Abstract
The general Visual Condition Index is an integral component of a Pavement
Management System. The current method in use in South Africa is based on an
aggregated visual condition. The method of analysis combines the visual pavement
condition data from individual distress types into an index representing the general
pavement condition, the Visual Condition Index. The calculation of the Visual
Condition Index requires the selection of a weighting factor for each type of distress.
The aim of this paper is to evaluate the capabilities of Artificial Neural Networks in
determining the general Visual Condition Index of Flexible Pavements using distress
data collected through visual assessments of the pavement surface.
1. INTRODUCTION
The purpose of this paper is to evaluate the capabilities of Artificial Neural Networks
(ANN) in determining the Visual Condition Index (VCI) of flexible pavements in a
pavement management system (PMS), using distress data collected through visual
assessments of the pavement surface and compare it with the current method of
calculating the VCI. The data collection method is according to the national standard
for the identification and collection of visual distress data for flexible paved roads,
TMH 9: “Pavement Management System: Standard Visual Assessment Manual”,
1992. By using some of the pavement management visual data from the existing
database as the training set and comparing it with the assessor’s “Overall Pavement
Condition” (OPC) categories (Very Good, Good, Fair, Poor and Very Poor), the VCI
was determined, using an ANN. Instead of relying on aggregated condition indices,
this paper examines the feasibility of using an ANN to determine the VCI of flexible
pavements.
TRB Committee AFD10 on Pavement Management Systems is providing the information contained herein for use by individual practitioners
in state and local transportation agencies, researchers in academic institutions, and other members of the transportation research
community. The information in this paper was taken directly from the submission of the author(s).
4th International Conference on Managing Pavements (1998)
Current practice
VCIp C N
F
1001 * n
n1 ..................................................................... (Eq.1)
Where:
VCIp = Preliminary VCI
Fn = Dn*(En)*W n............................................................................................................ (Eq.2)
n = Visual assessment item number as currently in use by PAWC
Dn = Degree rating of defect n
Range: 0 to 4 for functional defects
: 0 to 5 for other defects
En = Extent rating of defect n
Range: Default 3 for functional defects
: 0 to 5 for other defects
Wn = Weight per item number for defect n (Table 2)
C = 1/ N Fn max
n1
Fn(max) = Fn with degree and extent ratings set at maximum.
TRB Committee AFD10 on Pavement Management Systems is providing the information contained herein for use by individual practitioners
in state and local transportation agencies, researchers in academic institutions, and other members of the transportation research
community. The information in this paper was taken directly from the submission of the author(s).
4th International Conference on Managing Pavements (1998)
To try and improve the correlation of the VCI with the OPC based on the engineering
judgement of an expert panel, modifications to the formula is applied with an extent
weight factor (Yn) and a small degree factor (Sn). These factors are shown in Table
2. To accommodate these factors Equation 2 is adjusted as follows:
Fn = Dn*(En ^ yn) *W n * Sn ..................................................... (Eq.4)
The product of Dn*(En ^ yn) is limited to a maximum of 12 for functional defects, i.e.,
item no.’s 15, 16, 17 & 18 and to maximum of 25 for the other defects. The small
degree factor (Sn) is set to 1 for functional defects degree rating > 1, or for other
defects degree rating > 2, or else the Sn is according to Table 2. The values for the
constants “a” and “b” in the transformation formula 3 is set to 0,04 and 0,0006
respectively.
The purpose of the extent weight factor (Yn) is to provide weighting to the extent
rating for distress types. The purpose of the small degree factor (Sn) is to reduce the
contribution of certain defects to the total index value when their degree rating is
small.
TRB Committee AFD10 on Pavement Management Systems is providing the information contained herein for use by individual practitioners
in state and local transportation agencies, researchers in academic institutions, and other members of the transportation research
community. The information in this paper was taken directly from the submission of the author(s).
