Apart from my efforts, the success of this project depends largely on the
encouragement and guidelines of many others. I take this opportunity to express my
gratitude to the people who have been instrumental in the successful completion of this
project.
First and foremost, I would like to thank to my guide, Dr. Satish S. Chinchanikar
(Professor, Department of Mechanical Engineering, VIIT, Pune) for his valuable
guidance and advice. My special thanks to my Co-guide, Mr. Mahendra G. Gadge
(Assistant Professor, Department of Mechanical Engineering, VIIT, Pune) for his
valuable guidance and advice. He inspired me greatly to work in this project. His
willingness to motivate me contributed tremendously to this project.
I would like to thank Dr. Atul P. Kulkarni (Associate Professor and Head,
Department of Mechanical Engineering, VIIT, Pune) for allowing me to work on this
project and giving me valuable guidance and advice in my Project.
I am highly grateful to Dr. Bilavari S. Karkare (Principal) and I would also thank
our Institution, faculties and technical staff of mechanical engineering department who
helped me directly or indirectly during this project work. I also extend my heartfelt
thanks to my family, siblings, all my friends and well-wishers.
i
LIST OF FIGURE
ii
3.16 Average distance between consecutive generations (45HRC) 60
4.4 Feed forward neural network for AISI 4340 Hard turning 88
iii
4.6 Training state of Network at each epoch 89
iv
4.27 Training error in Tf 93
v
4.48 Validation Error Plots for Fr (Target vs Output) 107
vi
4.69 Regression Plots for Fa (Train /Test/Validate) 111
vii
4.90 Regression Plots for Ft (Train /Test/Validate) 118
viii
4.111 Response Surface Plot for Ra 122
ix
ANFIS FCM plots For 35 HRC
x
4.154 Testing Error Plots for Tf (Target vs Output) 134
xi
4.175 Validation Error Plots for Fr (Target vs Output) 138
FOR 35 HRC
For 45 HRC
xii
FOR 35 HRC
FOR 45 HRC
FOR 35 HRC
FOR 45 HRC
xiii
FOR 35 hrc
FOR 45 HRC
NSGA-NN 35HRC
xiv
5.13 Regression fit plot for Fr 152
xv
5.33 Regression fit plot for Ra 157
xvi
5.54 Response Surface Plot for Fa 163
xvii
5.75 Training Error Plots for Fr (Target vs Output) 169
xviii
5.96 Training Error Plots for Fr (Target vs Output) 175
xix
5.117 Testing Error Plots for Fr (Target vs Output 180
FOR 35 HRC
For 45 HRC
xx
5.135 Error Estimation Plots For F r 184
xxi
LIST OF TABLES
3.3 Results of NSGA II family of best solution for AISI 4340 35 HRC Steel 48
3.4 Results of NSGA II family of best solution for AISI 4340 45 HRC Steel 50
3.6 Results of SPEA II family of best solution for AISI 4340 35 HRC Steel 57
3.7 Results of SPEA II family of best solution for AISI 4340 45 HRC Steel 58
3.9 MOPSO family of optimal solutions for 35 HRC AISI 4340 steel 72
3.10 MOPSO family of optimal solutions for 35 HRC AISI 4340 steel 74
4.3.4(b) Statistical Results of ANFIS Grid Partioning Cluster for 35 HRC and 102
45HRC
4.3.16(b) Statistical Error analysis of ANFIS Subtractive clustering for 35 HRC 116
and 45 HRC
xxii
4.3.26 (a) Fuzzy structure 127
4.3.26(b) Statistical Results of ANFIS FCM for 35 HRC and 45HRC 129
5.2.1 (b) Calibrated weigths and bias for NSGA-NN 35 HRC and 45HRC Steel 150
5.3.2 (b) Calibrated weigths and bias for 35 HRC and 45HRCSteel 156
5.4.1 (a) Statistical Error analysis of GA and PSO based ANFIS (FCM) for 35 159
HRC and 45HRC
6.1 Tradeoffs among forces for Surface roughness and Tool life 190
6.2 Tradeoffs among forces for Surface roughness and Tool life 192
6.3 Tradeoffs among forces for Surface roughness and Tool life 193
6.4 Tradeoffs among forces for Surface roughness and Tool life 194
6.5 Mean error and Standard Deviation between Experimental and Predicted 201
statistics
6.6 Table.6.6 MSE and RMSE between Experimental and Predicted statistics 201
6.7 Mean error and Standard Deviation between Experimental and Predicted 205
statistics
6.8 Mean Square Error and Root Mean Square Error between Experimental 205
and Predicted statistics
xxiii
NOMENCLATURE
HRC Rockwell C Hardness
V, (vc) Cutting Speed
f Feed Rate
d Depth Of Cut
Ra Surface Roughness
Ft Tangential Force
Fa Axial Force
Fr Radial Force
Tf Tool Life
AI Artificial Intelligence
CI Computational Intelligence
HC Hard Computation
SC Soft Computing
FS Fuzzy Systems
NN Neural Network
EA Evolutionary Algorithm
GSO Global Search Optimization
SI Swarm Intelligence
GA Genetic Algorithm
DE Differential Evolution
CA Culture Algorithm
PSO Particle Swarm Optimization
FA Fire Fly Algorithm
BBO Biogeography Based Optimization
xxiv
W1/W2/W3 Weights
EA-NN Evolutionary Based Neural Network
NN-EA Neural Network Based Evolutionary Algorithm
NSGA Non-Sorted Domination Based Genetic Algorithm
PESA Pareto Envelope Based Selection Algorithm
SPEA Strength Pareto Based Evolutionary Algorithm
MF Membership Function
ANFIS Adaptive Neuro-Fuzzy Interface System
ACO Ant Colony Algorithm
HLGA Hybrid Adaptive Learning Based Genetic Algorithm
Pmt Probability of m bits in K sting length to Mutate
Sij Layers
Membership cluster vector
d ( xi , j ) Dissimilarity Function
J q ( , u) Cluster Function
xxv
CONTENTS
1 INTRODUCTION
1.1 Computational Intelligence (CI) 1
1.2 Approaches to Computational Intelligence 2
1.2.1 Fuzzy logic 4
1.2.2 Neural Network 4
1.2.3 Evolutionary Computing 4
1.2.4 Learning Theory 5
1.2.5 Probabilistic Methods 5
1.2.6 Swarm Intelligence 6
1.2.7 Global Search Optimization 6
1.3 Synergies of Computational Intelligence Techniques 6
1.4 Applications of Computational Intelligence 8
1.4.1 Application of NN 8
1.4.2 Application of Evolutionary Systems 9
1.4.3 Application of Fuzzy system 9
1.5 Overview of the chapter 10
1.6 Problem Statement 10
1.7 Objectives 10
1.8 Methodology 11
1.9 Thesis Outline 12
2 LITERATURE REVIEW
xxvi
2.1 Literature Review 14
2.2 Literature Summary 27
xxvii
3.7 Swarm Intelligence 60
3.8 Mathematical Formulation of PSO Algorithm 60
3.8.1 Typical Initialization strategy 62
3.8.2 Topologies of PSO 62
3.8.3 Definitions of Swarm topology 63
3.8.4 Convergence criteria 63
3.8.5 Criteria for Inertia clamping and acceleration co-efficeint 64
3.9 PSO Algorithm 66
3.9. Initialize population and Evaluate fitness 67
1
3.9. Create Grid Index 68
2
3.9. Select Leader 69
3
3.9. Delete extra elements 70
3.9.
4 Swarm Movement 71
6
3.9. Plots for MOPSO results (35 HRC) 73
73.9. Plots for MOPSO results (45 HRC) 75
3.10 8
Comparison between EA and SI technique 76
3.10. Comparison Based on Spectrum of solution space 77
1
3.10. Comparison Based on Diversity in solution space. 78
4 2
PREDICTION MODELS FOR MACHINING SYSTEM
THROUGH INTELLIGENT LEARNING TECHNIQUES
4.1 Introduction 80
4.2 Neural Network 81
4.2.1 Feed forward Neural network 81
4.2.2 Mathematical background of neural network 82
4.2.2.1 Gradient Descent Approach 83
4.2.3 Key notes form feed forward analysis 87
4.2.4 Multi-layer Perceptron for Turning of AISI 4340 Steel 87
4.2.5 Results of perceptron for 35 HRC Steel 89
4.2.6 Results of perceptron for 45 HRC Steel 92
4.3 Adaptive Neuro-Fuzzy Interference System (ANFIS) 94
4.3.1 Hybrid learning in ANFIS 96
xxviii
4.3.2 Backpropogation Learning 97
xxix
4.3.33 ANFIS FCM Cluster Plots For Ft 45 HRC 136
xxx
5.4 Synergies of EA and ANFIS 158
5.4.1 ANFIS GA 159
5.4.2 GA based ANFIS (FCM) Plots For Ra 35 HRC 169
5.4.3 GA based ANFIS (FCM) Plots For Ft 35 HRC 161
5.4.4 GA based ANFIS (FCM) Plots For Fa 35 HRC 162
5.4.5 GA based ANFIS (FCM) Plots For Fr 35 HRC 163
5.4.6 GA based ANFIS (FCM) Plots For Tf 35 HRC 164
5.4.7 GA based ANFIS (FCM) Plots For Ra 45 HRC 166
5.4.8 GA based ANFIS (FCM) Plots For Ft 45 HRC 167
5.4.9 GA based ANFIS (FCM) Plots For Fa 45 HRC 168
5.4.10 GA based ANFIS (FCM) Plots For Fr 45 HRC 169
5.4.11 GA based ANFIS (FCM) Plots For Tf 45 HRC 170
5.5 PSO based ANFIS (FCM) 35 HRC and 45HRC 171
5.5.1 PSO based ANFIS (FCM) 171
5.5.2 PSO based ANFIS (FCM) Plots For Ra 35 HRC 172
5.5.3 PSO based ANFIS (FCM) Plots For Ft 35 HRC 173
5.5.4 PSO based ANFIS (FCM) Plots For Fa 35 HRC 174
5.5.5 PSO based ANFIS (FCM) Plots For Fr 35 HRC 175
5.5.6 PSO based ANFIS (FCM) Plots For Tf 35 HRC 176
5.5.8 PSO based ANFIS (FCM) Plots For Ra 45 HRC 177
5.5.9 PSO based ANFIS (FCM) Plots For Ft 45 HRC 178
5.5.10 PSO based ANFIS (FCM) Plots For Fa 45 HRC 179
xxxi
5.6.6 Error Plots of GA based ANFIS (Fuzzy C-Mean Clustering)) Results with 185
Experimental Statistics for AISI 4340 Steel 45HRC
5.6.7 Error Plots of PSO based ANFIS (Fuzzy C-Mean Clustering)) Results with 186
Experimental Statistics for AISI 4340 Steel 35HRC
5.6.8 Error Plots of PSO based ANFIS Prediction Results with Experimental 186
Statistics for AISI 4340 Steel 45HRC
5.7 Conclusion 187
xxxii
6.3.3 ANFIS Synergies 204
6.3.4 Comparative Evaluation of the predictive technique on Experimental statistics 204
7 CONCLUSION 207
Appendix- C: Certificates
xxxiii
ABSTRACT
xxxiv
Evolutionary Algorithms for Multi-Objective Optimization: Modelling and Comparative Evaluation
Chapter-1
1. Introduction
The complexity of design optimization of any dynamic system has many aspects among
them the major facets are the ambiguity of objectives, conflicting nature of objectives and
many possible solutions this brings an issue in characterizing the difficulty of design
optimization task. Any design solution has combination of values for parameters of a
solution and the challenge lies in identifying the solution. Considering second issue
functionality of obtained solution, the solution should be practical enough, look appealing
and have moderate cost. Third issue is contributing to the ambiguity of design
optimization is conflicting objectives which inhibits unidirectional solution. To tackle
these complexity several attempts are made through conventional methods but the
solution in general are partially satisfactory To address this issue new computational
approaches are followed which has multi-agent system each agent is defined by its
behavior that are classified into various categories in Computational Intelligence.
Computational Intelligence deals with the design of intelligent agents which act
intelligently for goal attainment in any circumstances, flexible enough to adopt changing
environment and goals. Computational Intelligence has the ability to comprehend reason,
learn, and simulate intelligent behavior in systems for complete knowledge formulation.
Much real-time system behavior cannot be captured exactly through classical
mathematical description in spite of complex formulations; moreover complexity of
mathematical description inhibits development of system model. Hence it is really
advantageous to model a real time system with piece-wise linearity and non-linearity so
that the highly complex and un-anticipatory models can be captured by intelligent agents.
Any real time problem has uncertainties involved in it with multiple objectives and the
risk in decision making should be such that the performance criteria are maintained even
in drastic change, this necessity of capturing the dynamic behavior of system is replacing
conventional techniques with intelligent techniques. Computational Intelligent techniques
are thus an illustration of alternative methods to conventional technique when the system
knowledge is highly important in system modeling and control. The structure of such
systems is determined by experimental evidence where direct input-output response
behavior is utilized to develop system model. Intelligent systems are meant for the
processes that are not properly defined, complex and stochastic in nature, time varying.
The fundamental property of any intelligent system is that it must sense and reason
without prior knowledge about the environment and adapt to the control action in a robust
manner. Many attempts have been made to define by different researchers but the
property of a system to be computationally intelligent is if it deals with numerical data
and has ability of pattern recognition. CI is a subdivision of machine intelligence where
subtle difference between the techniques lies in the type of computing. Machine
intelligence has two constituents Artificial Intelligence based on hard computing (HC)
and Computational Intelligence based on soft computing (SC) [Fuzzy Sets (FS), Neural
Network (NN), Evolutionary Algorithms(EA)] (Fig. 1.1) distinguishes clearly the
components of Machine intelligence and their components.
The core of the computational intelligence is designing process or system model which is
not responsive to mathematical modeling since the process exhibits following attributes
The system is capable of learning to adapt to unknown situations and is able to make
predictions about the process status in future time step. CI is a combination of soft
computing and numerical technique with methods involving adaptive control, (Fig. 1.2)
optimal control, learning theory, fuzzy logic, neural-network, evolutionary computing.
All methods tuned to attain common goal set. There are five elemental methods to CI
Fuzzy logic
Neural network
Evolutionary computing
Learning theory
Probabilistic methods
Swarm Intelligence
In any real-time process, the measurement, process modeling, and control can never be
exact to the theoretical definitions [1]. There is always a certain amount of uncertainties
i.e., incompleteness, randomness of data. Fuzzy assimilates human experimental
knowledge converts it into engineering model and control. Most process which are ill-
defined with nonlinearity and uncertainties. The fuzzy logic is more of reasoning and
inference technique based on high level linguistic or semantic rules and operations.
Neural network neural network is a technique adopted from the biological brain which
involves a neuron as a fundamental building block [1]. These neurons receive signals
from neighboring neurons through their cell body and transfer the results through a long
fiber called an axon. The axons behave like signal conducting device. A similar electrical
analogy of biological neural network is artificial neural network which is characterized
by computational power learning of real-time data error tolerance, pattern recognition,
and generalization capabilities, low-level computational algorithms which manifest good
performance in numerical data processing. The learning is in different form supervised,
unsupervised, and competitive and reinforcement learning.
Learning theory is based on the human learning capabilities without much effort in a
conventional sense. The mechanism of learning in humans is the process of bringing
together cognitive, emotional and environmental effects to acquire enhance or change
knowledge, skills. In general, learning is characterized by how the information is input,
processed and stored. Learning theories fall into three framework behaviorism cognitive
theories and constructivism. Behaviorism is learning based on objectively observable
feature learning Cognitive learning is how learning occurs in brain. Constructivism
learning is a process in which permutation of existing idea builds a new idea. In most of
the machine learning four basic forms of learning si adopted i.e., supervised learning
where a mapping of input to desired output is done, unsupervised learning where a set of
input feature is modeled and mapping of input to output is done with similar pattern.
Semi-supervised learning, combination of both learned and unlearned datasets are used to
generate an appropriate classifier. Reinforcement learning involves decision making on
given observation and feedback is taken from the consequence to supervise the learning
process [1].
