frst and second NN input nodes; while a change Lom type 1 (1000) to type 3 (0010)
%ers changes in the hrst and thitd NN input nodes. Note tliat the input sensitivity of
back propagation hW for one input node is not a constant value but a distribution, such
combination effect makes it difficult to esp1,a.h the sensitivity of an "Binarp" type factor.
Fortunately, there is no such "Binary" type factor in spool fabrication
productiviq analysi where the sensitivity andysis was tested. For the field pipe
installation productivtg analysis, an esperment was conducted to treat such "Binary"
factors such as material type and project type as "Rank" factors. Various groups in each
factor were ranked on a 5-point scale by their relative difficulty based on the judgment of
domain expert such that a unit increase in the corresponding NN input node codd
represent the increase of degree-of-difhnilty factors. The results of the esperknent are
satishctoiry and the input sensitivity follows the correct direction for most factors.
However, the dran-back of such a heuristic method is that sometirnes even the domain
esTerts found it hard or impossible to weight the relative difficulty and rank each group
in a factor in a sensible way.
In short, more sophisticated methods such as &zzy set theory may be researched
and introduced into NN to convea textual or descriptive factors into numeSc formats.
Optimization of NN Structure
NN structure m d y concerns with the middle layers, for instance, the number
of hidden layers and number of hidden nodes in each for a BP NN; the number of
processing elements assigned to each output zone at Kohonen layer and the setup of
output zones for a PINN model. The determination of NN sarucme relies heavily on a
mal-and-enor based process, in which cornparhg the NN's outputs nrith actual outputs
on an independent tesMg data set serves as a yardstick for j us t i wg the structure. Such
optunization of NN structure tends to be hampered due to factors such as the stochastic
processes involved in NN leaming, the existence of multiple local optima in search of
the NN intemal parameters, noise mithin leamng data set, values of leaming rates and
me s of data transfer functioas
One appealing solution is to obtain the acnial distribution of input sensitioiv for
key input factors (if not dl) to be encouncered in operations. Matching the
conesponding Monte Cario distributions obtained fiom BP hW to the actual
distributions can seme as an effective means for optimization of the NN structure, in
addition to validation of the NN model as discussed in Chapter 4. Hence, gathering
quntitative information to fit actual disnibutions of input sensitivity could be included
as part of data collection for NN applications in the future if both data and resources are
available.
Sensitivity Analysis of PINN Mode1
In the thesis, the PINN model's sensitivity to changes in its parameters is sll
probed by tesMg the response of a manire n e ~ o r k on various input scenarios. One
approach that have been tried is to take advantage of the sensitivity- analysis method for
BP NN as proposed in the thesis to infer the input sensitivitg for a PINN model under
the following conditions:
1. The BP NN and PINN are trained with the same leaming data set and tested with
the same t e s kg data set, and both models are satisfactorily trauied;
2. And the point-value predictions of the BP NN and those of the PINN mode1 for the
testing data set are very close.
Thus, it can be assurned that the tnro models would "think alike" and have
common input sensitivity for a pa15cular input factor.
Table 5-1 shows the resulrs of five teshng records based on the training data for
spool fabrication pr oduct i ~t y after PXO models have satisfied the abore conditions.
Table 5-1: P NN vs. BP NN
It is noted that the &st condition is not hard to satisfy, but the second condition
rnay oot be readily met The BP NN and PINN may require to be trauied repetitively
using different smictures and leaming parameten in order to satisfy both conditions.
BPNN
0.239
O. 156
0.221
0.1 57
0.286
In the hure, it would be perfect if an independent approach could be found to
explain the input sensitivity of the PINN mode1 analyticaily.
PI NN
(Mode)
0.205
0.145
0.235
O. 1 45
0.265
The applications of AI\TN in es t i t na~g labor productivity of indusnial
consmiction prove that -ANN is effective in addressing the complesity and requiremenrs
in the problem domain. It is hoped that the contributions made in the thesis research
m d l make MWappealing to more engineering or business applications in the hture.
&PEND= A: USER'S MANUAL FOR PINN TRAINER
Probabilistic Inference Neural Nenvork (PINN) trainer is a genenc neural
nenvorli training and testkg program developed based on a new NN scheme as
proposed in Chapter 3 of the thesis.
Step 1. Prepare data and import data into trainer
The last column in a data table must be named as "Status", which hgs the
training/testing statu for each record. Status 1 stands for a training record, and Status 2
for a testing record, and Status O for an ignored record. The nest-to-last column in a data
table must be named as "Output", storing Actual Output Values of the target nsLy
variables such as actual production rates. Ali the rernaining columns in a data table d
be the input factors and no requirements are imposed on the narnes and relative order of
columns. The trainer nrill automatically count the number of total inputs and read in
data.
The prepared data table for PINN must be imported to the database file
'T1INN.mdb7'.
Step 2. Give a unique identifier key for a new training-testing trial
A unique identifier key is used to dis~guis1-i each training-testing scenario or
trial, whch is de6ned by the data source table to use, training / testing records within
the data, the setup of the output zones, trnining parameters, and number of training
iterations. hTaming convention requires no space in the key and numbers and short tests
are dowed such as cc081899aWeld".
For a previous trial, user can pick out the identifier key hom the drop-domn List.
Next, user may click the buttons on the switchboard to check trzining results
Figure A-1: Select an identifier key of one previous trial
("Check-Train" button), check t e s ~ g result ("Check-Test" button), and check global
report about training and testing ("Global Report" button) as shown in Figure A-1.
For a new &al, user needs to in a new idenfier key in the Identifier Key bos
first. And then click "T~ain-Test" button on the switchboard to activate the program.
Step 3 Select data table, edit training / testing status and setup output
zones
- -
Figure A-2: User select~ data table
User selects one data table fiom the &op-dom list of "Data Table Name" &sst
(Figure A-?), and clicks the "Edit Stanis: Trah or Test" button to Bags training and
r e s ~ g records (Figure A-3). The trainer d read in data and display the maximum and
minimum of the output values for user to setup the output zones.
Figure A-3: Flag status of records
Setting up the output zones properly- and adequately is crucial to PlNN's
performance. Too wde zone widths won't be adequate to help user make deasion, while
too nanom zone midths d probably sacrifice the accuracy of PINN's prediction. The
ollowing two issues should be taken into account:
-
Pression requirement of user. Here is a heuristic fonnula to appro-uuriate zone
wdth:
Zone Width= 0.4*0utpurRange*Ac~~1:aqThreshold
-
Distribution of actual output values over the output zones. A uniform distribution
generally yields better results.
Two approaches are avdable to set up the output zones:
1. User speuhes the number of output zones only. The trainer will evenly divide
the actual output range into the number of output zones as user has specified, and
automaticdly determine lower bound, upper bound, and mid value for each output zone.
2. User speuhes the lower and upper bounds of the output range and the width
of each zone as well. The trainer d start from the lower bound of output range and
detemine the boundaries of each output zone, und the upper bound of output value
range is esceeded.
Step 4. SpeciQ structure and learning parameters for PINN
Following the setup of output zones, user shifts focus to the next page to speS
a number of structure and leaming parameters for PINN iccluding the scale max and
151
min, the number of processing elements per output zone, the attraction rate and
repulsion rate and conscience hctor for leanllng, the smoothing factor for kemel
hct i on, and the accuracy threshold for performance rneasurement Figure A-4 shows
the program screen, user may take the default values or set new values for those
parameters. Refer to the technical paper and online help for detailed explanatioas of
Figure A-4: Setup structure and learning parameters for PINN
those parameters-
Step 5 Speciw training itetations and train-test PINN
Folloming settng the PLNN parameters, user shifts to the nest page to spec*
Figure A-5: Specify training iterations and train-test PINN
the training irerations by enterkg the "No. of Train Epochs", as shown in Figure A-5.
Step 6 Investigate whether the PINN mode1 has been successfully
trained
FoUowing training and testkg, user clicks the "Check-Train" button and the
"Check-Test" button on the switch board to view the results for maining data and
'1 Training Data: Actua VS N N -Rediction
f
Training Records Probabiliy Density Graph
10 .
Mode Avg E I ~ 1 0 1 5 1
I
Figure A-6: Check training results
testkg data respectively and investigate whether the neural netmork has been
successEully trained-
The majority of training records should indicate a centralized trend and the
dismbution generated by PINN f d s in the correct output zones for their actual output
values, as shown in Figure A-6. Othemvse, user should repeat fiom step 2 and perform
another trial using different network structure and learning parameters, or increase
leamng iterations. Note that on the *ht side of Figure A-6, the accuracy scatistics of
point predictions are also induded for user to judge the model' performance or maturity.