4th International Conference on Managing Pavements (1998)
This section provides an overview of neural network principles. Neural networks can
be used for a wide variety of learning tasks. They should be considered a
mathematical tool, similar to regression analysis. The key feature of neural networks
over regression analysis is that neural networks use non-linear mathematics and
therefore can be used to model highly complex and non-linear functions.
A learning task consists of a set of data from which training examples are formed.
Each training example usually comprises input data and the desired network
response. When applying a neural network to learn from this set of data, the training
of the network requires a number of decisions related to the use of the available
training methods. Selecting the appropriate method consists of configuring a neural
network and selecting algorithms (learning rules) for training the network.
The learning algorithms use sets of input and output vectors to train a network.
Using an error function that measures the distance between the desired output vector
and the current actual output vector, a learning algorithm adjusts the weights in the
network to reduce the average error over the set of training examples. Done
correctly, the trained network will reliably estimate the required response for new
inputs.
The activities in the training of an ANN are summarised as follows:
Determination of the task to be performed by the network in the application.
Analysis of the data available for the application.
Choice of the inputs to the neural network.
Proper pre-processing of the data for input to the network.
Choice of the desired outputs of the neural network, including post-processing of
the outputs for use in the application.
TRB Committee AFD10 on Pavement Management Systems is providing the information contained herein for use by individual practitioners
in state and local transportation agencies, researchers in academic institutions, and other members of the transportation research
community. The information in this paper was taken directly from the submission of the author(s).
4th International Conference on Managing Pavements (1998)
Choice of the neural network learning method and algorithm (learning rule) to be
used for training.
Setting of the parameters associated with the network chosen, including number of
the processing elements in each layer, type of processing elements and learning
constants.
Training the neural network on the training data.
Verification of the trained network on the test data.
Analysis of the results and possible re-training of the network or modifications of
parameters or pre-processing.
Integration of the trained network into the application software.
Initially the available neural network software on the World Wide Web was reviewed.
The software package ThinksPro™ (Logical Design Consulting Inc., 1996) was
selected. It operates in the Windows environment, including Windows 95 (Microsoft
Corporation) and Windows NT (Microsoft Corporation). ThinksPro™ provides
everything needed to develop powerful neural network applications in one easy-to-
use program. It allows the professional user to build custom Windows based
software.
TRB Committee AFD10 on Pavement Management Systems is providing the information contained herein for use by individual practitioners
in state and local transportation agencies, researchers in academic institutions, and other members of the transportation research
community. The information in this paper was taken directly from the submission of the author(s).
4th International Conference on Managing Pavements (1998)
TRB Committee AFD10 on Pavement Management Systems is providing the information contained herein for use by individual practitioners
in state and local transportation agencies, researchers in academic institutions, and other members of the transportation research
community. The information in this paper was taken directly from the submission of the author(s).
4th International Conference on Managing Pavements (1998)
A total of approximately 13 500 visual assessments (data points) are available in the
PAWC PMS database (the total population). An effort was made to evaluate all
these data points in the ANN. However, the magnitude of the data led to long
computing times. A screening process had to be implemented to reduce the data
used in the ANN model. An inspection of the data revealed that approximately 50
data points are available in the low (very poor) OPC category, see Figure 2. It was
6000
5000
Number of obs.
4000
3000
2000
1000
Good
Poor
Fair
Very poor
Very good
Figure 2: Number of Data Points per OPC
decided to select data uniformly from all OPC categories on a random basis. The
1996 data was excluded and used as the testing set.
The first runs of the ANN revealed certain anomalies in the data. Close investigation
showed that data errors were present in the assessment of the OPC. These errors
were corrected before being included in the final model.
Table 4: Data points per OPC category before and after data verification
OPC Pre verification Post verification
Very poor 32 24
Very poor/Poor 50 50
Poor 50 47
Poor/Fair 50 51
Fair 50 44
Fair/Good 50 54
Good 50 54
Good/Very good 50 53
Very Good 50 55
One can see from this how useful a tool the ANN can be by identified individual data
anomalies during the verification process. The distribution of the data points after
verification is shown in Table 4.