Probabilistic theory is methodology which guides in dealing with the uncertainties and
imprecisions. The probabilistic methods involve a space consisting probabilities of whole
system. The uncertainties of complex dynamic system are calculated and combined
behavior of system is analyzed for the degree of causticity. The chaotic behavior of
system is estimated by the past. In general, the chaotic behavior of system grows
exponentially with time [1].
Swarm systems are based on behavior of school of birds, insects, fireflies where a flocks
of birds twisting, V-shaped structure of migrating geese, winter birds hunting for food,
the synchronized flashing of fireflies are tried to imitating. The well-choreographed
collective behavior without any leader is adopted to search for optimal solutions for
instance, ants living in colony, their behavior is driven by the goal of colony survival
instead of individual survival, While searching for food ants initially explores
surrounding nests in random manner. A similar behavior is observed with flocks of birds
where a leader keeps guiding the flock to updated food location [1].
Both EA and SI together form a broader class of optimization driven search techniques
defined as global search optimization technique as shown in Fig.1.3 below.
The different combination of all the methodologies can be used to design intelligent
systems. Though a particular technique might be excellent in approximate reasoning and
modeling uncertainty but may not be so good at learning and adopting with experimental
data. A combined approach with computational intelligence technique and their
implementation can help in designing better intelligent agents.
Different forms of synergisms (Fig. 1.4) of fuzzy logic, neural networks and evolutionary
algorithms the common forms of weakly coupled synergism of neural network and
evolutionary algorithms include training designing, optimizing architecture and
parameters of neural networks and feature selection scaling training data for neural
network using evolutionary algorithms. A strongly coupled synergism between the two
methodologies where genetic operators are represented in the form of neural network and
the epochs are meant to be the generations of evolutions [1].
Synergisms of neural networks and fuzzy systems have proven to be very powerful for
system modeling and learning. In weakly coupled synergism, neural networks and fuzzy
system work independently towards a common goal where neural network assist fuzzy
logic to form rules and tuning membership functions. In strongly coupled synergism
fuzzy system assist neurons to assign weights to its membership functions where neural
network learns data over the epoch. There is other synergism possible between swarm
intelligence, fuzzy systems, evolutionary algorithms and neural network.
Neural network can be applied in five ways i.e., data analysis, classifier, clustering,
pattern recognition, control strategy neural network has been successfully applied in
problems behaving non-linearly whereas fuzzy logic has been applied to appliances
where a module control is required the most adaptive implementation is done on
stabilizing an unsteady image. Fuzzy expert systems are applied to medical systems,
diagnostic, scheduling, and financial systems.
1.4.1 Application of NN
1. In aerospace neural networks are applied to high performance autopilot flight path
simulation, aircraft control systems, fault detection system.
4. In defense, it is used for weapon steering, target tracking, object discrimination, facial
recognition.
6. In Medical cancer cell analysis, EEG and ECG signal analysis, optimization of
transplantation.
7. In speech, applied for speech recognition, compression, and text to speech synthesis.
8. Telecommunication applied for image and data compression, speech processing, real-
time translation of spoken language.
1. Fuzzy systems are used in automobiles and vehicle subsystems such as automatic
transmissions, ABS and cruise control.
3. In digital image processing, such as edge detection and video gam artificial
intelligence.
In the present work different soft computing technique is applied over machining system
to optimize machining performance and recognize machining pattern with a case study
from literature [2-9] in which extensive machinability aspects of AISI 4340 alloy steel
with different machining characteristics during, hard turning. Operation with different
faceted is discussed.
1.7 Objectives
In the present work, an attempt is made to apply the Computational Techniques and their
synergies with the objective to optimize and build prediction models for conventional
machining system. Different Computational methods are applied over the machining
system for optimization and learning of machining data Further, the optimized and learnt
data are compared with the result obtained from the literature
1.8 Methodology
The adapted methodology is developed (Fig. 1.5) to achieve the above mentioned
objective with focus to optimize and develop prediction models.
In Chapter 4, applied predictive models i.e., neural network and adaptive neuro fuzzy
based learning model are discussed. Mathematical aspects and pseudo code along with
their results is discussed.
CHAPTER-2
The literature was explored with an aim to gather research work of authors who utilized
different evolutionary and learning techniques. In the literature authors have utilized
optimization and predictive techniques exclusively; multi-objective machining systems
were converted to single objective for performance evaluation. Many authors have
optimized machining systems with unit control parameter to enhance machining
performance avoiding the complex nature of conflicting objective. Literature review was
emphasized on the process parameters and control parameters utilized to model
machining system, applied optimization techniques and their degree of accuracy in
comparison to conventional techniques.
Hesam et al. [11] executed EDM process on DIN 1.452 stainless steel in which surface
roughness and white layers were control parameters. The machining model was built on
Taguchi technique and NSGA II was applied for optimization which could produce
convincing results.
Garg et al. [12] improvised machining turning operation of AISI 1040 Steel with surface
roughness as a control parameter. Taguchi technique was applied to model surface
roughness apart from that (Artificial Neural Network) ANN, (Support Vector Regression)
SVR techniques were used to build regression. Genetic programming (C-GP) coupled
with classifier was used as optimization technique. The results suggested that C-GP was
on par with the ANN but SVR performed poorer than C-GP and ANN.
Khaider et al. [13] examined hard turning operation of AISI 52100 bearing steel with
CBN (7020). The machining performance was measured in surface roughness, tool wear
and material removal rate were modeled with Taguchi, RSM and grey-relation, these
models were utilizes to optimize performance by applying GA and the results obtained
from GA predicted parameters which gave better machining performance.
Ozel and Karpat [14] investigated on enhancing the performance of AISI H13 grade
steel turning with CBN tool. Prediction model for surface roughness and tool wear was
built on process parameters (cutting speed, feed rate, depth of cut). Experimental data of
AISI52100 steel was referred form literature and further data experiment was performed
on AISI H-13 steel these data was utilized to train neural network, Regression was also
carried out, two feed-forward neural networks was modeled. In the first model the input
layer constituted edge geometry, hardness, cutting speed, feed rate and depth of cut to
predict tool wear and surface roughness, while in the latter network material hardness,
cutting speed, feed rate, depth of cut and forces were utilized to model tool wear and
surface roughness. The latter network performed well than the former.
Alhameri et al. [15] studied multi pass turning of austenitic AISI 302 Steel. Box-Behken
design was utilized to develop model. Prediction models were also built by Regression
analysis and NN to predict Tool life and machining economics with motive to minimize
machine economics and maximize tool life.
Abbas et al. [16] carried out research on turning operation of J steel with Tungsten-
Carbide insert. Models were built to predict Surface roughness and Material removal rate
by applying Taguchi technique. The formulated regression equations were utilized as
objectives with appropriate constraints in process parameters. Multi-objective EGO
algorithm was implemented to optimize machining performance.
Zhenghua et al [17] investigated high speed milling aluminum alloy AlMN1CU with
Carbide tipped tool. Both Linear and quadratic regression models were built and
Bayesian Neural nets (BNN) was built using experimental data to predict surface
roughness, the regression model built was utilized as objective functions with precise
Yunguang et al. [18] worked on micro-grinding of nickel based super alloy (DD98),
surface roughness was modeled in linear and non-linear degree using CCD based RSM.
GA was applied to predict control parameters for best machining performance. The
results were verified experimentally and were found to show pretty good agreement.
N.Alberti and Perrone [20] worked on multi pass turning operation to predict least
power consumption, machine economics and surface roughness for which three different
modelling approaches was adopted viz., deterministic model, possibilistic model, a fuzzy
possibilistic-GA model with constrained and unconstrained search space. The results
established that fuzzy-possibilistic model predicted most failures and fuzzy-possibilistic-
GA optimized objectives to practically feasible solution.
Garge et al. [21] experimented on EDM of Titanium and Inconel alloy in which surface
roughness and cutting speed were control parameters modeled with process parameters.
NSGA II was applied to optimize performance.
Pramanic et al. [22] worked on EDM of ZrB2 where cutting speed, material removal
rate, and surface roughness was modeled with process parameters by applying Taguchi
technique and optimization based on Taguchi based grey relation. ANN was used to
predict cutting speed and surface roughness, the predicted accuracy was checked with
experimental statistics for confidence level and it gave the appreciative result.
Sahali et al. [23] worked on multi point turning operation, modeling machine economics
with process parameters and constrains in surface roughness, chip-tool temperature, tool
life and force was applied. Optimization techniques applied were viz., Deterministic
Dureja et al. [24] reviewed different optimization and modeling techniques used in hard
turning operations viz., RSM, Taguchi, Regression analysis, NN, Fuzzy modeling, GA.
Sundaraman et al. [27] in contrast worked on fixture design and layout of end milling;
quadratic model was built using RSM and optimization done by GA and PSO. The model
was built to predict work piece deformation with positions of clamp and location as
parameters. The results of optimization suggested that RSM-PSO gave better solution
than RSM-GA technique. Furthermore, these results were compared with FEM
simulation of fixture layout.
Costa [28] investigated on multi-pass turning with the objective to minimize unit
production cost. The objective was constituted of actual machining cost, machining idle
cost, tool replacement cost. The characteristic equation was built on process parameters
viz., cutting speed, depth of cut, feed rate both in rough and finish pass, operation
constraints were applied in tool life, cutting forces, power and surface roughness. A novel
hybrid technique in PSO was formulated for optimization. Furthermore, this solution was
compared with other techniques suggested 2.035 unit production cost, while FEGA gave
2.3057 as optimal cost SA gave 2.29, MGA gave 2.30, HC gave 2.27 and ACO gave
2.25.The hybrid PSO technique could optimize solution superior than other techniques.
Bharathi et al. [29] investigated on turning operation with diamond shape tungsten
carbide tool on four different materials i.e., brass, aluminium, copper and mild steel.
Machining was characterized by forces, surface roughness with cutting speed, feed rate,
and depth of cut as process parameters. Furthermore, these equations were optimized
using PSO technique; the optimal solution suggested the trend that higher cutting speed,
lower feed, and depth of cut gave better surface roughness. The optimal surface
roughness obtained from PSO for brass was 0.07m, for copper 0.08m, for aluminium
0.08m and mild steel 0.08m.
Bharathi et al. [30] carried out investigation on milling operation of aluminum bar with
carbide tool. Machining time and surface roughness were characterized with process
parameters such as spindle speed, feed rate, and depth of cut. The characterized equations
were optimized using PSO technique. The solution obtained from PSO was verified by
conducting confirmation test. The solution trend showed that higher speed, lower depth
of cut, lower feed rate gave better surface roughness and feed rate had a greater influence
on surface roughness. The prediction ability of present approach was found to be 96 %
for machining time and 85% for surface roughness.
Bharathi et al. [31] investigated on modeling and optimizing both turning and grinding
operation. The turning operation was done on single and multi-pass while grinding was
done in the single-pass. The performance of turning was measured on machining time
while grinding was done on machining time and material removal rate. Optimization
technique applied were PSO, GA, and SA whose optimal solutions were
comprehensively evaluated. The computational time obtained by PSO in single and
multi-pass turning was 11 sec and 12 sec respectively, while for grinding 4 sec. Similarly,
results of GA gave 15 sec in both single and multi- pass turning, while in grinding it gave
6 sec as optimal computational time. Likewise in SA, it was 12, 13 and 5 respectively.
Optimal material removal rate in grinding was in the range of 0.17-0.44 m, from the
results it could be inferred that PSO proved to be better than GA and SA.
Chandrasen et al. [32] reviewed different soft computing techniques that could be
applied to machining performance prediction. Any machining system could be
generalized by its corresponding inputs and outputs. Input in general are process
parameters, material properties, sensory feed and output of the system are concerned
about machining performance i.e., dimensional deviation, cutting forces and tool wear,
after machining characterizations done various soft computing techniques are applied to
optimize machine model. The review concluded that best strategy to predict performance
is to couple fuzzy with a neural network. Likewise, to optimize precisely GA, PSO and
similar heuristic techniques are the best technique.
Prabhakaran et al. [33] carried out work on machining fixture analysis where location
and displacement of clamp and locator were objective functions. Regression models were
developed for displacement and location and optimized using GA and ACA. Ant colony
algorithm gave near optimal solution than GA.
Farahnakian et al. [34] investigated end milling operation; performance was modeled in
cutting force and surface roughness with process parameters as cutting speed, depth of
cut, feed rate. The characterized equation was utilized to frame optimization problem,
coupled PSO-NN technique was applied to optimize. The applied technique gave better
Pareto-spread in solution space with good convergence.
Yang et al. [35] carried out worked on multi- pass face milling operation. Performance
was characterized by unit production cost with process parameters such as number of
pass, depth of cut, cutting speed, feed rate. Fuzzy based multi-objective PSO was applied
to optimize process parameters which gave better solution with fast convergence.
Sukla and Singh [39] investigated on Abrasive Water jet Machining (AJM) of
Aluminum alloy with garnet abrasive particles. Machining model was built for kerf width
and taper angle prediction with process parameters by applying Taguchi technique and
Optimization techniques applied, PSO, Firefly, Simulated Annealing, Black Hole, Bio-
Geographical, NSGA. PSO gave better results than other techniques.
Senthil et al. [41] predicted performance of cutting tool inserts using neural network.
Experiments were performed on workpiece with carbide inserts with process parameters
such as cutting speed, feed rate, depth of cut, material hardness, and cutting insert shape
(relief angle, nose radius) to model surface roughness and flank wear. The Taguchi based
ANN model was built with these process parameters as input layer. Results predicted by
neural network model were compared with experimental values which predicted values
close to experimental statistics.
Miron et al. [42] worked on dynamic characterization and vibration analysis of lathe
machining system by which the machine condition was determined. Modal analysis was
done to determine the natural frequencies, the frequency was compared with numerical
model and validation experiment was performed.
Dilbag and Venkateshwara [43] developed analytical tool wear model while turning
bearing steel with ceramic tool. The model incorporated abrasion, adhesion and diffusion
wear mechanism further it was validated by conducting experiment. The analytical model
had capability of predicting flank wear using cutting parameter and tool geometry.
Yahya et al. [44] worked on turning operation of steel at different conditions with P25
HSS tool at different working conditions. Surface roughness, flank wear, and crater wear
were modeled with process parameters to determine machinability of tool steel. The
relative degree of influence of each parameter on control parameters was quantified. This
work can help in sorting the priority of objective functions and the contribution to overall
machining performance.
Hamdi et al. [45] investigated behaviour of hard turning while machining AISI H11
Steel with CBN tool. Forces and surface profile were considered as process responses. A
CCD based RSM was applied to build machining model furthermore a comprehensive
analysis was done on influence of process parameters over machining quality.
Shihab et al. [46] conducted experiment on hard turning of AISI 52100 Steel alloy with
coated carbide tool in which surface roughness and micro hardness were modelled and
optimized utilizing CCD based RSM approach. The RSM based optimization technique
gave satisfactory results but by reducing multi objective to single objective.
Waleed et al. [47] worked on hard turning of AISI 4340 Steel with CBN tool; in his
work surface roughness and tool flank wear were modelled by Taguchi technique to form
multi-regression equation. This equation was used as objective functions along with
constraints in process parameters; the S/N ratio analysis was done on regression to
optimize control parameters.
brass wire and zinc coated brass wire by applying both RSM and multi criteria grey
relation. RSM was utilized to optimize the objectives individually for both the wires.
Emeryl et al. [49] worked on hard turning of Ni-Steel alloy (62HRC) with CBN tool
insert, the performance was modeled to predict cutting forces, surface roughness with
process parameters by applying Taguchi method and optimization was done by Taguchi
based S/N ratio. The results were in agreement with experimental data with good level of
confidence.
Ilhan and asks [50] worked on hard turning of AISI 4140 (51HRC) steel by carbide tool
coated with Al2O3 and TiC. A three-level full factorial with Taguchi based experimental
design was applied to model surface roughness by applying cutting conditions and
control factors. The process variability was measured by S/N ratio. Taguchi based S/N
response suggested that larger difference in S/N ratio will have more significant effect on
surface roughness. The result of process variables for optimum surface roughness was
120 m/min, 0.18 mm/rev and 0.4 mm cutting speed, feed rate and depth of cut
respectively.