User observes the tesng data in a very sirnilar manner. Testing data speaks more in
iaining Datas Actual VS N N Prediction
. .
Training Records Probability Oensiy Graph
t o
Figure A-7: Detected noise in training data
j udpg the nenvotk's performance thaa training data because the traines bas not seen
the t e s ~ g records in the learning process.
In case that after a nurnber of different train-test mal s, for a pmxicular training
record, the PINN indicates a very ar-off point prediction value (mode) comparing wth
the actual output value, or die PINN demonstrates a very dispersed dismbution as
shown in Figure A-7, such a record is likely tu be noise in the data and the data of the
record should be esarnined for errors.
Step 7 RecaU based on a trained PINN mode1
Once satisfactory results are obtained for both training and testing data, the
Figure A-8: Global report for a train-test trial
PINN mode1 is dedared to be crained and ready for developing a recd program. User
clicks the "Global Report" button to review the infomiation about this trial as shown in
Figure A-8.
If user does not want to keep a trial any more, select the identifier key for the
trial and click the "Delete Key" button on the switchboard to delete al I the data related
to the trial.
An on-Iine help is dso developed givng details about how to use the crainer
output
M ode
m
Weiphted
Average
Probabiiity
Distribution
Figure A-9: PINN Trainer on4ne help
program dong mith some technical descriptions about the PINN mode1 as shown in
Figure A-9. By selecting a category, related topics are fltered out; user chooses one topic
of interes& the description d be automatically displayed-
APPENDIX B: USERS' MANUAL FOR FAISMASTER
FabMaster is a histoncal project data warehousing system customized for the
Fabrication Facilities of PCL Industrial Constructors, Inc. It is a n automated data
processing tool to estract raw data Gom Fabrication Resources Planning System, Wcld
Tracking System, Labor Cost Contxol System, and conven raw data into aggregate
quantiq data at spool level and various ratios of productivity, quality control and
configuration complexity at cost-center level. A cost center, de he d by project number,
material type and size range of spools, is the level of detail that actud labor hours were
tracked in the corporate labor cost control systern. Two levels of compilation are
involved in FabMaster to convert raw data into item-coded productivity information, i.e.
the spool level and the cost-center level. The codiag systems for material type and size
range of spools used in the Company and Fabbhstrr are shown in Table B-1 and B-2.
The item codes for spool level compilaaon arc shown in Table B-3.
Table B-1: Size Range Codes
ID
1
7
-
G
16
30
O
Description
< 2"
- 7 -4"
6-14"
16-34"
30-48"
To td
Table B-2: Material Type Codes
Table B-3: Item Codes for Spool Level Data Compilation
ID Description
Iescrip tion
JO. of pipe pieces
'ootage of pipe pieces
liameter Inch Fr of pipe pieces
'ons of pipe pieces
cio. of pipe pieces longer hanG ft
'ootage of pipe pieces longer than G ft
>iaIriT;t of pipe pieces longer than G ft
'ons of pipe pieces longer than G ft
JO. of Stub Ends
JO. of Branches (Olets)
JO. of Caps/Plugs
\io. of EIbows
JO. of Swages
\To. of Blind Flanges
<o. of Dummy Legs
JO. of S'&'/TH Couphgs
JO. of Lap Joint Flanges
h. of hnchors/Shoes/Slider Supports
\To. of Nipples
go. of O s c e F h g e
\io. of Reducers
\30. of Slip-on/SW/TH FLznge
<o. of Tees
\To. of Unions
L'o. of Vdves
<o. of Weld Neck Fhge s
\To. of Laterds
\To. of Pvlisc Items
No. of Flanges
No. of In-Line Fittings
No. of Out-Line Fitrings
No. of Supports
No. of Design \Velds
Diamecer Inch of Design Welds
Equivalent Diameter Inch of Design Welds
1 ItemCode 1 Description
Volume of Design \Vdds
No- of BW Design \Velds
Diameter Inch of BW Design \Velds
Equivdent Diameter Inch of B\V Design Welds
\'olume of BW Design \Velds
No. of S\V Design \Velds
Diameter Inch of S\V Design \Velds
Equivalent Diameter Inch ofS\V Design \Velds
Volume of SW Design \Velds
No. of OL Design \VeIds
Diameter Inch of OL Design \Velds
Equivalent Diameter Inch of OL Design \Velds
IToIume of OL Design \Velds
No. of Pressure Amchmena in Design Welds
Diameter lnch of Pressure Attachments in Design Welds
Equivdent Diameter Inch of Pressure Attachments in Design \Velds
CToIurne of Pressure Artachments in Design LVelds
No. of Non Pressure Attachments in Design \Velds
Diameter Inch of Non Pressure ,-mchments in Design \Velds
Equivalent Diameter Inch of No Pressure ,.\mchrnents in Design \Velds
Volume of No Pressure Attachments in Design \Velds
No. of Positon Welds in Design \Velds
Diameter Inch of Positon \Velds in Design \Velds
Equivaient Diameter Inch of Positon \Velds in Design \Velds
\'olume of Positon \Velds in Design \Velds
No. of hidu-station Roii Welds in Design \Velds
Diameter lnch of MuIti-staaon Roll \Velds in Design \Velds
Equivalent Diameter Inch of hlulti-station Roll \Velds in Design \Velds
Volume of hiulti-station Roii \Velds in Design \cVelds
Volume ofTig Process in Design welds
Volume of &Lig Process in Design \Velds
Volume of FCAW Process ia Design \Velds
Volume of Stick Process in Design \Velds
VoIume of SubArc Process in Design Welds
Volume of Rotoweld Process in Desiga \Velds
No. of Reworked Welds
3iameter Inch of Remorked \Velds
3quivaient Diameter Inch of Reworked LVelds
Description
Volume of Reworked Welds
No. of Cut Sheet Revisions
Spool LVeight in Tons
RT percenc per Spec
MT percent per Spec
M' percent per Spec
Phi1 percent per Spec
P K " 1 / O per Spec
FT percent per Spec
BHN
\T percent per Spec
UT percenc per Spec
No. of Accepted Weids
\Veld Units in SpooI
Weight of Non-pipe in Spool
No of Spools in a matenal-size group
Step 1 Download Raw Data from Corporate Management Systems into FabMaster
Si - raw data tables of a project required for FabMaster to process are dkectly
downloaded Erom the corporate databases of various management systems in electronic
formats, namely RD-BranchPlant-SP (Spool) and RDBranchPlant-MT (Pieces) Grom J. D .
Edsvards matcrial resources planning sys tem, plus RDBranchPlant-DG (Dra\ving),
RD-BranchPlant-SC (Spec) , RDBranc hPlant-WD (Weld De tails) and
RDBranchPlant-WW Weld Welders) Gom WeldTrack quality control sys tem, as s homn in
the program flow chart of FabMas ter (Figure B-1).
User enters the project number and clicks the "Lnk RD Tables" to import raw data
tables. If al! needed tables are in phce, user kicks off the program Bow by hitting die
'Trocess if' button.
JDEdwards RD-B ranchPlant-SP (SpooI)
RDBranchPht-MT (Pieces)
based on 23
Pass aU checks-
1st level cornpihaon &
deril-ed qcy cdcuhtion
2nd level compilation
based on Spool Item
Code Structure
3rd level compilation
based on Job-Material-
Cost Code
Structure
Figure B-1: Program Flow Chart of FabMaster
Step 2. Data Validation Based on Pre-defined Rules
The foIlowing d e s were programmed in FabMaster to detect abnormalities and
prompt user to scrub raw data of errors prior to processing.
1. No blank is dowed in ptem Number] in RD-BranchPlanthfT.
2. No blank is allowed in p e l d Process] in RD-BranchPlant-WD.
3. PVeld Process] in RDBranchPlant-\JCTD must be a known process-combination in table
LP~WeldProcessCombo.
4. No blank is aliowed in goint Type] in RDBranchPlant-W.
5. (loint Type] in RDBranchPlantWD must be a known type in table
LP-Join tType-PANP .
6. No blank or O is dowed in p e l d Size] in RDBranchPlant-W.
7. [Weld Size] in RD-BranchPlant-WD mu t be a known s i x in table LP-PipeOD.
8. No blank or O is allowed in p e l d Thickness] in RD-BranchPlant-\W.
9. No redundant spool is dowed to exist in RDBranchPlant-SP.
10. No blank or O is allowed in [Spool Weight] in RDBranchPlant-SP.
11. No blank or O is allowed in [Spool Units] in RDBranchPlantSP.
12. No blank or O is allowed in [Size Group] in RD-BrmchPlant-SP.
13. [Size Group] in RD-BranchPht-SP must be a h o wn one in table tP-SizeGroup.