TRB Committee AFD10 on Pavement Management Systems is providing the information contained herein for use by individual practitioners
in state and local transportation agencies, researchers in academic institutions, and other members of the transportation research
community. The information in this paper was taken directly from the submission of the author(s).
4th International Conference on Managing Pavements (1998)
MSE D Y / M
ij ij
2
...................................................................... (Eq.5)
where:
Dij = the ith desired output for training example
j Yij = the ith actual network output for example j
M = the number of outputs times the number of examples
3.4.4. Number of processing elements in the hidden layer and the output layer
The number of processing elements in the hidden layer was set to 12. The MSE was
not significantly improved by varying the number of processing elements between 8
to 20 processing elements in the improvement of the MSE. The output layer is one
as this is the estimate of the Overall Pavement Condition.
TRB Committee AFD10 on Pavement Management Systems is providing the information contained herein for use by individual practitioners
in state and local transportation agencies, researchers in academic institutions, and other members of the transportation research
community. The information in this paper was taken directly from the submission of the author(s).
4th International Conference on Managing Pavements (1998)
3.4.5. Learning rule in the hidden layer and the output layer
It was decided to use the Jacob’s Enhanced Back Propagation (JEBP) learning rule
for both the hidden and the output layer. The software developers recommended the
use of this learning rule. The JEBP learning algorithm can be thought of in terms of
Back Propagation with added functionality. Back Propagation is probably one of the
most popular and commonly used algorithms. A detailed explanation of the JEBP
learning rule is available in ThinksPro™ User’s Guide. (Logical Designs Consulting,
Inc.)
3.4.6. Pre-processing function for the input layer
The input variables were normalised as explained in paragraph 3.2 and there was no
pre-processing.
3.4.7. Input function for the input layer and hidden layer
For both the output layer and the hidden layer the Dot Product Input Function was
used. The Dot Product function is a weighted sum of the inputs plus a bias value.
The Dot product function scales each input according to its relative influence in
increasing the net input to the processing element.
3.4.8. Transfer function for the hidden layer and output layer
The transfer function acts on the value returned by the input function. For the hidden
layer the Sigmoid Transfer Function was used. It is the most popular transfer
function among neural network users. The sigmoid acts as a "squasher,"
compressing the input function when it takes on large positive or large negative
values. Large positive values asymptotically approach 1, while large negative values
are squashed to 0. The sigmoid limits the output of the processing element to the
interval [0,1] and the function is given by Equation 6 and is shown in Figure 3.
tf 1/1 exp f ............................................................................... (Eq.6)
TRB Committee AFD10 on Pavement Management Systems is providing the information contained herein for use by individual practitioners
in state and local transportation agencies, researchers in academic institutions, and other members of the transportation research
community. The information in this paper was taken directly from the submission of the author(s).
4th International Conference on Managing Pavements (1998)
The transfer function for the output layer was set to the Threshold Linear Transfer
Function, which returns f when it lies in the interval [0,1]. 0 is returned when f<0 and
1 is returned when f>1. See Equation 7 and Figure 4.
4. FINDINGS
After training the ANN and using the training set (432 data points) described in the
previous paragraphs, the 1996 visual data (1912 data points) was used for the
testing set and evaluated. Figure 5 shows the MSE for both the training and test set
TRB Committee AFD10 on Pavement Management Systems is providing the information contained herein for use by individual practitioners
in state and local transportation agencies, researchers in academic institutions, and other members of the transportation research
community. The information in this paper was taken directly from the submission of the author(s).
4th International Conference on Managing Pavements (1998)
To compare the ANN output values, which are in the range [0,1], to the calculated
VCI as per Equation 3, the output value was multiplied by 100 to convert it to a
percentage scale [0,100], which is similar to the VCI scale.
4.2. Comparison of ANN with VCI
A graphical comparison of the data is shown in Figure 7, which is a scatter plot of the
TRB Committee AFD10 on Pavement Management Systems is providing the information contained herein for use by individual practitioners
in state and local transportation agencies, researchers in academic institutions, and other members of the transportation research
community. The information in this paper was taken directly from the submission of the author(s).