Gaurav and Choudhary [51] focused study on hard turning of EN31 bearing steel (58-
62HRC) with CBN tool insert. A three-level full factorial experimental design was
developed; ANOVA was performed to find the relative contribution. RSM was utilized to
build regression equation on cutting forces and surface roughness, further RSM
optimization was done. Results showed that depth of cut had more influence on cutting
forces and while speed had the least influence. Results also revealed that initially forces
decreased with increasing speed later on increased along with speed due thermal
softening of tool material.
Ashvin and Nanavati [52] enquired on turning operation of AISI 410 steel with carbide
inserts of TNMG series differing in nose radius. A three-level full factorial experimental
design was done, further, RSM was utilized to model and optimize surface roughness
which suggested optimal solution 225 m/min,0.1 mm/rev, 0.3 mm, 0.12 mm for cutting
speed, feed rate, depth of cut and tool radius respectively.
Asilturk and Suleyman [53] investigated hard turning of AISI 304 austenite stainless
steel with carbide inserts (SNMG series). A three-level full factorial Taguchi based
experimental design was built. RSM based regression equation was modeled for surface
parameters (Ra and Rz) using process variable, S/N ratio was determined then RSM
based optimization was done. The optimized control factors setting for Ra was found to
be cutting speed 50 m/min, feed rate 0.15 mm/rev, depth of cut 1.5mm and for Rz was
cutting speed 150m/min, feed rate 0.15mm/rev, depth of cut 1mm. These authors applied
Taguchi and RSM to model and optimize machining parameters. RSM relates response
based input parameters by experimental statistics and applying regression. RSM consists
of three stages design of experiment, regression, and optimization to find best optimal
solution RSM is coupled with meta-heuristic technique.
Aggarwal and Singh [54] reviewed different machining modeling technique for
conventional machining model and types of optimization methodology to optimize and
characterize machine model.
Chinmaya et al. [55] experimented on hybrid machining where laser assisted machining
(LAM) was coupled with turning operation. High strength alloy (Ti-6Al-4V) was
machined with cobalt bound tungsten carbide, liquid nitrogen was used coolant. The
LAM hybrid turning operation reduced specific cutting energy and improved surface
roughness when compared to conventional machining.
Wang et al. [56] worked on multi-pass turning operation of AISI 1045 with a different
set of tools (TNMG carbide inserts). A hybrid model was built to predict machining
performance, surface roughness, forces, and chip breakability were characterized by
process parameters with operation constraints such as surface roughness, forces, and tool
life. RSM was utilized to optimize process parameter. The hybrid model developed could
predict slip line field accurately, which was verified by Finite Element Modelling results.
it was found that weight forces have maximum impact on the cutting forces. A quadratic
model was developed using RSM. Optimal solution concluded that cutting forces was
majorly affected by the type of reinforcement.
EranAlsan et al. [58] investigated hard turning of AISI 4140 steel with ceramic tool
mixed with Al2O3 and TiCN. Machining performance was modeled using Taguchi based
RSM on flank wear and surface roughness as control parameters with process variables.
The optimization results obtained from RSM technique suggested cutting speed
250m/min, feed rate 0.1mm and depth of cut 0.25-0.4 mm for surface roughness and
flank wear.
Hashimoto et al. [59] identified the fundamental difference in the surface integrity of
hard turning and ground surface, their subsequent impact on rolling contact fatigue life.
The work concluded that the mechanical deformation could play a large role during hard
turning than grinding while the size effect in grinding introduced surface hardness
furthermore the hard turned surface may have more than 100% longer fatigue life than a
ground one with an equivalent surface finish due to very different characterization of
surface integrity. The effect of turned or ground surface free of white layer was clarified a
super finished turned surface may have twice a fatigue life than ground surface.
Ozel et al. [60] investigated on hard turning of AISI 4340 steel with uniform and variable
edge PCBN insert where the forces and tool wear was measured and 3D finite model was
utilized to predict chip formation, temperature and tool wear on both type of inserts the
predicted tool wear and forces were compared with experimentation. The result showed
that the variable edge tool insert has advantage of less tool wear and good temperature
distribution profile.
Ravinder and Santram [61] investigated the effects of cutting parameters on surface
roughness in turning of Al7075 hard ceramic composites and Al7075 hybrid composite
using polycrystalline diamond tool (PCD) dry turning was conducted to examine the trend
of roughness by using roughness tester for both composites. It was concluded that surface
roughness of hybrid composite was lesser in all combination of experiment. Further
RSM based artificial neural network was applied to validate the results obtained during
experimentation and to protect the behavior of the system under any condition within the
operation range.
Mia and Dhar [62] developed a predictive model of average tool-workpiece interface in
hard turning of AISI 1060 steel by coated-carbide insert. Cutting condition used were
cutting speed, feed rate, and depth of cut utilized to model temperature profile.
Experiment was conducted in both dry and high pressure coolant environment with full-
factorial design. Temperature was measured using tool-work thermocouple. Response
Surface Methodology (RSM) and Artificial Neural Network (ANN) were employed to
predict the temperature. The accuracy of both RSM and ANN model were in region of
acceptance. The regression coefficient of ANN for both the environment was greater than
99.8%. ANN model demonstrated a higher accuracy which was found convincing if
employed for controlling cutting temperature in turning of hardened steel.
Pontes et al. [63] worked on turning of AISI52100 hardened steel with multi-layered
coating (Al2O3+Tic+TiN) chamfer edge. Experiments were conducted with training sets
of different size to compare performance of best network in each experiment. Process
parameters considered were cutting speed, feed rate, depth of cut to model performance in
surface roughness. Radial base function (RBF) neural network was developed with the
use of Taguchis orthogonal array as a tool to design parameter of network. The factors
considered in designing RBF-NN were number of radial units, algorithm for selection of
radial center and algorithm for selection of spread factor for evaluating performance of
RBF-NN. The results revealed that algorithm for calculation of radial spread factor was
most influencing among the three factors RBF-NN trained gave least mean standard
deviation for worst trained case. The results suggested that DOE based RBF network are
more efficient and effective than trail and error based NN architect.
roughness and tool wear. Statistical comparison was done between the predicted results
and experimental results further more interaction effects among the process parameters
were studied.
Wang [65] developed neural network based optimal estimator for predicting CBN tool
wear during hard turning operation. Prediction model was based on fully forward
connected neural network, with inputs as cutting condition and machining time and
predicted output in tool flank wear. The feed forward fully connected neural network
(FFCNN) based optimization was validated with experimental data. Comparison showed
that the (FFCNN) estimated a close value to experimental tool wear and developed
FFCNN model was found to be faster, accurate than other neural network approaches.
Umbrello et al. [66] developed predictive hybrid model based on neural network and
finite element method with objective to predict residual stress profile in hard turning for
different combination of material properties, cutting tool geometry and cutting condition.
A converse prediction of cutting conditions and geometry was made for a given residual
stress profile which acted as constrained based process parameters determination.
Furthermore, this model was utilized as closed feedback where the predicted residual
stress of ANN were applied to simulate cutting condition in FEA and vice-versa The
results obtained from ANN based FE simulation gave practical results.
Ravinder and Santram [67] investigated the effect of cutting parameters (cutting speed,
feed rate and approach angle) on roughness while turning Al 7075 hard ceramic based
composite using polycrystalline tool diamond tool (PCD). The surface roughness was
modeled by both RSM and ANN. Moreover, the influence of parameters on surface
roughness was analyzed both RSM and ANN model correlated fairly to the experimental
data.
was compared with PSO, HEGA, Scattered Search (SS), Simulated Annealing (SA),
Pattern Search (PS), Floating encoding genetic algorithm (FEGA) and Hybrid Harmony
Search (HSS) HEDEA outperformed all other techniques.
Kara et al. [69] worked on turning of AISI 310L Stainless steel with both coated and
uncoated cutting tool (TiCN+Al2O3+TiN). The cutting condition (cutting speed, feed rate,
and depth of cut) were used to model tangential forces and feed force Prediction model
was developed for both the responses with ANN. Two learning methods were deployed
i.e., scaled conjugate learning and Lavenberg-Madquart learning. The predicted forces
were accurate with error within 5%.
Sener Karabulut [70] worked on milling of metal matrix composite (Aluminum Alloy
7039/Al2O3 powder metallurgy) with CVD carbide tools the process parameters were
material removal rate, cutting speed, feed rate and axial depth of cut to model machining
performance such as surface roughness and cutting force ANN model was developed
with cutting condition to predict performance. The predicted performance was compared
with experimental model which gave close results with 99.8% regression.
CHAPTER 3
3.1Introduction
Most of the practical problems are complex and their definition of optimality is not
simple as they need to satisfy multiple competing objective functions at the same time.
Moreover, some of these objectives may have conflicting relations with others, which
makes the optimization difficult. Problems requiring simultaneous optimization of more
than one objective function are known as multi-objective optimization problems
(MOOPs). They can be defined as problems consisting of multiple objectives, which are
to be minimized or maximized while maintaining some constraints. Formally, they can be
defined as:
g j ( x) 0 j 1, 2,3... j
hk ( x) 0, k 1,2, 3,..., K
equality constraints. This type of problem has no unique perfect solution. In traditional
multi-objective optimization, it is very common to simply combine all the objectives
together to form a single (scalar) fitness function. But the obtained solution using a single
scalar is sensitive to the weight vector used in the scaling process. This requires
knowledge about the underlying problem which is not known before in most cases.
Moreover, the objectives can interact or conflict with each other. Therefore, trade-offs
exists when dealing with such MOOPs, rather than a single solution. Most MOOPs do not
provide a single solution; rather, they offer a set of solutions. Such solutions are the
trade-offs or good compromises among the objectives. In order to generate these trade-
off solutions, an old notion of optimality called the Pareto-optimum set is normally
adopted.
maintain as diverse a set of solutions as possible. The first task ensures that the obtained
set of solutions is near optimal, while the second task ensures that a wide range of trade-
off solutions is obtained.
The Machining data from hard turning of AISI 4340 steel [2] is utilized as machining
objectives in which machining is performed on two different hardness and regression
equations were built using RSM with process variables involving in cutting speed, feed
rate and depth of cut to model surface roughness, cutting forces and tool life. The
machining constraints and objectives are as follows,
Depth of cut(mm) 1 2 1 2
3.3.1 Machining Model (Surface Roughness Cutting force components and Tool
Life work material hardness: 35 HRC).
3.3.2 Machining Model (Surface Roughness Cutting force components and Tool Life work
material hardness: 45 HRC).
423
Tf 3.9
(v) 0.59
( f )0.4697 (d )0.47
23135.13
Tf 3.10
(v) 0.59
( f )0.4697 (d )0.47
Optimum seeking is one of the central issue in Manufacturing system. Every problem
solved is outcome of best possible choice for which a variety of tools and techniques
have been developed and applied to systems for optimum seeking.
Meanwhile optimum seeking in nature, biological and social systems takes place in a
completely different way i.e., natural evolution they have adapted themselves to a
constantly shifting and changing environment in order to survive. Those weaker and
lesser fit members of species tend to die away leaving create stronger and fitter to mate
create to create offspring and ensure the containing survival of species and it is upon this
dictated idea that evolutionary computing is based on. Evolutionary computing is
emulation of the process of natural selection in search procedure (as shown in Fig.3.1).
( p 1) [ ( ( p))]
Fig 3.2 Evolutionary Model
The current EAs applied in Multi-Objective Optimization Problem (MOOPs) and the
combination became known as a multi-objective evolutionary algorithm (MOEA). An
MOEA will be considered good only if both the goals of convergence and diversity are
satisfied simultaneously. The MOEAs population-based approach helps to preserve and
utilize the non-dominated diverse set of solutions in a population. The MOEA converges
to a Pareto-optimal front with a good spread of solutions in some fixed number of
generations. Most MOEAs use the concept of domination to attain the set of Pareto-
optimal solutions.
EA system which does not use any specific properties of the set X, the only condition is
that the function f should be defined at every point of X i.e, X R n . Then the problem
of optimization can be defined as,
f x 0 X Rn
There are Variety of Evolutionary system with different types of selection, crossover and
mutation. This section discusses most generalized terms with no bias in different genetic
operation.
Consider the following events in evolutionary system with their respective properties let
P be a population and n be the size of population
P A | H Probability that the population does not contain solution after mutation
P A | H Probability that the no solution is found after mutation and crossover
P H Probability that solution will be found after a crossover.
_
P H No solution after crossover.
S is a finite S 2m
Let Pmt be the probability of mutation s S be selected individual, then the probability
*
of individual not getting mutated be P S x mut PD and not giving rise S * in the
Let K be the no. of different bits in S & S * 0 k m then the probability that S
P P , P PS S 1 P
k
m
k
mt
m
k
Pm Pmt
1 k
k m
P 1 Pmt : Pmk is The probability of K mutation
mk
Cm
1 Pmt
mk
is the probability that the remaining (m-k ) bits do not mutate
1
probability that precisely the necessary K-bits mutate but not other bits.
Cmk
m
pmt
Then P{s s ' s*} 1 P{s s*} 1 is the probability that s does not mutate
Cm[ m/2]
in S*
Pmtm
P S S *
1 k
Cm
mk 1 m
k Pmt 1 Pmt k Pmt ( m /2)
Cm Cm
For an individual to get mutated the minimum condition is that half of its bit length must
participate in mutation and in contradictory assuming that mutation does not happen then,
P S * mut P P{ S S ' S *}
SP
Pmt m P
P S S S * 1 ( m /2) 1 mmt/2
'
(3.2)
s p s p
Cm Cm
The above condition is valid when the probability of mutation pmt <0.5, if the probability
of mutation exceeds 0.5 then the probability that the population does not contain solution
S* is given as follows.
(1 Pmt )
P S * mut ( P) 1
Cmm /2
P S * mut ( P)
Now evaluating binary strings over the objective f P k max ( f ( S )) if elite method
SSp
is utilized for selection then,
f f f ( p kf ) f ( s ) f p
0
Re sults
p p 'f p kf Pf0 Pr obability
Here Pfk Probability that f ( P k ) f after k th iteration and Pfk is the probability that
f ( P k ) f * after the k th iteration now that the required expectation probability does not
decrease,
E P k E P k 1 E P k 2 ....... E P0
unchanged i.e, E[ P K ] f *
Consider situation when A =A that no solution is found after the first iteration,
consequently suppose that a hypothesis H stating at least one of the solutions results from
_ _
crossover, Then possible events P A P A | H P H P A | H P H
3.4.1.4 Criteria for Crossover
Estimating P{H} from the above which only differ from the pmt by the fact that
population contains one solution before mutation, where as it does not contain any in
second case. Applying from the above,
n
n l
Pmt m
PA | H P S .......S
*
1
*
P
mut cross( P 0 ) 1 mmt/2 1
m /2
C m
l
Cm
n
n l
Pmt m
S .......S
P
_
* *
P A | H P mut cross( P ) 1 mmt/2
0
1 m /2
1
C m
l
Cm
crossover it can be concluded that a pair is good if both the individuals in the pair
contains fragments of the same solution as sub-string.
C Event that all the pair is chosen for crossover are good.
P H P H | B P B P H | B P H | C
This can be concluded from the fact that if all the pairs are good then the probability is
maximum compared to other events,
So the probability that a good pair (s1,s2) yields a solution .can be written as
q
P{s1 s2 s*} pc pc q m 1,
m 1
P{s1 s2 s*} 1 pc
In contrary P{H | C}, the probability that a solution does not arise after a crossover,
provided that all pairs are good:
P{H | C}
( S1 , S2 )Cross ( P )
P{s1 s2 s1* ,...sl } (1 pc ) (1 pc ) n
P{H | C} 1 (1 pc ) n
Assuming that atleast n pairs take part in the crossover. Then probability of good pairs
after crossover
Now if Event A happens then the possibility of event A can be written as follows
m n
pmt pmt
(1 [ m /2]
) n
(1 (1 pc ) n
) (1 [ m /2]
)n *1
CM Cm
m
pmt
S (1 [ m /2]
)n (2 (1 pc ) n ) (3.3)
CM
Therefore probability in kth iteration can be written through mathematical induction as,
That no solution arises after the kth iteration now from above frame works the
expectation of solution after the kth iteration
E[ P K ] fp kf f * p kf f * (1 S k ) f *
With this following conclusions can be drawn from the expectation regarding the
parameters which influence mean convergence rate
m
pmt
(1 [ m /2]
)n 0 as (n infinity)
CM
and (2 (1 pc ) ) 2 as (n infinity)
n
In order extract the extreme limits of convergence we consider the extreme of the
function S(n).
m
pmt
Let a 1 [ m /2] b 1 pc
CM
S (n) a n (2 bn )
S ' (n) a n (2 bn ) ln a bn ln b
2log a
The optimum n is given by n= log b
ln ab
With increase in population size the string length increases, reaches optimum peek
and then decreases. The least string length possible for encoding is 1 for zero size
population.