14. No blank or O is allowed in pl at er d Group] in RDBranchPlant-SP.
15. platerial Group] in RDBranchPlant-SP must be a h o wn one in table
LP-LateridGsoup .
16. [Spool Nurnber] in RDBranchPlant-SP must have a corresponding record in
RDBranchPl ant DG.
17. No blank is allowed in [Spec] in RDBranchPlant-SC.
18. [Spec] in RDBranchPlant-DG rnust have a corresponding reference in
RDBranchPlantSC.
19. [Spool Number] in RDBranchPlant-SP must have a corresponding record in
R D B ranchPIant-MT.
20. PVeld Thichness] in RDBranchPlant-Cm like 3000/6000 must be able to be converted
to inches by tinding a reference Li LP-PipeThickness.
21. p e l d Thichness] in RD-BranchPlant_WD must not be abnormaily large (greater than
5").
22. PVeld Size] in RDBranchPlant-WD must be able to be converted to Equivalent
Diameter Inches by hnding a reference in LP-PipeThickncss based on size and
thickness.
23. For an Olet type weld, a reference in LP-OletDkn based on weId size must be found.
. .-
- , - .
. . . ,
- - _ r
- .
nter Branch Piant Nor . - 11709286
~hadc Q W ~ ~ M ~ T T , ~
-
- - -
- - - -
- -
-
. - - - . -- - - - - - - -- -. .. -
. . . , . .
Figure B-2: Main User Interface of FabMaster
FabMaster d hint user about the detected problem records, violated d e s and
solutions (either update cross-reference tables or correct ram data tables) in its progress
monitor mindows as showo in Figure B-2. To resume the program flow after hxing
problems, user needs to hit the "process it" button again to continue the process from
where it paused 1s t time.
Step 3 Unithe a project and perform spool Ievel compilation
Once dl the checks on raw data are passed, FabMaster runs its built-n prograrns to
automatically unitize a project into "Fabrication units", compute the quantities of various
work items in pre-specihed units of measure, and store the resdts into four temporaq
268
tables, namely RD-Spool, mePiece, RDomection, and RD-Weld. The hrt level
compilation is conducted based on those temporary tables to generate the item-coded
aggegate data for each spool and appended to spool s umr nq table "SMJternQty-SP".
Step 4. Compile data at cost centet level and compute ratios
The cost-center level data compilation and ratio computation follows the spool lerel
data processing and the resdts d be appended to tmo surnmarg tables "SM-ItemQty-CC"
for aggregate quantities, and "SM-ItemQy-RT" for final rsrios. Table B-4 shows samples of
Sh.I_ItemQty-RT" based on one small project nrith one materid type and one size range
only. A big project often has more than one material types and size ranges of spools.
FabMaster d generate valid ratios only for a speufic project and automaticdy handle the
roll-up of ratios to various total levels.
Table B-4: Sample of FabMaster Outputs
Matctial
To t d
To ta1
To td
To tal
Total
Tom1
Total
Total
Total
Total
Total
Total
T o d
T o d
Total
Total
To ta1
Total
Total
Total
To tal
Total
Total
Total
Total
Total
Total
T o d
Total
T o d
Totd
Total
To t d
Total
-
Size
T o d
Total
Total
Total
Total
Total
Total
Total
Total
To ta1
Total
Total
Total
Total
Total
Total
Total
Total
Total
Total
Total
Total
T o d
Total
Total
Total
To ta1
Total
Total
Total
T o d
To ta1
To ta1
To ta1
-
Ratio Description
Total hfanHours / DiaIn*Ft
Total LLfanHours / Equiv.DiaInrFt
Total h l d o u r s / Equiv-DiaIn
Totd Md o u r s / Volume
Total PvfanHours / Unit
No. of Pipe Pieces / Footage
No. of Pipe Pieces / DhInFt
No. of Pipe Pieces / Ton
No. of Pipe Pieces / Unit
No- of Pipe Pieces Over 3 ft / Footage
No. of Pipe Pieces Over 3 fr / DiaInFt
No. of Pipe Pieces Ov e 3 ft / Ton
No. of Pipe Pieces over 3 ft / Unit
No. of m g e s / Footage
No. of Flanges / DhlnFt
No. of Fianges / Ton
No. of k n g e s / Unit
No. of In-Liae Fittings / Footage
No. of In-liae Fittings / DiaInJ3t
No. of In-Line Finings / Ton
Yo. of In-line Filtings / Unit
No. of Non-In-Line Fittings / Foomge
No. of Non-In-Line Fimings / DiaInFt
'JO. of Non-In-Lide Fittbgs / Ton
go. of Non-In-Line FilMgs / Unit
'JO. of Valves / Foomge
So. of Valves / DiaInFt
'To. of Ir&-es / Ton
%o.ofVdves / Unit
%o. of Supports / Footage
go. of Supports / DialnFc
30. of Supports / Ton
Vo. of Supports / Unit
No. of PrLisc. / Footage
Ratio
Material
Total
T o d
To ta1
Total
Total
To td
Total
' Fod
Total
Total
T o d
Total
Total
Total
Total
Total
Total
Total
Total
Total
T o d
Total
T o d
Total
Total
Total
Total
Total
To ta1
To ta1
Total
Total
Total
Total
Total
Total
To ta1
Total
To ta1
Size
Total
Total
Total
Total
Total
Total
Total
Total
Total
Total
T o d
Total
Total
T o d
To ta1
Total
Total
Total
Total
Total
Total
Total
Total
Total
Total
Total
To ta1
Toul
Total
T o d
Total
Total
Total
Total
T o d
T o d
Toral
Total
Total
Ratio Description
No. of hkc. / Di dnFt
No. of LLLISc. / Ton
No-ofhlisc. / Unit
Pipe LVeight / Spool Weight
Non-Pipe \Veight / Spool LVeight
No. ofConnections (Design LVeIds) / Footage
No. of Co ~ e c a o n s (Design \Velds) / DiaInFt
No, of Comecuons (Design \Velds) / Ton
No. of Co ~ e c u o n s (Design \Velds) / Unit
No. of Multi-Station ROLL\Velds / Footage
No. of hldti-Station Roll \Velds / DiaInFt
No. of Mda-Station Roll CVelds / Ton
No. of Mda-Station Roll \Velds / Unit
No. of Repaired Welds / Footage
No- of Repriired \Velds / DiaInFt
No- of Repaked \Velds / Ton
No. of Repair \Velds / Unit
B\V DiaIn / Design \Vdd DiaIn
B\V Equiv-Dan / Design \Veld Equiv-DiaIn
BQ' Vol. / Design \Veld Vol.
SV D d n / Design \Veld DiaIn
SV7 Equiv-DiaIn / Design Weld Equiv-DiaIn
SW' 1'01. / Design \Veld Vol.
3L DiaIn / Design Weld DiaIn
3 L Equiv.DiaIn / Design \Veld Equv.DiaIn
3 L lrol. / Design \Veld lTol.
Pressure Atmchment / Design \Veld DiaIn
Pressure Anachrnent / Design \Veld Equiv.DiaIn
Pressure Anachment / Des& \Veld Vol-
Xon Pressure At t achent / Design \Veld DiaIn
Non Pressure Attachment / Design \Veld Equiv.DiaIn
%on Pressure Amchment / Design \Veld Vol.
Position Weld / Design \Veld DiaIn
Posiaon \Veld / Design \Veld Equiv-Ddn
Position \Veld / Design Weld Vol.
holl Weld / Design \Veld DiaIn
boll Weld / Design Weld Equiv-DiaIn
Kou \Veld / Design Wdd Vol.
khlti-Sraaon Roll Weld / Design Weld DiaIn
Ratio
9.232899E-05
0.0583273
5.403945E-04
0.9404383
5.956174E-02
5.786802E-O1
9.647334E-03
6.095 192
5.6471 22E-02
2.353484E-01
3.923557E-03
2.478906
2.296676E-02
3.322566E-03
5.539 139E-04
0.3499632
3.342367E-03
0.9782972
0.9782972
0.994041 3
1.335559E-O2
1 -335559E-02
3.170402E-O3
t3.34724GE-03
8.34734GE-03
2.78841 GE-03
0.9866444
0.986644
0.9968396
1.335559E-02
1 -335559E-02
3.1704OE-O3
O
O
O
Material
Total
Total
Total
Totd
T o d
Total
Totai
Total
Total
Total
Total
Total
Total
TGLZ
Total
Total
Total
Total
To td
Total
Total
To ta1
Total
Total
To ta1
Total
Total
Total
Total
Total
To-d
To td
Total
Total
Total
Total
Total
Total
Total
Size
-
To ta1
Total
Total
T o d
Total
Tord
Total
Total
Tod
Total
Total
T o d
T o d
To tai
Total
T o d
To ta1
To ta1
Total
To ta1
Total
Total
Total
Total
6- 14"
6-14"
6- 14"
6-14"
6-14"
6-14"
6- 14"
6-14"
6-14"
6-14''
6-14"
6-14"
6-1 4"
6-13"
6-14"
-
Ratio Description
Mulu-Station Roll \Veld / Design \Veld Equiv-DiaIn
hfuIu-S tation Roll \Veld / Design Weld VOL
Single-Station Roii \Veld / Design WeId DiaIn
Single-Station Roii \Veld / Design \Veld Equk-.DhLn
Single-Station Roli \Veld / Design \VeId Vol.