4th International Conference on Managing Pavements (1998)
TRB Committee AFD10 on Pavement Management Systems is providing the information contained herein for use by individual practitioners
in state and local transportation agencies, researchers in academic institutions, and other members of the transportation research
community. The information in this paper was taken directly from the submission of the author(s).
4th International Conference on Managing Pavements (1998)
3000
2500
2000
ANN
Km
1500
VCI
1000
500
0
Very P oo r F a ir G ood V e ry
poor good
OPC
5. CONCLUSION
The application of an ANN for determining the VCI of surface roads has been proven
a feasible method. The subjectivity in determining the OPC is minimised by the use
of an ANN with the only remaining subjectivity lying with the original evaluation of the
assessor. No subjective functions and factors are built into the final answer in
contrast to the method currently used to calculate the VCI. It is therefore
recommended that either of the following is to be done to calibrate the ANN:
TRB Committee AFD10 on Pavement Management Systems is providing the information contained herein for use by individual practitioners
in state and local transportation agencies, researchers in academic institutions, and other members of the transportation research
community. The information in this paper was taken directly from the submission of the author(s).
4th International Conference on Managing Pavements (1998)
Obtain results from an expert panel by visiting and evaluating at least the 5 major
categories of the OPC. The problem with this method is that it is very costly and
to obtain enough samples with the various combinations within each category is
an almost impossible task.
Use the existing ANN configuration and the existing complete database to
determine the VCI using an ANN. This is then updated annually with the addition
of new data, as they become available. In such a way more examples per
category can be identified. The disadvantage of this method is that the ANN
would become biased towards the Fair condition, as this is currently the condition
with the highest occurrence.
The latter method is recommended.
ACKNOWLEDGEMENTS
REFERENCES
TRB Committee AFD10 on Pavement Management Systems is providing the information contained herein for use by individual practitioners
in state and local transportation agencies, researchers in academic institutions, and other members of the transportation research
community. The information in this paper was taken directly from the submission of the author(s).
4th International Conference on Managing Pavements (1998)
APPENDIX 1
ROUTE/SECTION : ASSESSOR: .
DISTRICT:
CLIMATE: Very Wet Wet Moderate Dry TERRAIN: Mountains Rolling Flat
m: m: m: m:
ROAD WIDTH km: km: km: km:
DC:
SURFACING
TEXTURE VARYING FINE MEDIUM COARSE
CURRENT SURFACE VOIDS VARYING NONE FEW MANY
DEGREE EXTENT
Slight Severe Isolated Common
0 1 2 3 4 5 1 2 3 4 5
FAILURE/PATCHING
CRACKS
AGGREGATE LOSS
BINDER CONDITION
BLEEDING/FLUSHING
STRUCTURE DEGREE EXTENT
Small/Slight Large/Severe Isolated Common
0 1 2 3 4 5 1 2 3 4 5
BLOCK/STAB. CRACKS N M L
LONGITUDINAL/SLIP CRACKS
TRANSVERSE CRACKS
CROCODILE/FAILURE CRACKS
PUMPING
RUTTING
UNDULATION/SETTLEMENT
PATCHING
FAILURES/POTHOLING
OCCURRENCE STRUCTURES EDGES DEPRESSIONS GRADES CUTTINGS GENERAL
FUNCTIONAL
RIDING QUALITY VERY GOOD GOOD FAIR POOR VERY POOR
0 1 2 3 4 5 1 2 3 4 5
EDGE BREAKING
SUMMARY
TYPE ACTION NEEDED NONE STRUCTURE SURFACE ROUTINE
C B A C B A C B A
OVERALL PAVEMENT CONDITION VERY GOOD GOOD FAIR POOR VERY POOR
TRB Committee AFD10 on Pavement Management Systems is providing the information contained herein for use by individual practitioners
in state and local transportation agencies, researchers in academic institutions, and other members of the transportation research
community. The information in this paper was taken directly from the submission of the author(s).