From above result it can be concluded that n exists and is real and S attains its
maximum i.e, as n increase, S first increases and then decrease.
For accelerating convergence pc should be minimum while pmt should be maximum
but the drawback of having pc minimum is that the best solution are not inherited to
next generation and high pmt destroys best solution space.
With elitism the expectation of solution getting transferred to next generation
increases moreover if crossover probability is increased the density of best solution
increases including good solution in every iteration.
If the mutation probability is decreased then the string length participating decrease
which decreases the passivity of killing best solutions, but if mutation is accurately
tuned then there is quite a possibility that worst strings could give good solution.
If the crossover size is increased the convergence decelerates but the chances of
obtaining good solution increases. In contrast for accelerating convergence if the
mutation rate is increased then good solutions are lost leading to no solution so a
good balance between convergence time and crossover-mutation rate is essential.
It is always suggested that in order to obtain good solution if convergence rate is
allowed to float freely whenever possible and when convergence rate is strict criteria
then it is suggested that the string length is kept minimum so that humming effect and
relative degrees of change in string character is merely small.
With the key points in mind evolutionary algorithm is chosen with properties which
could overcome these drawbacks A fast elite NSGA II and SPEA2 algorithms are utilized
to optimize our machining system.
Initialize Variables uses the bounds of V decision variables and randomly generates N
number of population over the bound range and each objective M is evaluated over
this population pop for fitness through Evaluate Objectives.
1. For i :[1-N]
2. For j :[1-V]
3. V[i,j]= Rmin+ random(0,1)*Range
3.5.2 Non_dominated_sort
7. Initialize the front counter to one. i = 1 following is carried out while the ith
front is nonempty i.e. F[ ]= .
25. Q=[q,j]
26. q.rank=pop(F[i])+1
27. exchange q with pop[j]
28. F[pop[i]+1]=Q
Once the non-dominated sort is complete the crowding distance is assigned. Since the
individuals are selected based on rank and crowding distance all the individuals in the
population are assigned a crowding distance value. Crowding distance is assigned front
wise and comparing the crowding distance between two individuals in different front is
meaningless. The crowing distance is calculated as below
19. Assign infinite distance to boundary values for each individual in Fi i.e. I (d k ) =
and I (dn) =
20. for k = 2 to (n 1)
I (k 1).m I ( K 1).m
I (d k ) I (d k ) 3.4
f mmax f mmin
I(k).m is the value of the mth objective function of the kth individual in I
Distance (M, V, N)
3.5.3 Selection
Once the individuals are sorted based on non-domination and with crowding distance
assigned, the selection is carried out using a crowded- comparison-operator
Non-domination rank prank i.e. individuals in front Fi will have their rank
The individuals are selected by using a binary tournament selection with crowed-
comparison-operator
Real-coded GAs use Simulated Binary Crossover (SBX) operator for crossover and
polynomial mutation Simulated Binary Crossover. Simulated binary crossover simulates
the binary crossover simulated.
(u) (2u)1/( 1)
c
3.10
(u) (1/ 2(1 u)) 1/(c 1)
3.11
Crossover (parent pop, M, V, Rang, PC)
1. For i:[1-N]
2. If Random(0,1)<PC
3. //Child initiation child1 and child 2
4. Select parents
5. P1: round [N*random(0,1)]
6. P2: round[N*random(0,1)]
7. Parent 1=parent pop[P1,:]
8. Parent 2=parent pop[P2,:]
9. //Simulated Binary Crossover
12. Else
1
2(1 U i ) Pc 1
ck pk ( p u k pkl ) k
k (2rk )1/ m 1
1 if 0.5 3.12
1/m 1
k 1 [2(1 rk )]
(i ) 2 * m(i ) pm 1
else
1
(i )
2(1 m(i )) pc 1
The best parents from former generations and off springs in current generations are
combined to preserve parents so that the parents are conserved in the consecutive
generations
Replace pop (Intermediate pop,M,V,N)
1. For i : [1-N]
2. Sort pop
3. Max_rank=Intermediate pop[find(maxRank)]
4. For i: 1 to max_rank
5. J=max(find[sorted pop(max>rank)==i)
6. If (j>N)
7. //sorted with rank
8. //find the number of individuals with current rank
9. k=j-N
10. p=sorted_pop(k:N)
11. //Sort according to distance
12. For j : [1-N]
13. F[N+K:]=p(N:j)
14. Elseif j<N
15. F[N:j]=sorted_pop[j:N]
16. For i :[1-Kmax]
17. Pool=round(Np/2),tour=2
18. Parent pop=Tournament selection (pop, Tour)
19. [Child 1,Child 2]=Crossover (Parent pop, M, V, Range, PC)
20. [Child 3]=Mutate(Parent pop, M, V, Range, Pm)
21. Offspring pop=[Child 1, Child 2, Child 3]
22. Intermediate pop =[pop, offspring]
23. Replace pop(Intermediate pop, M, V, N)
Table.3.3 Results of NSGA II family of best solution for AISI 4340 35 HRC Steel
Vc f d Ra Ft Fa Fr Tf R D
,Ft ,Fa ,Fr ,Tf ).Out of the 1000 chromosome solutions few best chromosomes
Fig. 3.4 Rank and Pareto for Ra(35HRC) Fig. 3.5 Pareto-front for Tf (35HRC)
Table.3.4 Results of NSGA II family of best solution for AISI 4340 45 HRC Steel
Vc f d Ra Ft Fa Fr Tf R D
175 0.2500 2 5.4703 1.1135e+03 576.9927 494.4395 8.9403 1 65535
135.9606 0.1500 1 3.9895 519.5461 331.3059 282.9869 28.2786 1 65535
130.8263 0.1500 1 4.0725 524.2045 331.6248 282.5873 30.0149 1 65535
Table 3.4.consists optimized results for 45 HRC process parameters in first three
columns (vc ,f, d) and latter columns contains optimized results for objective functions
(Ra ,Ft , Fa , Fr , Tf ).Out of the 1000 chromosome solutions few best chromosomes are
listed in the above table.
Fig. 3.7 Rank and Pareto for Ra (45HRC) Fig. 3.8 Pareto-front for Tf (45HRC)
Initialize population structure with fields in position cost fitness variables, dominance
field and cumulative fitness. Then objectives are evaluated over random pop. Position
with initial fitness, pop.cost. A fitness subset archive for best individuals is initiated,
first, all non-dominated population members are copied to the archive; any dominated
individuals or duplicates are removed from the archive during this update operation. If the
size of the updated archive exceeds a predefined limit, further archive members are
deleted by a clustering technique which preserves the characteristics of the non-
dominated front. Afterwards, fitness values are assigned to both archive and population
members:
Each individual i in the archive is assigned a strength value S(i) [0, 1), which at the same
time represents its fitness value F (i).and S(i) is the number of population members j
that are dominated by or equal to i with respect to the objective values, divided by the
population size plus one.
S i {j j P.t Pt i j} |
R(i)
j Pt Pt , j i
S ( j)
On the basis of the S values, the raw fitness R(i) of an individual i is calculated:
That is the raw fitness is determined by the strengths of its dominators in both archive
and population. fitness is to be minimized here, i.e., R(i) =0 corresponds to a non-
dominated individual, while a high R(i) value means that i is dominated by many
individuals.
The density estimation utilizes kth nearest neighbor method, where the density at any
point is a (decreasing) function of the distance to the kth nearest data point. Here the
inverse of the distance to the kth nearest neighbor is used for as density estimate. For each
individual i the distances (in objective space) to all individuals j in archive and population
are calculated and stored in a list. After sorting the list in increasing order, the kth element
gives the distance sought, denoted as k. The kth nearest parameter is defined as square
root of the sample size, NN and density distribution as
1
D(i) (3.13)
k 2
In the denominator, two is added to ensure that its value is greater than zero and that
D(i) < 1. Finally, adding D(i) to the raw fitness value R(i) of an individual i yields its
fitness F (i).
2. Pop.position :random(range,N)
3. Pop.cost=evalautefunction(pop.position)
4. //Initialize archive
5. Archive={ }
6. I=random(N,2)
7. I1=I(1)
8. I2=I(2)
9. If F(i1)<F(i2)
10. P=pop[i1]
11. Else
12. P=pop[i2]
21. Rmax=max(range)
22. dr=Rmax-Rmin
23. : *dr
24. y=p1+*random(Nm)
25. y=min(max(y,range))
26. Main learning
27. Do until( max iteration IT)
28. P=[pop,archive]
29. //check for domination
30. [dom,p.S]=dominates (p[i])
31. S=[p.S]
32. P[I].R=sum(S[dom])
33. Q=[p.cost]
34. : Euclidiean distance[q]
35. : sort[]
36. p[i]. =
37. p[i].[k]=p[i].[k]
1
p[i].D
38. p[i]. [k 2]
39. //Fitness Calculation
40. P[i].F=p[i].R+p[I].D
41. Fit =sum(find(p.R==0))
42. P.F=[fit]
1. Archive =p[size[p.R]]
2. While[min()==max((k)&&k<size())]
3. Pareto front=archive[archive.R==0]
4. [p1,p2]=binary tournament selection(archive,[archive.F],N)
5. Popc.cost=evalautefitness[child1,child2]
6. [child 1 child2]=crossover(p1,p2,crossoverparameters)
7. [p3]=binarytournamentselection(archive,[archive.F],N)
Table 3.6 Results of SPEA II family of best solution for AISI 4340 35 HRC Steel
Position Cost S R sigmaK D F
[259.15 0.17 1.07] [1.88 514.11 136.22 225.95 35.51] [9.00] [0.00] 0.375 [0.32] [0.32]
[265.00 0.16 1.84] [1.60 674.93 238.28 350.68 27.63] [16.00] [0.00] 0.3563 [0.32] [0.32]
[265.00 0.24 1.81] [2.57 813.37 295.99 398.32 23.38] [17.00] [0.00] 0.2863 [0.34] [0.34]
[265.00 0.20 1.96] [2.05 757.72 292.73 391.58 24.60] [9.00] [0.00] 0.305 [0.33] [0.33]
[264.60 0.15 1.94] [1.51 683.95 250.23 366.11 27.68] [6.00] [0.00] 0.4035 [0.32] [0.32]
[264.83 0.17 1.74] [1.64 661.12 224.65 335.27 28.00] [26.00] [0.00] 0.3147 [0.33] [0.33]
[264.98 0.21 1.77] [2.11 737.27 265.23 367.09 25.17] [26.00] [0.00] 0.2558 [0.34] [0.34]
[265.00 0.19 1.79] [1.89 706.15 252.72 356.92 26.15] [20.00] [0.00] 0.3483 [0.33] [0.33]
[264.23 0.21 1.92] [2.25 776.56 295.04 392.75 24.13] [11.00] [0.00] 0.2511 [0.34] [0.34]
[264.85 0.22 1.47] [2.09 689.39 226.10 330.51 26.92] [34.00] [0.00] 0.3906 [0.32] [0.32]
[265.00 0.24 1.69] [2.57 802.55 280.54 388.16 23.84] [16.00] [0.00] 0.2971 [0.34] [0.34]
[264.99 0.18 1.88] [1.81 711.36 262.25 367.37 26.09] [17.00] [0.00] 0.3296 [0.33] [0.33]
[264.72 0.25 1.93] [2.91 874.97 328.11 427.52 22.13] [6.00] [0.00] 0.2971 [0.34] [0.34]
[256.63 0.16 1.09] [1.90 505.03 129.90 221.26 36.21] [13.00] [0.00] 0.3225 [0.33] [0.33]
[264.97 0.19 1.45] [1.83 636.05 200.63 303.73 28.72] [39.00] [0.00] 0.3818 [0.32] [0.32]
[263.42 0.20 1.31] [1.91 616.67 188.86 289.03 29.75] [30.00] [0.00] 0.368 [0.32] [0.32]
[264.62 0.21 1.91] [2.16 763.63 289.27 387.95 24.44] [12.00] [0.00] 0.2511 [0.34] [0.34]
[264.57 0.23 1.41] [2.16 695.37 225.90 331.84 26.94] [30.00] [0.00] 0.3986 [0.32] [0.32]
[265.00 0.21 1.97] [2.28 789.31 306.06 402.04 23.74] [8.00] [0.00] 0.2477 [0.34] [0.34]
Table 3.6 has results of optimized solution obtained from SPEA2 algorithm for 35HRC
position matrix represents the process parameters (vc,f,d), cost matrix represents
optimized solution for objectives (Ra ,Ft ,Fa ,Fr ,Tf ). and the remaining cells correspond to
SPAE2 parameters. Out of the 300 solutions in archive only few are listed in the above
table.
Table 3.7 Results of SPEA II family of best solution for AISI 4340 45 HRC Steel
Position Cost S R sigmaK D F
[175.00 0.22 1.90] [4.92 997.97 515.49 452.74 9.82] [14.00] [0.00] 0.238 [0.35] [0.35]
[174.72 0.23 1.54] [4.67 849.67 466.88 419.61 11.30] [10.00] [0.00] 0.271 [0.34] [0.34]
[174.92 0.18 1.34] [3.93 674.57 389.76 355.44 14.19] [29.00] [0.00] 0.427 [0.31] [0.31]
[174.83 0.23 1.70] [4.80 920.84 489.93 435.20 10.54] [14.00] [0.00] 0.247 [0.35] [0.35]
[174.85 0.18 1.55] [4.14 773.13 418.81 380.47 12.54] [21.00] [0.00] 0.299 [0.33] [0.33]
[175.00 0.16 1.33] [3.83 656.51 380.43 346.35 14.81] [26.00] [0.00] 0.482 [0.30] [0.30]
[174.99 0.19 1.70] [4.37 852.70 448.58 404.23 11.40] [21.00] [0.00] 0.253 [0.34] [0.34]
[175.00 0.19 1.44] [4.14 737.76 413.56 376.50 12.91] [24.00] [0.00] 0.373 [0.32] [0.32]
[174.99 0.22 1.47] [4.45 793.05 442.97 401.34 11.96] [11.00] [0.00] 0.278 [0.34] [0.34]
[175.00 0.21 1.84] [4.64 934.49 483.32 430.12 10.45] [20.00] [0.00] 0.238 [0.35] [0.35]
[175.00 0.22 1.51] [4.54 819.43 453.20 409.14 11.62] [13.00] [0.00] 0.26 [0.34] [0.34]
[173.87 0.19 1.34] [4.06 693.36 401.38 366.13 13.76] [26.00] [0.00] 0.435 [0.31] [0.31]
[175.00 0.21 1.40] [4.23 738.85 419.98 382.61 12.77] [23.00] [0.00] 0.348 [0.33] [0.33]
[175.00 0.24 1.29] [4.51 735.25 436.39 400.33 12.70] [7.00] [0.00] 0.346 [0.33] [0.33]
[168.21 0.16 1.02] [3.67 519.42 348.53 310.79 19.45] [15.00] [0.00] 1.069 [0.33] [0.33]
[175.00 0.24 1.19] [4.48 688.80 426.53 395.02 13.40] [8.00] [0.00] 0.425 [0.31] [0.31]
[174.95 0.23 1.26] [4.33 698.11 419.46 385.62 13.28] [16.00] [0.00] 0.347 [0.33] [0.33]
[175.00 0.22 1.70] [4.68 901.42 478.44 426.62 10.73] [18.00] [0.00] 0.266 [0.34] [0.34]
[175.00 0.24 1.82] [5.07 1000.50 524.52 459.35 9.80] [11.00] [0.00] 0.353 [0.32] [0.32]
Table 3.7 has results of optimized solution obtained from SPEA2 algorithm for 45HRC
position matrix represents the process parameters, (vc, f, d) cost matrix represents
optimized solution for objectives (Ra ,Ft ,Fa , Fr , Tf ) and the remaining cells correspond to
SPAE2 parameters Out of the 300 solutions in archive only few are listed in the above
table
Fig. 3.14 Rank and Pareto for Ra (45HRC) Fig. 3.15 Pareto-front for Tf (45HRC)
Swarm systems are based on behavior of school of birds, insects, fireflies where a flocks
of birds twisting ,V-shaped structure of migrating geese, winter birds hunting for food,
the synchronized flashing of fireflies are tried to imitating. The well-choreographed
collective behavior without any leader is adopted to search for optimal solutions. For
instance Ants living in colony, their behavior is driven by the goal of colony survival
instead of individual survival, while searching for food ants initially explores surrounding
nests. In random manner A similar behavior is observed with flocks of birds where a
leader keeps guiding the flock to updated food location.