Ti Process iveid / Design \Veld Vol.
hLig Process Weld / Design \Veld Vol.
FU\ V Process \Veld / Design \Veld VoI.
Stick Process \Veld / Design Weld Vol.
SubArc Process Weid / Design \Veld Vol.
Rotoweld Process Weld / Design \Veld 1'01.
Repair Rate (No. of R / R+A)
No. of Cut Sheet Revision / No. of Spool
RT rate /Spool
MT rate /Spool
P T rate /Spool
PMI rate /Spool
P\VHT rate /Spool
F rate /Spool
B K N rate /Spool
V T rate /Spool
UT rate /Spool
Non-Welded Spool/\VeIded Spool (Weight)
Non-LVelded Spool/WeIded Spool pnits)
No. of Pipe Pieces / Footage
'io. of Pipe Pieccs / DiaInFt
'JO. of Pipe Pieces / Ton
Xo. of Pipe Pieces / Unit
?JO. of Pipe Pieces Over 3 ft / Footage
'JO. of Pipe Pieces Over 3 ft / DiaInFt
Yo. of Pipe Pieces Over 3 ft / Ton
30. of Pipe Pieces over 3 ft / Unit
go. of FLmges / Footage
go. of Flanges / DiainFt
90. of FIanges / Ton
30. of Fianges / Unit
30. of In-Line Fittings / Footage
90. of l n - h e Fitnngs / DiaInFt
30. of In-Line Fittings / Ton
Ratio
0.4257095
0.432653
0.5742905
0.5742905
0.567347
0.106721 1
0
0
0.893279
0
0
2.933985E-O2
0
100
100
0
1 O0
1
0
100
100
100
0
0
5.3 1 61 OSE-02
8.812613E-03
5.59941 1
5.187787E-02
5.0669 13E-02
8.M7 187E-03
5.336939
4.944609E-02
5.53761E-03
9-33 1899E-04
0.583272
5.303945E-03
l.lO752Z-O3
1 -84638E-04
0.1166544
Material
Total
Total
T o d
Tord
Total
Total
Total
Total
Total
Total
Total
Total
Total
T o d
Total
T o d
To ta1
Tord
Tord
Total
Total
Total
Tord
To ta1
Total
Total
Total
T o d
To ta1
Total
Tord
Total
To ta1
To ta1
To td
Total
Total
Total
Total
Ratio Descripaon
No. of In-Line Filtings / Unit
No. of Non-in-Le Fittings / Footage
No. of Non-In-Le Fittings / DiaInFt
No- of Non-In-Line Fitngs / Ton
No. of Non-l n-Le Filtings / Unit
No. ofTrdves / Footage
No. of lrdves / DiaInFt
No. of Vaives / Ton
No-of Valves / Unit
No. of Supports / Footage
No. of Supports / DiaInFt
No. of Supports / Ton
No. of Supports / Unit
No. of blisc. / Footage
No. of Ilfisc. / DialnFt
No. of blisc. / Ton
No-of blisc. / Unit
Pipe Weight / Spool LVeight
Non-Pipe LVeight / Spool Weight
No. of Connections (Design \Velds) / Footage
No. of Connections (Design \Velds) / DiaInFt
So. of Connections (Design \Velds) / Ton
30. of Connections (Design \Velds) / Unit
No. of Mulu-Station Roll \Velds / Footage
30. of hlulti-Station Roll \Velds / DiaInFr
No. of Mulu-Station Roll !Velds / Ton
No. of blulti-Scation Roll Welds / Unit
%o. of RepaLed \Velds / Footage
'sio. of RepaLed \Velds / DidnFc
go. of Repaired Welds / Ton
I\To. of RepaL Welds / Unit
3W DiaIn / Design Weld DiaIn
3\V Eq~~iv.DhIn / Design Weld Equiv.DiaIn
3W Vol. / Design Weld Vol.
SWDiaIn / Design Weld DiaIn
;W Equiv.DkIn / Design Weld Equiv.DiaIn
337 Vol. / Design Weld Vol.
3L Di dn / Design Weld DiaIn
3L Equiv-DiaIn / Design Weld Equiv-DiaIn
Ratio
1.080789E-03
5.2053536-02
8-6779856-03
5.482757
5.079708E-02
0
0
0
0
0
0
0
0
5.5376 lE-04
9.23 1899E-05
0.0583272
5.403945E-04
0.9404383
5.956 174E-02
5.786802E-02
9.647334E-O3
6.095 192
5.64712Z-O2
2.353484E-03
3.923557E-03
2.478906
329667GE-03
3.333566E-03
5.539139E-04
0.3499632
3.242367E-03
0.9782972
0.9782972
0.9940413
1.335559E-02
1.335559E-02
3.170402E-O3
8.347246E-03
8.347246E-03
Material
T o d
T o d
Total
Tord
Total
Total
Total
Total
Totaf
Total
Total
Total
Tord
Total
Tocd
Total
Total
Total
Total
Total
Total
Total
Total
Total
Total
To ta1
Total
Total
Total
Total
Total
Total
Total
Total
Toul
Total
T o d
Total
AS WOY)
Ratio Description
OL Vol. / Design LVeld VOL
Pressure Attachment / Design \Veld D d n
Pressure Amchment / Design Weld Equk.DiaIn
Pressure +.\nachment / Design Weld VOL
Non Pressure Amchment / Design Weld Didn
Non Pressure Attachment / Design Weld Equiv-Dialn
Non Pressure Atmchment / Design \Veld Vol.
Position \Veld / Design Weld DiaIn
Position \Veld / Design LVeld Equk-Dialn
Posiuon \Veld / Design Weld VOL
Roll Weld / Design Weld DiaIn
Roll LVeld / Design \Veld Equiv-DiaIn
Roll \Veld / Design Weld Vol.
Multi-Station Roll Weld / Design \Veld DiaIn
blulti-Station RoU [Veld / Desip WeId Equiv-DiaIn
Mula-Smtion RoU \Veld / Design T'Veld Vol.
Single-Station Roll Weld / Design \Veld DiaIn
SingIe-Station RoU Weld / Design Weld Equiv-DiaIn
Single-Station Roll Weld / Design \Veld 1'01.
T'ig Process Weld / Design \Veld 1'01.
Mig Process Weld / Design \Veld Vol.
FCAW Process Weld / Design Weld Vol.
Stick Process \Veld / Design Weld 1'01.
SubAxc Process \Veld / Design \Wd 1'01.