Where
X t X tn,d 1 n Nt & 1 d D
Vt Vt n,d 1 n Nt & 1 d D
Lt Lnt ,d 1 n Nt & 1 d D
Gt Gtn,d 1 n Nt & 1 d D
n 1
Furthermore Gtn ,d = Gt if (n<N) and satisfying initial and final conditions Gt Gt 1
N ,d l ,d
with given distribution for initial position and velocity (x0,v0) Initial grid index G0I is
determined by minimum argument of function.
vtn,1d vtn,1d C1.r1n,d Lnt ,d X tn,d C2 .S1n,d Gtn,d X tn,d (3.15)
Where C1 & C2 control the influence of personal best of particle and the common
Stn.d & rtn,d are randomness which are drawn uniformly at random [0,1]
If a particle velocity component exceed a certain interval [-Vmax, Vmax]. It is set back to
interval found. The movement equation is altered by clamping inertia.
vtn,d 1.vtn,d C1.rt n,d Lnt ,d X tn,d C2 .Stn,d Gtn,d X tn,d
and position is altered as
0.72 C1 C2 1.49
v n,d .vn,d C1rand (0,1) Ln,d X n,d C2 rand (0,1) G n,d X n,d
If f X n L Ln then Ln X n
f X n L G then G X n
To better predict social learning process the global best particle (Gd) is replaced by the
local guide particle (Ld) Topology is typically represented as graph whose nodes are
particles and edges connect neighboring particles. The edge connections between any two
grid points n1 & n2 as a particle for its own local guide, the edge connections are
determined by minimum argument of function
n ', n
Pt n Euclidean(arg min f ( x) where x [ Lt | n' N (n) |]
n ', n
Lt is local guide attraction of particle n at the time step t when particle n
makes its move.
n ' n ',n
n ', n
Lt
L n ' L If n n
L t 1
For a Swarm, definition of topology and its potential drives solution space, leader of
swarm both locally and global depends on the topology, though topology remains static
the co-ordinates of particle attractors monitors the distribution of swarm. In general the
global and local attractor is determined by fitness augments i.e.
dimension d is determines the swarm fitness level at interval time step t. It is determined
by global best and personal best of swam at all interval.
n 1 n N
nt ,d ( | vtn ,d | | Gtn,1d X tn ,d |) ( | v
' '
n' , d '
It is important to control swarm topology for determining the desired solution space
which depends on the movement constants applied on swarm so for a swarm to converge
the inertia damping co efficient which keeps the swarm in bound and the acceleration
coefficient keep drives the co-ordinates of local and global attractors.
At the topological development both global and local attractors are bound to same value
t
X t1,1 X 01,1 v1,1
s
t 0
t 1
X 01,1 v0 s
t 0
t 1
Lt s X 1,1
t Lt s [ X 1,1
0 v s
]
0 t 0
1
X 01,1 v0 [ ]
1
So for swarm to be in bounds the inertia damping co-efficient should be between [0<
<1]. Now for acceleration co-efficient the movement equation is analyzed
vtn,1d X tn,1d X tn ,d
vtn ,d X tn ,d X tn,1d
vtn,1d vtn ,d [ X tn,1d X tn,1d ]
Now to calculate the error in position between the expected and attained the expentancy
operator is applied on position vector.
Assuming the expectance to be , now for a particle in swarm it is expecte that the
calculated and expected positions are equal so for optimum goal attainment the
expectancy is equal to zero which converts the position expectancy equation to
2 [1 0.5*(C1 C2 )] =0
0 (C1 C2 ) 4(1 )
Initialize population structure with fields Particle. Evaluate fitness over particle
position and determine the dominance levels of each particle by dominates function
1. For i :[1-N]
2. Pop[i].position=random[Range,N]
3. Pop[i].velocity : zeros[N]
4. Pop[i].cost :evaluate (pop[i].position)
5. //update personal best
6. Exchange pop[i].best position with pop[i].position
7. Exchange popi[i].best cost with pop[i].cost
Initiate repository element which is subset of all particles with best position and cost
then select leader for swarm at every iteration through select leader function and
i=update the particle structure for current leader swarm.
Pop : domination(pop,N)
1. For i:[1-N]
2. For j:[i+1-N]
3. If dominates (pop[i], pop[j])
4. True(pop[j] Is dominated)
5. Else if dominates(pop[j],pop[i])
6. True(pop[i] Is dominated)
7. b=dominates(pop[i],pop[j])
8. b= all(x<y)&&any(x<y)
9. //Initiate repository element
10. Rep=pop[~dominated pop)
Apply mutation operator on the updated particle structure then calculate the
dominance level for current structure. Create neighborhood for swarm by initiating
grid topology for swarm. Update the swarm in the repository element with current
dominance level.
Now the topology is built for swarm which was initialized through particle structure
topology is static and remains unchanged at every generation. Naumann /Grid based
topology is built which utilizes Euclidean co-ordinates based on position of each swarm
Grid=create grid(rep,ngrid,)
1. P=[pop.cost]
2. Rmin=min(p,[],2)
3. Rmax=max(p,[],2)
4. dr=Rmax-Rmin
Rmin Rmin * dr
5. Rmax Rmin * dr
6. //Initiate grid index
7. Grid.L : [] Grid.H :[]
8. Object : size[p,1]
9. For j: [1-object]
10. p.object[equal spacing in R]
11. Grid[j].L =[-, p.object]
12. Grid[j].H=[p.object, ]
13. For i :[1-size(Rep)]
14. Rep[i]=find Grid index (Rep[i],Grid)
15. Obj=size(particle.cost)
16. Grid size=size(Grid.L)
17. Particle.Grid sub Index :zeros(p.object)
18. For i : [1-obj]
19. Particle.Grid sub Index[i]=find(particle.cost[i]<grid]
20. Particle.Grid Index=N*Grid size*particle grid index+ particle grid sub index
21. //Initiate repository element
22. Rep=pop[~dominated pop)
After building topology each particle in repository is recognized with its position and
velocity coordinates the identity of each particle is recognized through this co ordinate
Swarm is lead most fittest particle/particles At each generation the swarm changes its
leader according to the swarm velocities and position evaluated through swarm
movement equation.
//select leader
Select leader(rep, )
Excess particles in the repository are either deleted or replaced by better fit particles in
each generation if repository exceed its size then the interia damping factor reduces the
velocity of swarm resulting in poor convergence hence at aevery generation repository is
checked for its size.
[3] Deletion={ }
Now that the swarm and its topology is built the swarm is allowed to move over the
topology with governing movement equation followed by a slight mutation in fit particles
which would accelerate motion of swarm by randomly changing the velocity and position
coordinates.
Main Learning
2. Newpop=mutatute(pop,pm,Range)
3. newpop.cost=evaluate(newpop.position)
4. Determine domination(Rep)
5. If dominates(New.pop.position, pop.position)
6. True(Is dominated pop.position)
7. Else if dominates(pop.position, New.pop.position)
8. True(Is dominated New.pop.position)
9. Grid=Create Grid(Rep, grid size,)
10. Check if resize>maxrep
Table 3.9 MOPSO family of optimal solutions for 35 HRC AISI 4340 steel
Grid
Position Velocity Cost Best Position Best Cost Index GridSubIndex
[265.00 0.16 [1.55 702.05 266.48 379.44 [263.49 0.22 [2.01 595.25 181.77 277.57
2.00] [1.62 -0.06 1.07] 26.93] 1.01] 31.89] 10668 [2,6,6,7,3]
[212.34 0.15 [-43.26 -0.04 - [2.99 426.70 123.66 206.18 [255.60 0.19 [2.03 659.89 217.50 320.77
1.00] 0.97] 43.69] 1.59] 28.39] 33631 [6,2,2,2,7]
[193.76 0.15 [137.29 -0.02 - [3.43 417.33 134.23 215.15 [193.76 0.15 [3.43 417.33 134.23 215.15
1.00] 0.35] 46.12] 1.00] 46.12] 40282 [7,2,3,3,7]
[174.19 0.15 [-85.70 -0.01 - [3.91 413.06 148.34 227.70 [174.19 0.15 [3.91 413.06 148.34 227.70
1.00] 0.24] 49.11] 1.00] 49.11] 46844 [8,2,3,3,8]
[265.00 0.15 [1.45 652.07 221.19 340.52 [265.00 0.15 [1.45 652.07 221.19 340.52
1.82] [7.93 -0.01 0.20] 28.93] 1.82] 28.93] 9850 [2,5,5,6,4]
[174.72 0.15 [-32.71 -0.00 - [3.89 413.10 147.91 227.31 [207.43 0.15 [3.10 427.11 127.30 209.77
1.00] 0.01] 49.02] 1.01] 44.08] 46844 [8,2,3,3,8]
[215.30 0.15 [-30.50 -0.03 - [2.92 428.66 122.23 205.01 [215.30 0.15 [2.92 428.66 122.23 205.01
1.00] 1.23] 43.34] 1.00] 43.34] 27070 [5,2,2,2,7]
[211.21 0.15 [3.01 425.98 124.22 206.64 [211.21 0.15 [3.01 425.98 124.22 206.64
1.00] [-1.61 -0.04 -0.73] 43.83] 1.00] 43.83] 33631 [6,2,2,2,7]
[194.42 0.15 [-14.35 -0.05 - [3.42 417.57 133.80 214.78 [194.42 0.15 [3.42 417.57 133.80 214.78
1.00] 0.63] 46.02] 1.00] 46.02] 40282 [7,2,3,3,7]
[265.00 0.20 [1.95 694.04 238.22 342.10 [265.00 0.20 [1.95 694.04 238.22 342.10
1.66] [37.22 -0.00 0.20] 26.52] 1.66] 26.52] 17139 [3,6,5,6,3]
[171.25 0.15 [-32.00 -0.03 - [3.98 412.91 150.71 229.85 [203.25 0.18 [3.13 673.17 231.63 322.24
1.00] 0.62] 49.60] 1.60] 33.15] 46844 [8,2,3,3,8]
[265.00 0.15 [1.66 509.07 120.80 217.67 [265.00 0.15 [1.66 509.07 120.80 217.67
1.13] [32.08 -0.01 0.13] 36.15] 1.13] 36.15] 14684 [3,3,2,3,5]
[265.00 0.15 [26.00 -0.05 - [1.49 603.83 180.32 296.71 [265.00 0.15 [1.49 603.83 180.32 296.71
1.59] 0.03] 30.84] 1.59] 30.84] 9031 [2,4,4,5,4]
[249.95 0.16 [2.10 480.93 125.34 211.22 [249.95 0.16 [2.10 480.93 125.34 211.22
1.00] [8.34 -0.09 -0.99] 37.96] 1.00] 37.96] 21236 [4,3,2,2,5]
[265.00 0.16 [1.50 696.28 262.59 377.89 [265.00 0.16 [1.50 696.28 262.59 377.89
2.00] [32.06 0.00 1.15] 27.25] 2.00] 27.25] 10659 [2,6,6,6,3]
[265.00 0.16 [1.57 608.71 184.51 296.17 [265.00 0.24 [2.38 749.62 248.99 359.27
1.55] [2.84 -0.08 0.06] 30.28] 1.49] 25.43] 9031 [2,4,4,5,4]
[265.00 0.15 [1.48 610.04 185.18 302.22 [265.00 0.15 [1.49 603.83 180.32 296.71
1.62] [1.51 -0.02 0.03] 30.57] 1.59] 30.84] 9031 [2,4,4,5,4]
[265.00 0.20 [1.96 698.30 241.07 344.82 [265.00 0.20 [1.96 698.30 241.07 344.82
1.68] [10.04 0.01 0.04] 26.38] 1.68] 26.38] 17139 [3,6,5,6,3]
[265.00 0.17 [125.17 -0.00 - [1.74 530.80 141.76 233.24 [265.00 0.17 [1.74 530.80 141.76 233.24
1.10] 0.30] 34.33] 1.10] 34.33] 15494 [3,4,3,3,5]
[265.00 0.16 [1.61 708.49 270.60 381.20 [265.00 0.15 [1.49 603.83 180.32 296.71
2.00] [0.68 0.01 0.47] 26.59] 1.59] 30.84] 10668 [2,6,6,7,3]
[265.00 0.24 [28.76 -0.01 - [2.70 846.43 323.67 419.78 [265.00 0.24 [2.70 846.43 323.67 419.78
1.96] 0.04] 22.55] 1.96] 22.55] 31169 [5,7,7,8,2]
[265.00 0.15 [1.48 694.34 261.24 377.37 [265.00 0.15 [1.48 694.34 261.24 377.37
2.00] [3.61 0.00 0.67] 27.37] 2.00] 27.37] 10659 [2,6,6,6,3]
[250.88 0.17 [2.09 486.86 128.60 214.65 [250.88 0.17 [2.09 486.86 128.60 214.65
1.00] [-7.54 -0.02 -0.11] 37.52] 1.00] 37.52] 21326 [4,3,3,3,5]
Table 3.9 has results obtained from MOPSO for 35 HRC the position, Best position matrix
represent process parameters (vc,f,d),. and Cost, Best Cost represents objective fitness (Ra ,Ft ,Fa
,Fr ,Tf ) and the latter columns are corresponding to grid index and grid sub index of PSO
topology Out of the 500 repository elements few are listed in above table.
Fig 3.18 Pareto spread surface roughness and Fig 3.19 3D surface plot of optimal Ra with
cutting force 35HRC best position 35HRC
Fig 3.20 3D surface plot of Tf with best Fig 3.21. Depth of cut influence on cutting
position 35HRC forces 35 HRC
Table 3.10 MOPSO family of optimal solutions for 35 HRC AISI 4340 steel
Table 1 Repository elements Solution for Population
Position Velocity Cost Best Position Best Cost Grid Index GridSubIndex
[131.18 0.15 [4.07 523.87 331.58 282.59 [169.76 0.15 [3.91 730.05 393.60 357.11
1.00] [-38.59 -0.02 -0.73] 29.89] 1.55] 14.41] 20510 [4,2,2,2,8]
[145.95 0.15 [3.85 512.27 332.51 286.06 [145.95 0.15 [3.85 512.27 332.51 286.06
1.00] [-29.05 -0.08 -1.19] 25.34] 1.00] 25.34] 13948 [3,2,2,2,7]
[175.00 0.24 [5.24 1075.52 554.08 479.04 [172.92 0.18 [4.53 953.91 482.02 431.27
2.00] [2.94 0.05 0.95] 9.20] 2.00] 10.62] 45011 [7,8,7,7,2]
[175.00 0.15 [3.88 720.39 390.67 356.83 [175.00 0.15 [3.88 720.39 390.67 356.83
1.52] [18.14 -0.00 0.06] 13.99] 1.52] 13.99] 15501 [3,4,3,4,3]
[175.00 0.19 [4.67 977.77 494.94 440.17 [175.00 0.19 [4.67 977.77 494.94 440.17
2.00] [2.26 0.01 0.10] 10.11] 2.00] 10.11] 31070 [5,7,6,6,2]
[157.18 0.15 [3.72 506.91 336.75 293.12 [157.18 0.15 [3.72 506.91 336.75 293.12
1.00] [-17.82 -0.04 -1.28] 22.59] 1.00] 22.59] 13947 [3,2,2,2,6]
[175.00 0.19 [4.62 969.60 489.97 436.99 [136.93 0.15 [4.01 617.69 362.52 315.19
2.00] [56.89 -0.05 0.12] 10.21] 1.26] 23.55] 31070 [5,7,6,6,2]
[175.00 0.24 [4.99 944.81 506.19 447.74 [175.00 0.24 [4.99 944.81 506.19 447.74
1.68] [4.45 -0.01 0.22] 10.29] 1.68] 10.29] 36911 [6,6,6,7,2]
[132.82 0.15 [4.04 522.32 331.42 282.65 [132.82 0.15 [4.04 522.32 331.42 282.65
1.00] [-53.49 -0.04 -0.47] 29.32] 1.00] 29.32] 20510 [4,2,2,2,8]
[149.38 0.15 [3.80 510.32 333.49 287.81 [149.38 0.15 [3.80 510.32 333.49 287.81
1.00] [-7.05 -0.03 -1.52] 24.44] 1.00] 24.44] 13947 [3,2,2,2,6]
[154.57 0.15 [3.75 507.89 335.50 291.14 [154.57 0.15 [3.75 507.89 335.50 291.14
1.00] [-20.43 -0.12 -0.45] 23.19] 1.00] 23.19] 13947 [3,2,2,2,6]
[162.55 0.15 [3.67 505.40 339.86 297.86 [166.04 0.15 [3.64 504.79 342.25 301.40
1.00] [-3.48 -0.02 -0.06] 21.45] 1.00] 20.75] 7395 [2,2,2,3,6]
[173.21 0.15 [0.62 -0.06 -0.09] [3.64 549.52 354.83 317.92 [172.59 0.20 [4.02 635.55 389.86 357.26 8204 [2,3,3,3,5]
Table 3.10 has results obtained for MOPSO for 45 HRC the position, Best position matrix
represent process parameters (vc,f,d),. and Cost, Best Cost represents objective fitness (Ra ,Ft ,Fa
,Fr ,Tf ) and the latter columns are corresponding grid index and gris sub index of PSO topology
Out of the 500 repository elements few are listed in above table.