Rotoweld Process [Veld / Design Weld TToL
Repair Rate (No. of R / R+A)
NO. of Cut Sheet Revision / No. of Spool
RT rate /Spool
MT rate /Spool
PT rate /Spool
PMI rate /Spool
PWWT rate /Spool
FI' rate /Spool
B H N rate /Spool
VT rate /Spool
U T rate /Spool
%on-Welded Spool/\Velded Spool (Weight)
Son-Welded Spool/Welded Spool (Units)
rotal bIaaHours / Di;iIn*Ft
Ratio
2.78841 6E-03
0.986644
0.986644
0.9968396
1 -335559E-02
1 -335559E-02
3.170402E-O3
0
0
0
1
1
1
0.4257095
0.4257095
0.432653
0.5742905
0.5742905
0.567347
O. 106721 1
0
0
0.893279
0
0
2.933985E-O2
0
1 O0
1 O0
0
1 O0
1
0
1 O0
1 O0
100
0
0
1 -705658E-04
Material
AS (MoY)
AS (Moy)
AS (Moy)
AS (,.Vloy)
AS (AUoy)
*AS(.!!oy)
AS (Moy)
AS (MOT)
AS (-Uoy)
AS (hlloy)
AS ( M~ Y )
-1s (,-Uioy)
AS (Alloy)
-AS (-Moy)
AS (Aoy)
AS (Mo);)
AS (Moy)
AS (Ailey)
AS (AUoy)
*AS(.Woy)
-4s (Moy)
AS (Alloy)
AS (Moy)
AS (Moy)
AS (Alloy)
AS (AUoy)
AS (ruloy)
AS (Alloy)
-4s (-Uoy)
AS (Moy)
AS (Moy)
AS cMoy)
-AS (May)
.!AS (-Uoy)
AS WOY)
AS W ~ Y )
AS ( M~ Y )
AS (Moy)
AS W ~ Y )
Size
Total
Total
Totd
Total
Total
T o d
T o d
Total
Total
Total
Total
To td
T o d
T o d
Totai
Total
Total
Total
Total
Total
Tom1
To ta1
Tocal
Total
Total
Totai
To td
Total
Total
To td
T o d
Total
Total
Total
Total
Total
Total
T o d
T o d
Ratio Description
Total hfanHours / Equiv.Ddn*Ft
Tot d hlaaHours / Eqvlv-DiaIn
Total hianHours /
Total ManWours / Unit
No. of Pipe Pieces / Footage
No- of Pipe Pieces / DiaInFt
No. of Pipe Pieces / Ton
No. of Pipe Pieces / Unit
No. of Pipe Pieces Over 3 ft / Footage
No. of Pipe Pieces Over 3 t / DiaInFr
No. of Pipe Pieces Over 3 ft / Ton
No. of Pipe Pieces over 3 ft / Unit
No. of Fimges / Footage
No. of b g e s / DtaInFt
No- of b g e s / Ton
No. of b g e s / Unit
No. of In-line Fittings / Footage
No. of In-Line Fittings / DiaInFt
No. of In-line Fittings / Ton
No. of In-Line Filtings / Unit
No- of Noa-In-Line Fittings / Footage
No. of Non-ln-Line Fittings / DiaInFt
No. of Non-In-Line Fittings / Ton
No. of Non-In-Line Filtings / Unit
No. of Valves / Foomge
No. of Valves / DhInFt
No. of lrdves / Ton
No-of Valves / Unit
No. of Supports / Footage
No. of Supports / DiaInFt
No. of Supports / Ton
No. of Supports / Unit
No. of bLisc. / Footage
No. of blisc. / DhInFt
No. of PvL-c. / Ton
No-of PvLisc. / Unit
Pipe Weight / Spool LVeight
Yon-Pipe Weight / Spool Weight
30. of Connections (Design Welds) / Foomge
Ratio
1 -705658E-04
0-6 160267
2.657094
0.1394056
5.316105E-03
8.863633E-03
5.59941 1
5.1 87787E-02
5.0663 13E-02
8.U7 187E-03
5.336939
4.9UG09E-02
5.5376 1 E-03
9231899E-04
0.583272
5.103945E-O3
1 .IO7533E-O3
1 -84638E-O4
0-1 166544
1 -080789E-03
5205353E-02
8.677985E-03
5.482757
5.079708E-02
O
O
O
O
O
O
O
O
5.53761E-04
9.232 899E-O5
0.0583772
5.403945E-04
0.9404383
5.956174E-O2
5.786803E-02
Material
As (MoY)
AS (Moy)
AS Woy)
AS (Moy)
AS (-Uoy)
AS (MoY)
AS (Alloy)
AS (MoY)
AS (Moy)
AS (Moy)
AS (Mol-)
AS (Moy)
AS (Moy)
AS (UOY)
AS (-Uoy)
AS (Moy)
AS (rUoy)
AS (Moy)
AS (Noy)
-4s (i-uloy)
AS (Moy)
AS (Moy)
AS (iUoy)
AS (*AlIo).)
AS (hlloy)
AS (Alloy)
AS (LUoy)
AS (-Uoy)
AS (-AUoy>
AS (Moy)
AS (Moy)
-AS (rvloy)
AS ( M~ Y )
AS (Moy)
AS Wo y )
AS (May)
-4s (Moy)
AS (Ailey)
AS (Moy)
Size
-
Total
Total
Total
Total
Total
Total
T o d
T o d
Total
Total
Total
Total
To tai
Total
Total
Total
T o d
Total
T o d
Total
Total
T o d
Total
T o d
T o d
To ta1
Totai
Total
Total
Total
Total
Total
T o d
Total
Total
Total
r o ta1
Total
rotal
-
Rao Description
No. of Connecuons (Design LVelds) / Dk-JInFt
No- of Co~e c t i ons (Design \Velds) / Ton
No. of Co ~ e c o n s (Design Welds) / Unit
No- of Multi-Station Roll Welds / Footage
No- of Pvfulti-Station Roll WeIds / Dd n F t
No. of Mulu-Station Roll LVelds / Ton
No. of Mulu-Station Roll [Velds / Unit
No. OF Repaired !Velds / Footage
No. of Repaired Welds / DiaInFt
No. of Repaired Welds / Ton
No. of Repair \Velds / Unit
BW' DiaIn / Design Weld DiaIn
BW Equiv-Diain / Design Weld Equiv-DiaIn
BK' Vol. / Design WeId Vol.
SLVD a n / Design Weld DiaIn
S\V Equiv-DiaIn / Design \Veld Equiv-DiaIn
SV7 1'01. / Desiga Weld Vol.
OL DiaIn / Design Weld Di ah
OL Equiv-DiaIn / Design \Veld Equi vDkIn
OL Trol. / Design Weld Vol.
Pressure At t achent / Design \Veld DiaIn
Pressure Attachment / Design LVeld Equiv-DhIn
Pressure Attachment / Design \Veld Vol.
Son Pressure Amchment / Design \Veld DiaIn
Non Pressurc Attachment / Design Sreld Equiv-DiaIn
Non Pressure Attachment / Design \Veld Vol.
Posion \Veld / Design \Veld DiaIn
Position \Veld / Design \Veld Equiv-DiaIn
Position \Veld / Design Weld Vol.
Roll Weld / Design \Veld DiaIn
Roll Weld / Design Weld Equiv.DiaIn
Roll Weld / Design Weld Vol.
Mulu-Staaon Roll Weld / Design \Veld DiaIn
bluia-Station Roll 'LVeld / Design WeId Equiv-DiaIn
Uula-Station Roll Weld / Design Weld Vol.
Single-Station Roll Weld / Design Weld DiaIa
Single-Station Roll Weld / Design Weld Equv.DiaIn
Single-S tauon RolI \Veld / Design \Veld TroL
'ig Process Weld / Design WeId Vol.
Ratio
9.647334E-03
6.095 192
5.647 133E-02
2-353484E-O2
3-923557E-O3
2.478906
2296676E-02
3.333566E-03
5.5391 39E-04
0.3499633
3.2423G7E-03
0.9782972
0.9782972
O.9WO413
1.335559E-03
1.335559E-02
3.170402E-03
S.34724GE-03
8.347246E-03
2.78841 GE-O3
0.9866-l-N
0.9866CW
0.8968296
1.335559E-02
1.335559E-02
3.1 70402E-03
O
O
O
1
1
1
0.4257095
0.4257095
0.432653
0.5743905
0.5742905
0.567347
0.206721 1
Material
AS (Moy)
AS (Noy)
AS Woy)
AS (Moy)
AS (Mo).)
AS (Mo).)
AS (Moy)
AS (rvloy)
AS (Moy)
AS (Aioy)
AS (Moy)
AS (Noy)
Lis (Moy)
AS (-Uoy)
Lis (hlloy)
Lis (Moy)
AS (Uoy)
AS (rvloy)
AS (Moy)
AS (Noy)
AS (May)
AS (Moy)
AS (May)
,AS (Moy)
AS (Moy)
AS (Moy)
AS W ~ Y )
AS (AUoy)
AS (Moy)
AS (ALioy)
AS (LUloy)
AS (Alloy)
AS (Alioy)
AS (Uoy)
AS (May)
AS (Alloy)
AS P ~ Y )
AS (Aiioy)
AS (Alloy)
Size
-
To tai
Total
Total
Total
Total
Total
Total
Total
Total
Total
Total
Total
Total
Total
Total
Total
Total
Total
6-14"
6-14"
6-14"
6-1 4"
(5-14"
6- 14"
6-14"
6-14"
6-14"
6-14"
6- 14"
6-14"
6- 14"
6- 14"
6-14"
6-14"
6-14"
6-14"
6-14"
6-14"
6-14"
Ratio Description
Mig Process \Veld / Design Weld Vol.
F=\V Process Weld / Design Wdd Vol-
Stick Process !Veld / Design Wdd Vol.