Fig 3.22 Pareto spread surface roughness Fig 3.23 3D surface plot of optimal Ra
and cutting force 45HRC with best position 45HRC
Fig.3.24 3D surface plot of Tf with best Fig.3.25 Depth of cut influence on cutting
position 45 HRC forces 45 HRC
Fig.3.26. Solution Spectrum for 35 hrc Fig.3.27. Solution Spectrum for 45 hrc
Fig.3.28 Solution Spectrum for 35 hrc Fig.3.29 Solution Spectrum for 45 hrc PSO
Fig 3.30. Solution Spectrum for 35 hrc Fig. 3.31 Solution Spectrum for 45 hrc
3.1 The solution spectrum distribution in NSGA II (as shown in Fig. (3.26-3.27)) has
attained uniform amplitude with periodic crusts and troughs in all the objectives
exhibiting quit a good saturation level in the demography in solution space for
both hardness levels.
3.2 Solution Spectrum in SPEA 2 While in SPEA 2 the demography of solution space
(as shown Fig. (3.30-3.31)) is quite different from the NSGA II as the solution
space is built on niche Pareto sharing
3.3 The spectrum is pretty disturbed with low levels of saturation and crusts and
troughs varying throughout the wavelength of the data. for both the hardness
levels.
4. Solution Spectrum in PSO
4.1 The trend in PSO shows moderate level of disturbance in spectrum of solution
space (as shown Fig. (3.28-3.29)) but the level of saturation is appreciable when
compared to the SPEA 2 and change in demography is not as amplifying as SPEA 2
while in comparison with the NSGA II it is inferior in terms of saturation levels.
From the nature of spectrum in solution space conclusion can be condensed to as follows.
NSGA II relatively better compared to PSO and SPAE 2, PSO is better compared to
SPEA 2.
1. The diversity of solution space is evaluated through the average Pareto spread in
each generation, diversity measures exploration potential in search space.
2. Higher diversity in generations gives greater chances of solutions getting retained
from varying locals of search space hence increasing the strength of solution.
3 The diversity of NSGA II is superior when compared to SPEA 2 the range of search
exploration between the generations in NSGA II varied from Euclidean spread of [0.01 to
1] (Fig.3.6 and Fig. 3.9) which shows that the exploration happened between two
extremum and in mid generations solution spread curled towards mean solution and
drifted away from the mean solution.
4 While in the SPEA 2 the Euclidean spread was in a short range of 0.27-0.37 (as shown
in (Fig.3.26 and Fig. 3.27). with all the equal dominating and non-dominated solutions in
the spread, the diversity oscillated in short range at each generation with most of the
generations between average spread [0.3-0.35].
5 The overall analysis between EA and PSO suggests that NSGA II performs well in
terms of diverse solution while SPEA 2 and PSO performs well when the solutions
spread is short In NSGA II more diverse solutions are preserved while in PSO and SPEA
2 neighborhood solutions are preserved.
Further potentials of both EA and SI are explored with synergism with prediction models
in chapter 5 where EA and PSO are utilized to enhance learning in prediction models.
CHAPTER 4
4.1 Introduction
Neural network consists of nodes connected by direct links each link has numeric
weight wij associated with it which determines the strength and configuration
between links. A threshold function activation function f(.) is applied to model
which alters the topology of links, these connection between nodes forms layer
pattern so called network architecture. Depending on the direction of propagation of
weights in the link neural model is classified into feed forward and feed backward
/recurrent network.
Feed forward networks are arranged in layers such that each unit receives input only
from the units in the preceding layer. The architect of the feed forward network has
following composition.
First layer is input layer receiver of data or input from the external stimuli,
the incoming data is then sent to the next layer where the number of layers
can be more than one.
Second layer consists of hidden layer in which the number of nodes depends
on the complexity and non-linearity of data to be handled with weights
defining connections between node and bias at each node A single hidden
layer constituents a network activated by threshold /activation function which
takes augment of weights and bias matrix from the net This augmented net is
propagated to the subsequent layers depending number of hidden layers.
Data processed in hidden layers are routed to the output layer. This layer
plays a role in determining the validity of data that are analyzed based on the
existing limits in the activation function.
Neural network runs training examples through the net one at a time,
adjusting the weights slightly after each example to reduce the error. Each
cycle through the examples called epoch.
Since the essential elements of neural network are discussed now the working matrix
by which each layer learns is discussed through simple network.
Let X be input vector and Y be output layer vector with weight matrix mapping
between the input and output layer. Then the neural network is characterized by the
learning model
n
Yi Wx Xi bi
i 1
Where the input layer and output layers are defined by the vectors
X {x1 , x2 , x3 ,......xl }
Y { y1 , y2 , y3 ,....... yl }
Now for training and mapping between the input and output layer learning law which
describes that prediction accuracy to increase the weights of nodes should be
correlated to attain minimum error in predictor consequently to store a prototype
( xi , yi ) .The weights are altered by weight matrix
w .yi x iT
Where is learning factor and is generally kept positive, elements of weight matrix w
starts from zero to a perfectly associative neuron weight. And inverse mapping is possible
at any instance of learning stage by recalling the weight matrix as
L
W Yi .Xi T
i 1
W . X K YK K 1, 2...l
Now that the weight matrix is defined and one to one mapping between the input and
output layer is established the next step is to minimize the error of mapping in weight
matrix by applying gradient descent approach.
According to gradient descent the mean square E(w) associated gradient of expected
error .The error gradient ( points in the direction in which E(W) will decrease at fastest
possible rate
w k 1 w k (E)
arbitrary constant similarly the lest mean square error for predicted output and weight is
calculated as
1
ek [w(k), y(k)] [| w(k).y(k) | w k .y k ]
2
To minimize this error the error gradient showed be headed in the direction where error is
least
ek w k y k 1
y k .Tb w k .y k y k
w k 2
w k 1
0
From extremum of error gradient we get the
Tb w k 1 if w(k).y(k) 0
1 if w(k).y(k) 0
w k 1 w k
2
y k y k Tb w k y k
Now validating weights for desired output d(k), the augments w(k) and y(k)determines
the rate at which the desired output is approached .
ek d k w k * y k
2
e2k d k w k .y k d k w k .y k
2 T
To minimize the least square error expectation operator E[ek2:]is applied over the squared
gradient of error
e2k
k 2e k y k
w
E k 2E ek y k
E(K ) 2E y k .yT k .w k d k .y k
By estimating the mean of the gradient the direction of least error propagation can be
known
K 2E y k .yT k .w k d k .y k
the expectancy of gradient is expected to be zero and with few manipulations weight
matrix which has least error propagation is obtained.as
K 0
P y k .yT k .
Q d k .y k
W* Q1.P
Now that the weights matrix for a single layer is established an extension of weight
matrix for consecutive layers are evaluated for the above multi-layer network.
y j k d k
1 m
E k
2
2 j1
N
ET E k
k 1
Where ET is the expectation of all the layers similarly evaluating the gradients of
expectation
E k
yj dj
dy j
s j yi wil j
The activation function or transfer function characterizes the input output relationship
y j f j (s j )
Most common choice of activation function is sigmoidal function which satisfies the
Now for the above function the error propagation for every input to layer can be written
by chain rule as
E k E k y j
s j y j s j
After few assumptions in the activation function and applying the above chain rule
dyi d 1
( ) y j (1 y j )
ds j ds j 1 exp( s j )
E (k )
( y j d j ) y j (1 y j )
s j
Where j are the threshold value generally referred as bias and yj in the consecutive layers
is determined by transfer function.so at each layer the targets changes with the weights
which are expected to approach to desired matrix. Thus the error gradient at each layer
with respect to the weights can be written by chain rule as
E k E k yl si
w m yl sl w ml
There are two approaches to apply gradient descent method to the training method of a
multi-layer feed forward neural network .The first is based on periodic updating and
second is based on continuous updating. In both the cases the weights are repeatedly
monitored either sequentially or randomly until the convergence criteria is satisfied
With m donating the number of weights in the network and weights are updated only
once every epoch after all the training patterns are evaluated the weights get updated by
the generalized fixed increment/decrement rule.
E k
w new w old n
w
Where is small constant referred to as learning rate. Wnew and wold are weights vector
at epoch k+1 and k respectively.
In order to build accurate prediction model sufficient input and output vectors is
necessary to be included in the network.
With the above fundamentals a prediction model for machining system [1] is built for
both the steels i.e., 35 HRC and 45HRC with process parameters as input vectors and
cutting speed, forces and tool life as output vectors. In the Fig a schematic description of
developed prediction model is explained in detail. This model is adopted for predicting
both the steels.
Fig.4.4 Feed forward neural network for AISI 4340 Hard turning
Table 4.2.4 (b) Calibrated weights and bias above Neural Network
Calibrated weights and bias for 35 HRC Steel
The hypothesis of network for both the steels were quite accurate with the man
square error approaching to order of 10 -3.Error gradients and mean error gradients in
learning stage is found to converge to minimum criteria. Error analysis for each
target output is evaluated at learning stages i.e., training testing and validating and
targets are fir to linear regression model with regression coefficient approach to 1.
There are two ways of implying fuzzy neural systems. In the first method the fuzzy
rules are modified with no change in the input and output membership functions. In
the second method fuzzy neural systems with learning algorithms such as back-
propagation or hybrid learning are applied to learn and adjust the membership
function parameters.
Different combinations of fuzzy neural systems are possible with varying input
output membership functions. The applied adaptive neuro fuzzy interference system
is explained through a two input-single output model utilizing Sugeno-type fuzzy
system also known as Takagi-Sugeno-Kang type fuzzy system where rule base is
replaced by neural network weights and output membership are defined by linear
function instead of fuzzy linguistic model.
Layer 1 Every node in this layer is defined by membership function where each node
gives membership values after evaluating inputs over the membership function. The
applied membership function can be linear or exponential with each input defined by
desired subsets in membership function In most cases Gaussian membership function
is applied
x ci 2
A,i exp[( ) ]
2ai
O1,i A,i ( x), i 1, 2..k
O1,i B ,i ( y ), i 1, 2..k
Layer 2 In this layer the node is fixed and takes the fuzzified value as input from the
layer 1. The output of this node is the result of fuzzy multiplication of membership
function which goes into the next node. Each node represents the firing strength of
each rule in the second layer. The T-norm operator with and operation is applied to
obtain the output. This layer is known as andecent layer
Layer 3 In this layer the normalized weight for each firing strength with respect to all
the cumulative fire strength is calculated. The resulting weights is called normalized
firing strength
wi
O3,i wi
wi
Layer 4 Every node in this layer is adaptive node known as consequent layer and
gives output with node function defined as
O4,i wi fi
O4,i = wi ( pi x qi y ri )
Layer 5 Single node fixed node that calculates overall output of from consequent
layer
O5,i wi fi
The adaptive layers i.e., first and fourth layer contain parameters which can be
modified at every iteration. The antecedent and consequent parameters can be
updated through learning method. There are two paths of learning forward and
backward path. In the forward path recursive least square method is used to alter
consequent parameters. While in the backward path the antecedent parameters are
changed through gradient descent method.at each iteration which is also called
epochs.
f w1 f1 w2 f 2
w1 ( p1 x q1 y r1 ) w2 ( p2 x q2 y r2 )
( w1 x) p1 ( w1 y )q1 w1r1 ( w2 x) p2 ( w2 y )q2 w2 r2
When N training data are given as input vector the n the consequent function changes
to
A y
The parameters in Gaussian parameters are trained for minimizing error For a given
adaptive network where the network consists of five layers and has total of N (L)
node in layer L then the square error in the L layer to n data is 1 n N the error at
each node can be written as
N (l )
En d k X kL,n
k 1
Where dk is the k-th component of the vector of the desired output while X kL,n is k-th
component of the vector of actual output generated by adaptive network with input
from the input vector n. The main aim of adaptive learning system is to reduce error
that occurs .i.e.,
En
L ,i 2(di ,n X iL,n )
iL,n
Applying chain rule for consecutive layers for error propagation we get
l 1
En N ( l 1)
En X m, p
l 1 l 1
X l ,i m 1 X m , n X m , n
With 0 l L 1 internal node error is a cumulative node error in the layer l+1. For a
specific node in adaptive layer the error rate corresponding to parameter is given as
En E x
n
xS x
Where S is the set of nodes containing the parameter the error specific to this
parameters is given as
En En
N
=
n 1
4.3.3 Fuzzy clustering Algorithms
Fuzzy clustering algorithms are utilized to discretize the membership function into
subsets in the input vectors so that each input vector is sub divided into topologies
defined by densities of points in respective region. Building a fuzzy set requires
following key points.
X (t 1) f ( x(t ), u (t ))
Y (t 1) f ( x(t ), u (t ))
f &g
f : X *U X
G : X *U Y
[m1T , m2T ,.....mmT ]T
c) Clustering in mfs
In any clustering technique the goal is to estimate that characterizes the best
cluster for input vector X. The parameter vector is sensitive to the shape of
clusters. To define the topology of clusters a set of m points mi in the l-dimensional
space is required which corresponds to a cluster
d) Definition of cluster
Let X be data set, for which m clustering is defined in R partition of X into m sets so
that the following three conditions are met
X {x1 , x2 ,......xn }
3.Ci C j , i, j 1,......, m
u j : X [0,1], J 1,....m
And
m N
u j ( xi ) 1, 2,.....N ,
j 1
0 u j ( xi ) N , j 1, 2,....m
i 1
These are called membership functions. The value of fuzzy membership function is a
mathematical characterization of clusters which is not precisely defined and each
vector x belongs to more than one cluster simultaneously.
X {x1 , x2 ,......xn }
Ci , i 1,........m
m
Ci X
i 1
Ci C j , i, j 1,......, m
u j : X [0,1], J 1,....m
m N
u ( x ) 1, 2,.....N ,
j 1
j i 0 u j ( xi ) N , j 1, 2,....m
i 1
e) Proximity measures: This parameter measure will quantify the similarities and
dissimilarities between the two clusters and within the clusters with no bias in
selected clusters that each cluster should contributed equally with no domination
among each other. The proximity measure has two property functions which measure
dissimilarity and similarity between two vectors.
Dissimilarity measure
d : X *X R
d 0 R : d 0 d ( x, ) , x,
d ( xi , j )
N m
J q ( , u ) uijq d ( xi , j )
t 1 j 1
Where J q ( , u) is clustering structure and uijq is membership function for each input
Now that the fuzzy clustering is defined the clusters are applied to adaptive layer of
neuro fuzzy interference model.
N m
J q ( , u ) uijq d ( xi , j )
t 1 j 1
Table 4.3.4 b)Statistical Results of ANFIS Grid Partioning Cluster for 35 HRC and
45HRC
Statistical Results of ANFIS Grid Partioning Cluster for Results of ANFIS Grid Partioning Cluster for 45
35 HRC HRC
ANFIS Grid- Error Error MSE RMSE Error Error MSE RMSE
35 mean() STD() mean() STD()
Test Ft(N) -0.0501 0.5101 0.2610 0.5109 7.612*10-4 0.3501 0.128 0.3148
Validation -0.0023 0.714 0.41 0.57 1.62*10-4 0.153 0.0235 0.153
Ft(N)
Axial force (Fa)
Train Fa(N) 1.54*10-5 0.0847 0.0072 0.0847 3.76*10-5 0.0675 0.0045 0.0674
Test Fa(N) 0.0203 0.3453 0.118 0.3448 0.0055 0.1399 0.0195 0.1396
Train Fr(N) 2.4*10--5 0.082 0.0068 0.0823 3.59*10-5 0.0654 0.00043 0.0654
Test Fr(N) 0.0567 0.3802 0.1468 0.3831 -0.0207 0.2173 0.0473 0.0654
Test Tf (min) -6.69*10-4 0.0271 7.28*10-4 0.0270 0.0014 0.0235 5.487*10-4 0.0234
Fig.4.31 Training Error Plots for Ra (Target vs Fig.4.32 Testing Error Plots for Ra (Target
vs Output)
Fig.4.33 Validation Error Plots for Ra (Target vs Fig.4.34 Regression Plots for Ra (Train
Output) /Test/Validate)
Fig.4.36 Training Error Plots for Ft (Target vs Fig.4.37 Testing Error Plots for Ft (Target vs
Output) Output)
Fig.4.38 Validation Error Plots for Ft (Target vs Fig.4.39 Regression Plots for Ft (Train
Output) /Test/Validate)
Fig.4.41 Training Error Plots for Fa (Target vs Fig.4.42 Testing Error Plots for Fa (Target vs
Output) Output)
Fig.4.43 Validation Error Plots for Fa (Target vs Fig.4.44 Regression Plots for Fa (Train
Output) /Test/Validate)
Fig.4.46 Training Error Plots for Fr (Target vs Output) Fig.4.47 Testing Error Plots for Fr (Target vs Output)
Fig.4.48 Validation Error Plots for Fr (Target vs Fig.4.49 Regression Plots for F (Train /Test/Validate)
r
Output)
Fig.4.51 Training Error Plots for Tf (Target vs Fig.4.52 Testing Error Plots for Tf (Target
Output) vs Output)
Fig.4.53 Validation Error Plots for Tf (Target vs Fig.4.54 Regression Plots for Tf (Train
Output) /Test/Validate)
Fig.4.56 Training Error Plots for Ra (Target Fig.4.57 Testing Error Plots for Ra (Target
vs Output) vs Output)
Fig.4.58 Validation Error Plots for Ra Fig.4.59 Regression Plots for Ra (Train
(Target vs Output) /Test/Validate)
Fig.4.61 Training Error Plots for Ft (Target vs Fig.4.62 Testing Error Plots for Ft
Output) (Target vs Output)
Fig.4.63 Validation Error Plots for Ft (Target vs Fig.4.64 Regression Plots for Ft (Train
Output /Test/Validate)
Fig.4.66 Training Error Plots for Fa (Target Fig.4.67 Testing Error Plots for Fa
vs Output) (Target vs Output)
Fig.4.68 Validation Error Plots for Fa (Target Fig.4.69 Regression Plots for Fa (Train
vs Output /Test/Validate)
Fig.4.71 Training Error Plots for Fr (Target Fig.4.72 Testing Error Plots for Fr (Target
vs Output) vs Output)
Fig.4.76 Training Error Plots for Tf (Target vs Fig.4.77 Testing Error Plots for Tf
Output) (Target vs Output)
Fig.4.78 Validation Error Plots for Tf (Target Fig.4.79 Regression Plots for Tf (Train
vs Output) /Test/Validate)
N m
J q ( , u ) uijq d ( xi , j )
t 1 j 1
exponential function
N || xi x j ||
d ( xi , j ) exp[ ]
J 1 j (ra / 2)2
Where j (ra / 2) is the radius of xi & x j are membership function points if a point has
many other points surrounded around itself then point has highest density point.
The highest density point is as first cluster center xc and in the consecutive the
iteration the density measure of each point is obtained by subtracting cluster points
by applying the following equation.
N || xi x j ||
di di d . exp[ ]
J 1 j (ra / 2)2
After the calculation of the dissimilarity at each point first cluster center is identified
as the point having highest density.
Eliminate all points in the vicinity around the first cluster center of its defined radius
value. For the next iteration update the dissimilarity and apply the cluster function.
Table 4.3.15 (b) Statistical Error analysis of ANFIS Subtractive clustering for
35 HRC and 45 HRC
Statistical Results of ANFIS Subtractive Cluster for 35 HRC Results of ANFIS Subtractive Cluster for 45
HRC
ANFIS Grid- Error Error MSE RMSE Error Error MSE RMSE
35 mean() STD() mean() STD()
Surface roughness (Ra)
Train Ra (m) 1.0458*10-7 0.0044 1.922*10-5 0.0044 6.64*10-6 4.32*10-4 2.63*10-7 5.12*10-4
Test Ra (m) -2.82*10-5 0.0052 2.86*10-5 0.0052 1.165*10-5 0.0039 3.48*10-5 5.89*10-3
Validation Ra -4.121*10-6 0.0045 2.036*10-5 0.0045 5.58*10-5 0.0035 1.25*10-5 0.0035
(m)
Tangential force (Ft)
Train Ft(N) -4.28*10-6 0.5928 0.3510 0.5924 -1.29*10-6 0.4917 0.2415 0.4914
Test Ft(N) -0.0135 0.6604 0.433 0.658 0.0128 0.6335 0.3988 0.6315
Validation -0.0020 0.6031 0.3633 0.6028 0.0019 0.515 0.2651 0.5149
Ft(N)
Axial force (Fa)
Train Fa(N) 2.98*10-6 0.3191 0.1017 0.318 3.621*10-6 0.2806 0.07806 0.2805
Test Fa(N) -0.0233 0.4351 0.1886 0.434 0.0414 0.4176 0.1749 0.4183
Validation -0.35 0.3389 0.1147 0.3387 0.0062 0.305 0.0931 0.3051
Fa(N)
Radial force(Fr)
Train Fr(N) -3.86*10-6 0.3432 0.1176 0.3430 4.85*10-6 0.3020 0.0911 0.3013
Test Fr(N) -0.0484 0.380 0.1458 0.3818 -0.0051 0.352 0.1233 0.3512
Validation -0.0073 0.3849 0.121 0.349 -7.62*10-4 0.3099 0.096 0.3098
Fr(N)
Tool Life (Tf)
Train Tf (min) 1.49*10-6 0.0325 0.011 0.0325 7.78*10-4 0.0249 6.19*10-4 0.0249
Test Tf (min) 0.0048 0.0358 0.0013 0.0360 0.0018 0.0283 7.96*10-4 0.0282
Fig.4.82 Training Error Plots for Ra (Target Fig.4.83 Testing Error Plots for Ra (Target
vs Output) vs Output)
Fig.4.84 Validation Error Plots for Ra Fig.4.85 Regression Plots for Ra (Train
(Target vs Output) /Test/Validate)
Fig.4.87 Training Error Plots for Ft (Target vs Fig.4.88 Testing Error Plots for Ft
Output) (Target vs Output)
Fig.4.89 Validation Error Plots for Ft (Target vs Fig.4.90 Regression Plots for Ft
Output) (Train /Test/Validate
Fig.4.92 Training Error Plots for Fa (Target vs Fig.4.93 Testing Error Plots for Fa
Output) (Target vs Output)
Fig.4.94 Validation Error Plots for Fa (Target Fig.4.95 Regression Plots for Fa
vs Output) (Train /Test/Validate)
Fig.4.97 Training Error Plots for Fr (Target Fig.4.98 Testing Error Plots for Fr
vs Output) (Target vs Output)
Fig.4.102 Training Error Plots for Tf (Target Fig.4.103 Testing Error Plots for Tf
vs Output) (Target vs Output)
Fig.4.104 Validation Error Plots for Tf (Target Fig.4.105 Regression Plots for Tf (Train
vs Output) /Test/Validate)
Fig.4.107 Training Error Plots for Ra Fig.4.108 Testing Error Plots for Ra (Target
(Target vs Output) vs Output)
Fig.4.109 Validation Error Plots for Ra Fig.4.110 Regression Plots for Ra (Train
(Target vs Output) /Test/Validate)
Fig.4.112 Training Error Plots for Ft Fig.4.113 Testing Error Plots for Ft
(Target vs Output) (Target vs Output)
Fig.4.114 Validation Error Plots for Ft Fig.4.115 Regression Plots for Ft (Train
(Target vs Output) /Test/Validate)
Fig.4.117 Training Error Plots for Fa (Target Fig.4.118 Testing Error Plots for Fa (Target vs
vs Output) Output)
Fig.4.119 Validation Error Plots for Fa (Target Fig.4.120 Regression Plots for Fa (Train
vs Output) /Test/Validate)
Fig.4.122 Training Error Plots for Fr (Target vs Fig. 4.123 Testing Error Plots for Fr (Target
Output) vs Output)
Fig.4.124 Validation Error Plots for Fr (Target Fig.4.125 Regression Plots for Fr (Train
vs Output) /Test/Validate)
Fig.4.127 Training Error Plots for Tf Fig.4.128 Testing Error Plots for Tf
(Target vs Output) (Target vs Output)
Fuzzy C-Mean clustering is another circular invariant clustering technique the radius
of clusters is calculated by membership functions. Recall the cluster function
N m
J q ( , u ) uijq d ( xi , j )
t 1 j 1
u
j 1
ij
m
* xj
Compute the cluster mean i M
(u
j 1
ij )m
1
Update the member partition matrix uij by uij M
dik
(d
k 1
) 2/ m1
kj
Table 4.3.26 b) Statistical Results of ANFIS FCM for 35 HRC and 45 HRC
Statistical Results of ANFIS Fuzzy C -Mean for 35 HRC Results of ANFIS Fuzzy C -Mean Cluster
Test Ft(N) -0.0224 1.897 3.577 1.893 -0.1034 1.4454 2.085 1.44
Test Fa(N) 0.0835 0.9753 0.953 0.976 -0.0626 0.59206 0.3500 0.5916
Radial force(Fr)
Train Fr(N) -2.149*10- 0.863 0.744 0.862 -3.36*10-7 2.12*10-4 0.0041 0.064
6
Test Fr(N) 0.1296 0.935 0.886 0.941 -0.0112 0.559 0.3132 0.559
Fig. 4.133 Training Error Plots for Ra Fig.4.134 Testing Error Plots for Ra
(Target vs Output) (Target vs Output)
Fig.4.135 Validation Error Plots for Ra Fig.4.136 Regression Plots for Ra (Train
(Target vs Output) /Test/Validate
Fig.4.138 Training Error Plots for Ft (Target Fig.4.139 Testing Error Plots for Ft
vs Output) (Target vs Output)
Fig.4.143 Training Error Plots for Fa Fig.4.144 Testing Error Plots for Fa
(Target vs Output) (Target vs Output)
Fig.4.148 Training Error Plots for Fr Fig.4.149 Testing Error Plots for Fr
(Target vs Output) (Target vs Output)
Fig.4.153 Training Error Plots for Tf Fig.4.154 Testing Error Plots for Tf
(Target vs Output) (Target vs Output)
Fig.4.158 Training Error Plots for Ra Fig.4.159 Testing Error Plots for Ra
(Target vs Output) (Target vs Output)
Fig.4.160 Validation Error Plots for Ra Fig.4.161 Regression Plots for Ra (Train
(Target vs Output) /Test/Validate
Fig.4.163 Training Error Plots for Ft Fig.4.164 Testing Error Plots for Ft
(Target vs Output) (Target vs Output)
Fig.4.165 Validation Error Plots for Ft Fig.4.166 Regression Plots for Ft (Train
(Target vs Output) /Test/Validate
Fig.4.168 Training Error Plots for Fa (Target Fig.4.169 Testing Error Plots for Fa (Target vs
vs Output) Output)
Fig.4.173 Training Error Plots for Fr Fig.4.174 Testing Error Plots for Fr
(Target vs Output) (Target vs Output)
Fig.4.175 Validation Error Plots for Fr Fig.4.176 Regression Plots for Fr (Train
(Target vs Output) /Test/Validate)
Fig.4.178 Training Error Plots for Tf Fig.4.179 Testing Error Plots for Tf
(Target vs Output) (Target vs Output)
4.4.1 Statistical Comparison of Neural Network and ANFIS Prediction Results with
Experimental Statistics for AISI 4340 Steel 35hrc
NN35 Error Error MSE RMSE Error Error MSE RMSE
NN VS Experimental 35 HRC
STD() NN VS Experimental
STD() 45 HRC
Ra 0.0442 1.0814 1.1130 1.0550 -0.1700 0.2769 0.1018 0.3190
Ft (N) 0.186.6 126.4 594.81 24.38 -57.23 128.3 1.8929*104 137.5
4
Fa (N) 103.98 33.77 1.897*10 109.07 54.67 33.57 4.0607*103 63.72
Fr (N) 1.1318 93.95 93.95 9.69 0.415 14.0720 188.29 13.72
ANFIS grid partitioning VS Experimental 35 HRC ANFIS grid partitioning VS Experimental
Ra -0.2308 0.5964 0.3912 0.6255 -0.9278 0.5964 1.198 1.0949
4 3
Ft (N) 17.49 106.9 1.116*10 105.66 -22.35 77.58 6.128*10 78.85
Fa (N) 157 80 3.09*104 176 -7.5709 40.08 1.5841*103 39.8
Fr (N) 32.57 41.29 2.6*103 51.78 -7.69 38.35 1.457*103 38.17
ANFIS Subtractive VS Experimental 35 HRC ANFIS Subtractive VS Experimental 45
Ra -0.226 0.5964 0.3875 0.6225 -0.947 0.594 1.23 1.109
Ft (N) 15.29 105.36 1.07*104 103 -23.106 76.64 6.1143*103 78.19
4
Fa (N) 157.79 82.3 3.133*10 177.016 -6.9912 39.58 1.537*103 39.213
Fr (N) 27.89 40.81 2.3604*103 48.58 -8.525 36.57 1.343*103 36.65
ANFIS FCM VS Experimental 35 HRC ANFIS FCM VS Experimental 45 HRC
Ra -0.2339 0.5964 0.392 0.6266 -0.856 0.596 1.0714 1.0351
Ft (N) 18.19 104.9 1.08*104 103.9 -21.7 78.8 6.38*103 79.8
Fa (N) 157 79.42 3.073*104 175.32 -6.8003 38.45 6.36*10-6 0.4664
Fr (N) 32.23 42.20 2.73*103 52.25 -6.63 36.89 1.33*103 36.56
Fig.4.189 Error Estimation Plots For F a Fig.4.190 Error Estimation Plots For F r
Fig.4.191 Error Estimation Plots For R a Fig.4.192 Error Estimation Plots For F t
Fig.4.195 Error Estimation Plots For R a Fig.4.196 Error Estimation Plots For F t
Fig.4.197 Error Estimation Plots For F a Fig.4.198 Error Estimation Plots For F r
Fig.4.199 Error Estimation Plots For R a Fig.4.200 Error Estimation Plots For F t
Fig.4.201 Error Estimation Plots For F a Fig.4.202 Error Estimation Plots For F r
Fig.4.203 Error Estimation Plots For R a Fig.4.204 Error Estimation Plots For F t
Fig.4.205 Error Estimation Plots For F a Fig.4.206 Error Estimation Plots For F r
Fig.4.207 Error Estimation Plots For R a Fig.4.208 Error Estimation Plots For F t
Fig.4.209 Error Estimation Plots For F a Fig.4.210 Error Estimation Plots For F r
Fig.4.211 Error Estimation Plots For R a Fig.4.212 Error Estimation Plots For F t
4.5 Conclusion
From the Comparison PLOTS and table the mean error and RMS of neural
network was found to be lower compared to ANFIS models
Though the errors in neural network was less the prediction curve for surface
roughness should poor match with the experimental curve while in ANFIS the
curve match for surface roughness was better in comparison to neural
network.