Subhrc Process \Veld / Design Weld VOL
Rotoweld Process Weld / Design Weld 1701-
Repait: Rate (No. of R / R+A)
No. of Cut Sheet Revision / No. oESpool
RT rate /Spool
MTrate /Spool
PT rate /Spool
Ph11 rate /Spool
PWHT rate /Spool
FT rate /Spool
B HN rate /Spool
''T rate /Spool
UT rate /Spool
Non-LVelded Spool/Welded Spool (Weiglit)
Non-LVelded Spool/LVelded SpooI (Units)
No. of Pipe Pieces / Footage
No. of Pipe Pieces / DialnFt
30. of Pipe Pieces / Ton
No. of Pipe Pieces / Unit
No. of Pipe Pieces Over 3 Ft / Footzge
'JO. of Pipe Pieces Over 3 ft / DaInFt
No. of Pipe Pieces Over 3 Ft / Ton
So. of Pipe Pieces over 3 ft / Unit
Mo. of Fianges / Footage
'To- of Fhnges / DiaInFr
30. of Fhnges / Ton
No. of Fhnges / Unit
Mo. of In-Line Fitlingj / Footage
No. of In-Line Fittings / DialnFt
No. of In-line Fittings / Ton
qo. of In-Line Piltings / Unit
go. of Non-In-Lne Fitaogs / Footage
qo. of Non-la-Line Fittligs / DiaInFt
go. of Non-In-Line Fittings / Ton
go. of Non-In-Line Filtings / Unit
\JO. of Valves / Footage
Ratio
Materid
AS (Moy)
AS (Moy)
AS (Alloy)
AS (Ailey)
AS (r\Uoy)
AS (AlIoy)
AS (Moy)
AS (,-Vloy)
AS (~Uoy)
AS (AUoy)
AS (,-Vloy)
AS (Alloy)
AS (Moy)
AS (-Uoy)
AS (Moy)
-4s (Alloy)
AS (-Uoy)
AS (Moy)
AS (Auoy)
AS GWoy)
AS (Moy)
AS (Moy)
AS (AUoy)
AS (Alloy)
AS (,illoy)
AS (Moy)
AS (Moy)
AS Woy)
AS (Alloy)
Lis (Auoy)
AS (Lioy)
AS WOY)
AS (,.Vloy)
AS P O Y )
AS (Moy)
AS Woy)
AS (Mo y)
AS (May)
AS (May)
Size
-
6-14"
6-14"
6- 14"
6-24"
6-14"
6-14"
6-14"
6-1 4"
6-14"
6- 14"
614"
6-14"
6-14"
6-14"
6-14"
6-14"
6- 14"
6-1 4"
6-24"
6- 13"
6-14"
6-13"
6-1 4"
6-14"
6-14"
6-14"
6-14"
6-14"
6-14''
6-14"
6-14"
6-13"
6-13"
6-14"
6-14''
G- 14"
6-14"
6- 14"
G- 14"
Ratio Description
No. of Valves / DialnFt
No. of Valves / Ton
No-of Trdves / Unit
No. of Supports / Footage
No. oESupports / DiaInFt
No. of Supports / Ton
No. of Supports / Unit
No. of Mise. / Foomge
No. of hlisc. / Di aI f i t
No. of bLisc. / Ton
No-of hiisc. / Unit
Pipe Weight / Spool Weight
Non-Pipe \Veight / Spool LVeight
No. of Connections (Design \Velds) / Foocage
No. ofcomections (Design \Velds) / DiaInFt
No. of Co~ect i ons (Design \Xelds) / Ton
No. of Comrctions (Design Welds) / Unit
No. of hldti-Station Roll \Velds / Foomge
No- of bfulu-Scation Roll \Velds / DiaInFt
No. of Mdt i - kaon RoU \Velds / Ton
No. of hfulu-Station Roll \Velds / Unit
No. of Repaired \Velds / Footage
No. of RepaLed Welds / DiaInFt
No. of Repaired \Velds / Ton
No. of RepaL \Velds / Unit
B\V DiaIn / Design \Veld DiaIn
B\V Equiv-DiaIn / Design \Veld Equiv.Dia1n
B\V Vol. / Design \Veld Trol.
S\V DiaIn / Design \Veld DiaIn
SWEquiv.Didn / Design Weld Equiv-Ddn
S\V Vol. / Design \Veld Vol.
OL DiaIn / Design \Veld DiaIn
OL Equiv.DiaIn / Design Weld Equiv.DiaIn
OL Vol. / Design \Veld Vol.
Pressure At t achent / Design Weld Dialn
Pressure Attachment / Design Weld Equiv.DiaEn
Pressure Attachrnent / Design \Veld VOL
Non Pressure Attachmenr / Design 'WeId DiaIn
Non Pressure Attachment / Design WeId Equiv.Dialn
Ratio
Material
AS (LLUoy)
AS (U~OF)
AS (Moy)
AS (Moy)
AS (MoY)
AS (Uloy)
AS (Moy)
A4S (,.Ulo:.)
AS (Moy)
AS (AUoy-)
-4s (-Uoy)
-4s (Moy)
AS (+Uoy)
AS (Uloy)
AS (Moy)
AS (-Uoy)
-4s (Moy)
AS (,-Uioy)
AS (Uo)
AS (Alloy)
'4s (*Uoy)
AS (Moy)
AS (Noy)
AS (Alloy)
AS (Alloy)
AS (-Uoy)
AS (Noy)
,AS (Mo).)
'4s (-Uoy)
AS (-Uoy)
AS (Moy)
AS (floy)
Size
-
6-14"
6-14"
6-14"
6- 14"
6-14"
6-14"
6- 13"
6-14"
6-14"
6-14"
6-14"
6-14"
6- 14"
6-14''
6- 14"
6-13"
6- 14"
6- 14"
6-14"
6-14''
6-14"
6-14"
6-14"
6-14"
6-14"
6-14"
6-14"
6-14"
6-14"
6-14"
6-14"
G-14"
Ratio Description
Non Pressure Attachent / Design \Veld Vol.
Position !Veld / Design \Veld Diain
Position LVeld / Design Weld Equiv-DiaIn
Position \Veld / Design \Veld Vol.
Roll \Veld / Design \Veld D d n
RoU\Veld / Design LVeld EqukDiaIn
Roll Weld / Design \Veld VOL
blulti-Station Roll \Veld / Design \Veld DiaIn
hlulti-Staaon Roll \Veld / Design Weld Equiv-DiaIn
hlulti-Station Roll Weld / Design \Veld Vol.
Single-Station Roll \Veld / Design \Veld DiaIn
Single-Station Roll \Veld / Design Weld Equiv-Dialn
Single-Station RoU \Veld / Design \Veld Vol.
rig Process Weld / Design \Veld Vol.
4% Process Weld / Design Weld Vol.
FCALVProcess \Veld / Design Weld VOL
Stick Process \Veld / Design \Veld Vol.
SubArc Process \Veld / Design Weld Vol.
Xotoweld Process Weld / Design \Veld Vol.
XepaL Rate (No. of R / Rt-\)
go. of Cut Sheet Revision / No. of Spool
XT rate /Spool
b1Trate /Spool
?T rate /Spool
'MI rate /Spool
?\VHT rate /Spool
T' rate /Seo01
3HhTrate /Spool
4T rate /Spool
JT rate /Spool
\ion-Welded Spool/CVelded Spool (YVeight)
\ion-ivelded Spool/Welded Spool (Uni=)
Ratio
3-170402E-03
O
O
O
1
1
1
0.1257095
0.4257095
0.432653
0.5742905
0.5732905
0.567347
0.106721 1
O
O
0.893279
O
O
2.933985E-02
O
100
100
O
1CO
1
O
100
1 O0
100
O
O
FabMaster processed project data individually and warehoused the item-coded
project information in an easy-to-access format. Fab-OLAP provides the functionaliq
of vieming and analyzing productiviq-related information across dl projects that have
been processed bp FabMaster. Fab-OLAP is an On- he Andytical Processing System
custom-developed for the Fabrication FaUlities of PCL Indusmal Constructors, Tnc. The
system feanires dynamic query, graphic presentation, and the functionality of statistical
anaiysis on 105 ratios of labor productivity/spool configuraton compleSq/qualiq
control. It is an advanced decision-support tool for management to grasp the trend in
the historical project data and identifg exceptional problems in the work at hand.
Figure C-1: Select one ratio
Step 1. Load the program and select one ratio
User selects one ratio from the "Select Ratio" dropdomn Est, which includes all
the 105 ratios computed in FabMaster, as shown in Figure C-1.
Step 2. Apply Filters on Material Type and Size Range
Fab-OLAP uses the standard codes of the Company for the material types and
size ranges. Fab-OWP helps user e-xplore data in decision-oriented mays and allows
user to view data and get at them fomi different perspectives dong the dimension of
181
Figure C-2: Trial on ccnumber of pipe pieces per foot"
material and size. The histogram dong &th ~ t a t i ~ t i d analysis results for the selected
ratio is presented on screen and updated automatically. Figure C-2 shows the nial based
on "carbon steel 6-14 inch spool, nurnber of pipe pieces per foot of pipe".
Step 3. Drill into details of data
Fab-OUP doms user to drill dom to details of data by clicking the Ti e w
Data" button. Figure C-3 shows the data behind the selected ratio.
1 7250 / CS (Carbon] 6-1 4" j No- af Pipe Pieces / Footage i 0.13346798!
-
1 700204j CS (Carbon] 6-1 4" I + No. of Pipe Pieces / Footage : 3.660536~-02;
1 7002551 CS [Carbon) 1 6-1 4" i No- af Pipe Pieces / Faotage 1 O.OU4265;
1 7002651 CS (Carbon) 1 6-1 4" !No. of Pipe Pieces / Footage ' 4.8767 1 7E-02;
17002341 CS (Carbon) 1 6-1 4" 1 No. of Pipe Reces / Foatage 0.0488468 3
17004781 CS (Carbon] 16-1 4" No. of Pipe Pieces / Foatage 6.061 01 9E-M!
I
1 700205; CS (Carbon] / 6-1 4" ' No. of Pipe Pieces / Footage O. 0656395 3
i
1
' 1 700474[C [Carbon]
6-1 4" j No. of Pipe Pieces / Footage p 6.587098~-021
1
17004621 CS [Carbon] 6-1 4" : l No. of Pipe Pieces / Footage 16.762063E-021
1700466 j CS [Carbon] 6-1 4" i No. of Pipe Pieces / Footage 7.231 822E-Mf
1 700242! CS [Carbon) / 6-1 4" i No. of Pipe Pieces / Fooiage / 7.61 0802E-021
1700206j CS [Carbon] 1 6-1 4" b No. of Pipe Pieces / Footage 7 2931 4~-021
1
1700481 j CS [Carbon) i 6-1 4" No. af Pipe Pieces / Footage 7.750466E.O2]
'
1700211 i CS [Caban) j 6-1 4" No. af Pipe Pieces / Fmtage .8.051168E-021
i
1700464 [ CS [Cabon) 1 6-1 4" 'No. of Pipe Pieces / Footage 8 496705E-Mi
1700491 ! C (Carbon) 1 6-1 4" ' No. of Pipe Pieces / Footage 8 67351 4E-02:
1
1 70021 3; CS (Caban] 16-14'' :No. of Pipe Pieces / Footage 1 8.859606E-021
Figure C-3: View details of data
Step 4. Prht out the trial and statistical analysis resdts
User clicks the 'Trint out" button to psint a hard copy of the current trial and
stausti cd analysis results including the histograrn for record.
&PENDIX D: USER'S MANUAL FOR PIPINGMASTER
PipingMaster is a historical project data =-arehoushg system customized for the
field construction systems of PCL Industrial Constructors, Inc. It is an automated data
processing tool to estract ram data from Labor Cost Control System, Estimating System,
and Quality Contcol Spstem, and convert raw data into useful pruductiviq information
based on embedded expert de s . Pipe handling and welding are processed by
Pipinghlaster independendy, but in simila fashions including the user interfaces and
prograrn logic. Thus, Pipe handling is selected to illustrate the program flow in the
following steps.
Step 1. Impoa raw data in standard format
User irnports three ram data tables for each project into the database manudy to
d o w for PipingMaster to calculate the quantity of piping work, namely,
RDJroject#Hand table for pipe hanrlling, RD-Project#Detl for pipe work
components, RD-Project#Weld for pipe melding. The table structures are shown in
Figure D-1.
RD-Pro j#Hand
Project #
Nominal Size
Schedule
Ciassifcation
hhterial Type
Length (fi)
Es tUnitPvLH
RD-Pro j#Detl
Project #
Detl Type
Nominal Size
Classifca tion
Mzterial Type
Quantity
RD-Pro j#Weld
Project #
Nominal Size
Schedule
Joint Type
Classification
Material Type
# \cVelds
Es tUnitlhIH
Figure D-1: Structures of Raw Data Tables for A Projcct
The detailed quantity take-off (in footage) for pipe handling of one project is
available in the project estimate only. L'sually information is known and complete o n the
size, the thickness, the material tgpe, and the location classification of each individual pipe
section.
The detailed quantity take-off in number of welds for pipe melding of one project is
availaole either in the project estimates or in the field quality control system. In most cases
the pipe ske, pipe thickness, pipe material type, location classification and meld joint type are
knomn for each individual weld.
Installation of other piping moxk components (or piping details) includes pipe
supports, bolt-ups, valces, screw joints, and misceUaneous items like flanges, specialties,
185
elbows, cuts and bevels. The number and type ofwork components and estimated unit man-
hours for one project me available in the project estimates. However, information on the
size, material type, location dassihcation may not be found in the estimate. Therefore, we
need to check the ram data integrty of the piping work components prior to processing.
Step 2 Raw Data Integrity Check
The raw data integrity check is controlled by the entered project setting regarding the
raw data integiq and methods of actual mm-hour cost coding as shown in Figure D-2. User
Mat ui d Type
- ----
Sue Ra ge [<TT'-16'>16'1 y=
OIoaae Ho If Co& To Total ri NO
Lwd
Figure D-2: Main user interface of FabMaster
enters the project number to be processed and answer a number of Yes/No questions about
the project Nesq user clicks "Check Raw Data Integritg" button to s t m the program. User
dl be prompted to correct any problems due to failure to pass the checks.
The PipingMaster is capable of identfyng missing data or incorrect data in the raw
data tables. For esample, if actual labor hours in the labor cost system were tracked to the
level of various classifications of location, then a null in the "Classihcation" field of the raw
data tables d be detected as invalid data and must be corrected for W e r processing.
Three valid types ofmeld joinh i.e. BW putt Weld), SW (Socket Weld), OL (Olet Weld) and
five valid types of piping work components are allowed lo the RD-Project#Ded table, 1.e.
bolt-up, valve, screw joint, support, and misc.
Step 3 Check cross-reference Data for Quantity Calculation
User hrst chooses one of four options and then click the "Check Cross Reference
Integriy" button to perform the check for the selected option. Four options should be
checked through one by one. User d be prompted to correct raw data or update cross
reference tables in case Pipinghlaster finds a problem. The "Action" button d only be
activated when aLi the ram data checks and cross reference checks are passed.
In PipingMaster, a number of cross-reference tables are involved to cal dat e the
quantitg in various units of measurement, i.e. five units of measurement for pipe handling:
DiametePLength (Tnch*Feet), Equivalent Diameter*Length (Inch*Feet), Length (Fee t),
Weight (p.lI.T), and Base Manhours (MH); five units of measurement for pipe welding:
Diameter (Inch), Equivalent Diameter (Inch), Volume (Cubic Inch), Volume/Thickness
(Square Inch), and Base bfanhours ($El). Cross-reference integrity check is performed to
ascertain that each record in the raw data table can h d the needed information in the
conesponding cross-reference tables so as to calculate accurate quantities. The forrnulae
187
used are commonly found in an industriai maoual or piping handbook- The relationships
between raw data tables and cross-reference tabIes are shomn in Figure D-3 and Figure D-4.
NominalSize
Schedule
RD-Proj#Hand
Project #
Nominal Size
Schedule
Classiacation
Ma t e d Type
Length
OuterDiameter (inch)
Figure D-3: HandLuig: X-Refercncc Information Integrity Check
Schedule
Thicknes (inch)
Project #
Nominal Size
Schedule
Joint Type
Classification
Ma terid Type
# Welds
Es tUnitkfH
- 1 Schedule
JT 1s OL
tblOletDim
Nominal Oudet
Dimension B
Figure D-4: Welding: X-Reference Information Integrity Check
Step 4 Generate Aggregate Cost Codes and Calculate Quantities
User hits the "Action" button to generate aggregate cost codes to the level of project
nurnber, classiication of location, material type, size range, activity, and unit of measure. The
total quantities and quantities breakdown for size ranges, dong with the generated cost
codes 1 . be appended a summary table called "tblQuanutyLMaster". Table 1 shows sample
records in the summary table for one relatively s md job.
Table D-1: Sample of Quantity Calculation Surnmary Table in PipingMaster
Pro ject# Material
CS
CS
CS
CS
AS
AS
SS
SS
TOT
TOT
Class
41 0
41 0
460
460
460
460
460
460
31 O
460
CostCode Description
Welding Total
Volume/Thickness
Handling Total Feet
Handiing Total Feet
Welding To ta1
Volurne/Thickness
Handling Total Feet
Welding Total
Volume/Thickness
Handling To tal Feet
LVeIdrng To t ri
Volurne/Thickness
Hand Tot Mati Tot Size Ft
Hand Tot Mat l Tot Sizc Ft
Step 5 Enter ActuaI Hours and Compute Actual Degree-of-difficulty Factors
Folloming generating the cost codes and calculating the quantities, PipingbIaster
Show Gnnpkd Records aid A d ln Actual Mhs h m LCS Report
Proje& 1 Materiaflypa 1 Classification 1 CostCode 1 QtyTotal 1 BaseMH 1 Adual Mhn 1 Statris [
Shan U35 Act MHr la Girrcnt Roiect
Show Campikd Han- MuPipiicrt:
1 Proje- lateria4 Classl CostCode 1 QtyBetow2 ( Qty2TolS 1 QtyAbovelS 1 QtyTotal 1 ES!MUI~ M u l t l Statur
15L30486 CS 410 302151-02 11 5623 O 5634 100 388 No
Figure D-5: Productivity Analysis Page for Pipe Handiing
sliifis focus to productiviry analysis page as shown in Figure D-5. User reads acrual
manhours and enters into the "actuai manhours" column for corresponding records. Nesr,
user hits the "Analyze" button to let PipingMaster figure out the actud labor hours for pipe
handling based on the project setting about actual labor cost t r a c h g practice and the
ernbedded espert d e s for handling different scenarios. Evennidy, the acmal degree-of-
difficulty factors are computed for each record and listed against the factors estimators have
used for comparison. After comparison, user decides on which records are valid for NN to
use by switching the status of one record from No to Yes.
Step 6 Make Questionnaires for Valid Records
T L INDUSTRIAL CONTRACTORS INC. Pipe Handling Report
Prtpaied By: I #i t di Sotaert Report Date: 15/7/59
1 .Handling Job Group by Prqect # 1500484_007 CosiCode: m-
Figure D-6: Sample of Pipe Handling Questionnaire
PipingMaster m&es questiomaires for those valid records as conhmied by user.
Figure D-6 shows a sampIe questionnaire.
Followng the above six steps, Pipngbkister processes one project and convert ram
data into accurate cost-coded productiviy information for f5uther productivity analysis.
Figure D-7 shows the flom chart of the whole program.
1 Check Raw Data Intecgrity
1
[ Correct Missng/
Compile
Estimated Man-
hours for Pipe
Work
Component from
the Raw Data
N 1 Incorrect Data
1s Data Complete and
Correct?
\1/
Prorate
Estimated Man-
hours for Pipe
Work
Component
Based On Pipe
A
1s Piping Work Component
included in the "Pipe
Install~' Cost Code?
N
Cost-Coded MA-hours &
Genqrate Indices
Extract Es timated Man-
hours fcom "Pipe Instd"
1s Output Data Valid?
Y
Y
k/ v
<
Generate Historical Cost
/ Codes/Qty for NN
Y
I
Quantitative Input and 1
Questiomakes for 1
1 Subjectivq Data CollecMg 1
1s ~ocat i & Classification,
Matenal Type, and Size
Ranges Known For Piping
Work Component?
9
Feed Valid Data To NN for
Training
I
Figure D-7: Program Flow Chan of PipingMaster
N
Generate Indices
for Piping \Vork
Components
APPENDIX E: USER'S MANUAL FOR SENSIT~VENN
SensitiveNN is a back propagation Neural-Network based system to analgze the
sensitivity of input factors in some comples engineering and management problems that
are not amenable to analy-sis ushg conventional mathematical models- The sensitivity
analysis method is proposed in this thesis.
Step 1 Prepare data for SensitiveNN
The last column in a data table must be named as "Status", which flags the
teaining/testing status for each record. Status 1 stands for a training record, and Status 2
for a tesring record, and Status O for an ignored record. The nest-to-last N columns in a
data table contain Actual Output Values of the target nsky variables such as actual
nroduction rates, N being the number of outputs. rill the remaining columns in a data
table will be the input factors. There are no requirements imposed on the column names.
The trainer dl count the number of inputs and outputs according to user's setup of the
netsvork, which is discussed in Step 3.
Figure E-1: Splash Screen of SensitiveNN program
The prepared data table for PINN must be imported to the database file
"FE3PNN.mdb" prior to analysis, which is instded nrith the program and by default
under the program folder.
Once the data table is imported, user can st ar t up the program
"SensitiveNN.exe". The splash screen shows up as in Figure E-1. By hittng the forin,
user proceeds to &.e nest step.
Step 2 Link SensitiveNN to data tables and select the one for analysis
Figure E-2 shows the switchboard of the program. User needs to link to
"FFBPNN.rndb" h s t in order to load up data. By clicking the "Link FFBPNN.mdb"
button, an "Open File" dialog form pops up as show in Figure E-3.
NOTE TO USERS
Page(s) missing in num ber only; text follows. Page(s) were
microfilmed as received.
196
This reproduction is the best copy available.
* - . _ _ -
. -
.- .
- .
-linport& .-.
- -
~ h e ~ ~ a t a set fa NN Anabsis shdd
have been lmportedinto -
FFEPNN;mdb ih theformat of a
Table with a Status Field!
Figure E-2: Program Switchboard
, .
. r Open aspad-only
Figure E-3: Open FFBPNN-mdb Fust
Once FT;SPNN.mdb is linked, all the data tables are listed in a combo bos for
197
user to select the one for analysis. AU the M d s in the selected table are numbered and
Listed in the list box captioned "Field List of Selected Datay', as shonm in Figure E-4. By
diclckg the "Show Datayy button, user can examine the details of data and edic the
train/test status for each record, which is shomn in Figure E-5.
- .
Show Data
Enter 1
Importa&
T ~ - D & 'Set for NN Analpis .should
have been Irnported nto .
FFBPNN.mdb h the Format of a
Table wth a S tatus Field!
Fiold List of
Selected Data. -
Figure E-4: Select data source table
. . . . ..
Figure E-5: Examine details of data and edit record status
Step 3 Set up NN structure and Iearning parameters to train-test BP NN
Followkg linktng to the data source, user click the "Enter" button on the
switchboard to enter the main interface of the program, as shown in Figure E-6.
User enters the tal ID, the leaming rate, the momentum rate, the number of
inputs, the number of hidden processing elements in the rniddle layer, the number of
outputs, the training iterations, and the threshold of global error to terminate learning. Ir
is important to match the number of inputs and outputs with the number of columns of
the linked data table in the previous step. User may revert to the switchboard (Figure E-
4) my t h e for double check bp clicking the "Exit" button, and clicks the "Enter" button
on the switchboard to restore the main interface. The mal ID is used to identify a
spedfic tnin-test trial and store the n e ~ o r k parameters and weights.
I 1
f tain 'n Test
- - I
Figure E-6: Program main interface of Sens itiveNN
User may refer to the pertinent paper for details of those NN parameters.
Once the network is set up, click the "Train n Test" bunon to start the leaming
peocess, mhich can be monitored through a progress bar. The current iteration and
global error are also shonm at the top and bottom of the progress bar dynamically during
the leaming.
Step 4 Training tenninates and investigate leamng results
Figure E-7: Check learning results when NN training terminates
The trainkg process terminates when any of the foilowing conditions is satisfied:
1. The current training iteration hits the uses-specified total iterations.
2. The current global enor is lomer than the user-specified threshold of global error.
3. User hits the "Stop" button.
When training terminates, user investigates the leaming results by checking final
global enor and compwing the actual outputs against the NN outputs for both training
and testing data as shown Li Figure E-7. Note that the average absolute errors for both
training data and leaming data are cornputed and shown in the screen as well. If the
average absolute enors for both training data and lemming data are reasonably small, the
netwolk is declared to be crained and the program flow moves to the next step.
Othernrise, repeat step 3.
Step 5 Perform Monte Carlo simulation to analyze the sensitivity of input factors
Based on a mature network obtained from Step 4, user spedies the total number
of simulation nrns in the left lower corner of the main interface and clicks the
"Sensitivity Analysis Simulation" button. m e n the simulation is done, the statisticai
analysis results of simulation about the input sensiavity berneen each input-output pair
are shown on screen as in Figure E-8- X tab-delimited test 6le called "SenNNFile-txt" is
also generated in the program folder recording the simulation results, mhich can be
imported to Excel for p l o t ~ g . Note tiiat the text H e ~vi l l be erased nest tkne the
simulation is perfomed, thus user should back it up if iieeded.
Step 6 Save a trial
User cari Save a mal including the network structure, l e d g parameters and
final weights of trained netsvork by clicking the "Save Trial" button. The trial ID d be
the key for access the network at later times, hence must be remembered.