With this results learning techniques for were further attempted to improve with
synergies.
CHAPTER 5
5.1 Introduction
In the previous chapters optimization and prediction techniques were applied exclusively
and the results obtained from them were found to be pretty convincing when compared
with the experimental statistics In this chapter hybridization of these techniques is applied
to our current machining problem with objective to improve the ability of techniques
through mutual assistance and improve prediction and optimization ability of exclusive
techniques. A exposition of adapted synergies is illustrated in brief. In the first section the
combinations of Neuro-Evolutionay and Neuro-Swarm techniques is implemented , in the
second section combination of Evolutionary - Neuro fuzzy and Swarm-Neuro fuzzy is
exercised and in the third segment a .comparison is made between the predicted results
obtained from synergism and experimental statistics. The objective of this chapter is
depicted through the flow chart.
Various types of EAs and NNs synergies are possible which can be broadly classified in
three combinations supportive combination, collaborative combination and amalgamated
combination
Finding an appropriate topology of NN for a given problem is trial and error task.
Synergies between EAs and NN assists in determining optimal network architecture and
then evaluate neural network.
2. Set up the training data for neural network as received from the EA results by
permuting train and target sets and dividing them for training testing and
validation
3. Apply learning criteria for network and evaluate the weights and bias for the
targets. Test the targets against the expected outputs
4. Evaluate the training errors and fitness of network for current learning, transfer
back weights and bias along with the targets to EA. Repeat the step2-4 until the
convergence or max generation is reached.
Similar to EA-NN synergies, SI-NN is has three class of synergies which differ in the
degree of coupling and interdependency in working towards solution. The collaborative
combination strategy is analogues to EA-NN strategy. The figure below illustrates the
collaborative combination of SI and NN.
The synergism between EA and ANFIS is strongly coupled here the EA technique is
toapplied on the membership function which optimizes fitness value of fuzzy output. The
below flow chart outlines the applied strategy in synergism .Membership function are
clustered through Fuzzy C-mean clustering and the fuzzy structure utilized is same as
previously applied (refer chapter 4 table (4.3.27 (a) )). A detailed description is given in
appendix.
5.4.1 ANFIS GA
5.4.1 (a) Statistical Error analysis of GA based ANFIS (FCM) for 35 HRC and
45HRC
Results of GA based ANFIS (FCM) for 35 HRC Results of GA based ANFIS (FCM) for 45 HRC
ANFIS Grid- Error Error MSE RMSE Error Error MSE RMSE
35 mean() STD() mean() STD()
Surface roughness (Ra)
Train Ra -1.5*10-15 0.1016 0.0103 0.1015 3.14*10-16 0.0682 0.0046 0.0681
Test Ra (m) -0.0020 0.1004 0.0100 0.1002 0.0014 0.069 0.0048 0.0692
Tangential force (Ft)
Train Ft(N) 9.22*10-3 23.7 561.42 23.69 -2.69*10-13 18.31 335.08 18.30
Test Ft(N) -0.3057 25.68 657.6 25.64 1.3611 19.58 384 19.59
Axial force (Fa)
Train Fa(N) 1.21*10-13 11.57 133.7 11.56 -1.69*10-13 8.42 70.8 8.4176
Test Fa(N) 0.0194 10.79 116.09 10.77 1.681 9.79 98.45 9.92
Radial force(Fr)
Train Fr(N) 4.037*10-13 7.568 57.2 7.56 -4.3*10-13 6.266 38.70 6.22
Test Fr(N) -0.0449 7.90 62.314 7.894 1.082 7.41 55.97 7.481
Tool Life (Tf)
Train Tf (min) -3.06*10-14 0.7850 0.6153 0.7844 2.69*10-14 0.6304 0.39 0.62
Test Tf (min) 0.034 0.80 0.6431 0.0802 -0.0157 0.16 0.38 0.618
Results of PSO based ANFIS (FCM) (35 HRC) Results of PSO based ANFIS (FCM) 45 HRC
Surface roughness (Ra)
Train Ra (m) -8.43*10-17 0.0988 0.2168 0.4656 1.4*10-15 1.4*10-15 0.0042 0.0681
Test Ra (m) -4.79*10-4 0.1010 0.0102 0.1008 1.4*10-15 1.4*10-15 0.0048 0.0692
Tangential force (Ft)
Train Ft (N) -3.3*10-13 23.87 589.7 23.86 -4.98*10-13 18.57 344.4 12.55
Test Ft (N) -2.214 23.79 569.38 23.86 1.178 18.90 357.6
Axial force (Fa)
Train Fa (N) 1.02*10-13 10.98 120.46 10.96 -5.59*10-13 8.78 77.08 8.78
Test Fa (N) -0.2655 11.98 143.22 11.96 0.080 9.045 81.54 9.03
Radial force(Fr)
Train Fr (N) -2.36*10- 7.57 57.22 7.56 -3.23*10-13 6.57 42.79 6.54
13
Test Fr (N) 0.221 7.71 59.42 7.708 -0.1446 6.73 45.25 6.72
Tool Life (Tf)
Train Tf (min) 3.03*10-14 0.796 0.6335 0.795 2.90*10-14 0.6186 45.25 6.72
Test Tf (min) 0.0042 0.88 0.774 0.8817 -0.0036 0.64 0.416 0.645
Fig.5.43 Training Error Plots for Ra (Target Fig.5.44 Testing Error Plots for Ra (Target
vs Output) vs Output)
Fig.5.47 Training Error Plots for Ft (Target vs Fig.5.48 Testing Error Plots for Ft (Target vs
Output) Output)
Fig.5.51 Training Error Plots for Fa (Target vs Output) Fig.5.52 Testing Error Plots for Fa (Target vs
Output)
Fig.5.55 Training Error Plots for Fr (Target Fig.5.56 Testing Error Plots for Fr (Target
vs Output) vs Output)
Fig.5.59 Training Error Plots for Tf (Target Fig.5.60 Testing Error Plots for Tf (Target
vs Output) vs Output)
Fig.5.63 Training Error Plots for Ra (Target Fig.5.64 Testing Error Plots for Ra (Target
vs Output) vs Output)
Fig.5.67 Training Error Plots for Ft (Target Fig.5.68 Training Error Plots for Ft (Target
vs Output) vs Output)
Fig.5.71 Training Error Plots for Fa (Target Fig.5.72 Testing Error Plots for Fa (Target
vs Output) vs Output)
Fig.5.75 Training Error Plots for Fr (Target Fig.5.76 Testing Error Plots for Fr (Target
vs Output) vs Output)
Fig.5.79 Training Error Plots for Tf (Target Fig.5.80 Testing Error Plots for Tf (Target
vs Output) vs Output)
The synergism between SI and ANFIS is also strongly coupled where the SI technique is applied
on the membership function to optimize fitness value of fuzzy output. The below flow chart
outlines the applied strategy in synergism .Membership function are clustered through Fuzzy C-
mean clustering and the fuzzy structure utilized is same as previously applied (refer chapter 4
table (4.3.27 (a) )). A detailed description is given in appendix.
Fig.5.84 Training Error Plots for Ra (Target Fig.5.85 Testing Error Plots for Ra (Target
vs Output) vs Output)
Fig.5.88 Training Error Plots for Ft Fig.5.89 Testing Error Plots for Ft (Target
(Target vs Output) vs Output)
Fig.5.92 Training Error Plots for Fa (Target Fig.5.93 Testing Error Plots for Fa
vs Output) (Target vs Output)
Fig.5.96 Training Error Plots for Fr (Target Fig.5.97 Testing Error Plots for Fr
vs Output) (Target vs Output)
Fig.5.100 Training Error Plots for Tf Fig.5.101 Testing Error Plots for Tf (Target
(Target vs Output) vs Output)
Fig.5.104 Training Error Plots for Ra Fig.5.105 Testing Error Plots for Ra (Target
(Target vs Output) vs Output)
Fig.5.108 Training Error Plots for Ft Fig.5.109 Testing Error Plots for Ft (Target
(Target vs Output) vs Output)
Fig.5.112 Training Error Plots for Fa Fig.5.113 Testing Error Plots for Fa (Target
(Target vs Output) vs Output)
Fig.5.116 Training Error Plots for Fr Fig.5.117 Testing Error Plots for Fr
(Target vs Output) (Target vs Output)
Fig.5.120 Training Error Plots for Tf (Target Fig.5.121 Testing Error Plots for Tf (Target vs
vs Output) Output)
5.6.1 NSGA-NN
Fig.5.124 Error Estimation Plots For R a Fig.5.125 Error Estimation Plots For F t
Fig.5.126 Error Estimation Plots For F a Fig.5.127 Error Estimation Plots For Fr
Fig.5.128 Error Estimation Plots For R a Fig.5.129 Error Estimation Plots For F t
Fig.5.130 Error Estimation Plots For F a Fig.5.131 Error Estimation Plots For F r
5.6.4 Error Plots of PSO-NN Results with Experimental Statistics for AISI 4340
Steel 35 HRC
Fig.5.132 Error Estimation Plots For R a Fig.5.133 Error Estimation Plots For F t
Fig.5134 Error Estimation Plots For F a Fig.5.135 Error Estimation Plots For F r
5.6.5 Error Plots of GA based ANFIS (Fuzzy C-Mean Clustering)) Results with
Experimental Statistics for AISI 4340 Steel 45hrc
Fig.5.136 Error Estimation Plots For R a Fig.5.137 Error Estimation Plots For F t
Fig.5.138 Error Estimation Plots For F a Fig.5.139 Error Estimation Plots For F r
5.6.6 Error Plots of GA based ANFIS (Fuzzy C-Mean Clustering)) Results with
Experimental Statistics for AISI 4340 Steel 45hrc
Fig.5.140 Error Estimation Plots For R a Fig.5.141 Error Estimation Plots For F t
Fig.5.142 Error Estimation Plots For F a Fig.5.143 Error Estimation Plots For F r
5.6.7 Error Plots of PSO based ANFIS (Fuzzy C-Mean Clustering)) Results with
Experimental Statistics for AISI 4340 Steel 35hrc
Fig.5.144 Error Estimation Plots For R a Fig.5.145 Error Estimation Plots For F t
Fig.5.146 Error Estimation Plots For F a Fig.5.147 Error Estimation Plots For F r
5.6.8 Error Plots of PSO based ANFIS Prediction Results with Experimental
Statistics for AISI 4340 Steel 45hrc
Fig.5.148 Error Estimation Plots For R a Fig.5.149 Error Estimation Plots For F t
5.7 Conclusion
Furthermore the predicted results are tested against experimental statistics for evaluating
the prediction accuracy of each learning model. The prediction models demonstrated
relatively varying accuracy results which are tabulated in each sections.
1. The results of EA NN were more accurate than SI-NN and less accurate when
compare to the synergies of ANFIS.
2. EA-NN accuracy was convincing but the accuracy of ANFIS-EA and ANFIS-SI
were better than the EA-NN and SI-NN prediction models which can be clearly
commented from the statistical error table.
3. Between the ANFIS EA and ANFIS-SI the relative difference of accuracy is
negligible as both the techniques demonstrated almost similar results in prediction
4. The Adaptive neuro fuzzy combination proved to be better than the neuro
computing combination this difference is possible due to the adaptive layers
introduced in the neuro-fuzzy inference. Though in both the techniques back
CHAPTER 6
6.1.1 NSGAII
NSGA II utilized non-domination technique for its population selection and mating, the
fitness of individuals were determined by two criteria (a) Pareto-Individual rank (b)
Pseudo Euclidean distance
According to criteria an individual with least rank and maximum Euclidean distance is
the fittest and these individual/individuals represents globally optimized solutions. The
degree of optimality depends on their relative function of rank and distance which
represents diversity of individuals in each generation. The results of both the steels AISI
4340 35HRC and 45HRC had initial population of 1000 and elitism was applied at each
generation.
The Table 3.3 and Table 3.4 records results of family of optimized individuals for both
steels after hundred generations arranged in descending order of fronts and fitness levels
which are calculated by rank and distance. Only the first 20 solutions are listed out of the
1000 population in the tables.
From Table 3.3 the first front has maximum distance and minimum rank the solutions
and these solutions shows a good tradeoffs between the surface roughness and tool life
among them fifth individual with fitness which gives both good surface finish and
maximum tool life .This individual corresponds to process parameters (cutting speed
,feed rate and depth of cut) 170m/min, 0.15mm/rev, 1mm respectively with machining
objectives(Ra, Ft, Fa, Fr, Tf) 4.0m, 412N, 152N, 231N and 50 minutes. The subsequent
individual with process parameter (cutting speed, feed rate and depth of cut)
231m/min,0.15mm/rev,1mm corresponding to machining performance (Ra, Ft, Fa, Fr, Tf)
2.5m, 442N, 115N, 199N and 42.49 minutes can be picked out as optimal solution.
Though the fitness of first front is fittest among others, the individuals from subsequent
front can also be chosen from the solution space with permitted tradeoffs.
The Fig.3.3-3.5 constitutes plots of NSGA II for 35HRC, in the Fig.3.3 the first subplot
shows the rank of individual at each generation, from the figure it can be inferred that the
rank converges to one for most of the population as non-dominated technique is applied
for sorting, in the second subplot of Fig 3.3 Pareto plots between the objectives are
plotted to evaluate relative tradeoffs between two objectives.
Pareto front depicts plots for elite members which can be between two or three objectives
that are non-inferior. The subplots 2, 3 and 4 in the figure 3.3 are pareto plots between
force and surface roughness, the second subplot represents pareto-front between Ra-Ft, in
third and fourth sub plot pareto front between Ra-Fa and Ra-Fr.
The Fig 3.4 contains pareto-fronts between tool life and forces; first subplot is pareto-
front between Tf -Ft, second subplot between Tf-Fa and third sub plot between Tf-Fr. From
the pareto fronts for minimum surface roughness and maximum tool life following region
range of forces were found to be favorable (approximately) optimal surface profile.
Table.6.1 Tradeoffs among forces for Surface roughness and Tool life
Forces Surface Roughness(1-2.5 m) Tool life (Tf>40 mins)
Tangential Forces 800-600 N <470N
Axial Forces 300-200 <160N
Radial forces 400-300 <250N
In fig. 3.5 average pareto diversity plots in the consecutive generations is plotted which is
determined by average pareto (Euclidean) distance among individuals The diversity of
population varied from 1-0.001 The population in overall generations exhibited vivid
diversity. Initially the average distance between the generations was maximum after 20
generations a shift in distance was observed with distance spread to 0.6 in mid generation
and as final generations are approached the distance between generations converges to
0.001 showing a good migration of individuals across generations.
For AISI 4340 45 HRC steel the family of solutions is listed in table 3.4 out of the 1000
elite solutions only 20 solutions are tabulated. The best individual in the first front is
corresponding to process parameters (cutting speed, feed rate and depth of cut)
130.8m/min, 0.15 mm/rev, 1mm and objective fitness (Ra, Ft, Fa, Fr, Tf) 4.0 m, 542N,
331N, 282N, 30 minutes is fittest. Similarly preceding individual with process parameters
(cutting speed, feed rate and depth of cut) 135m/min, 0.15mm/rev, 1mm corresponding to
objective fitness (Ra, Ft, Fa, Fr, Tf) 3.98, 519N, 331N, 282N, 28 minutes can be picked as
best solution. Likewise favorable solutions can be chosen with the degree of tradeoffs
permitted among the objectives.
The Fig.3.6 - 3.8 are plots of NSGA II for 45HRC, in the Fig.3.6 the first subplot depicts
rank of individual at each generation and the rank converges to one for most of the
population, in the second subplot of Fig.3.7 Pareto plots between the objectives are
graphed and in Fig.3.8 diversity plot at each generation is drawn.
The subplots 2, 3 and 4 in the figure 3.6 are Pareto plots between surface roughness and
force, the second subplot represents pareto-front between Ra-Ft, the third and fourth sub
plot shows pareto front between Ra-Fa and Ra-Fr.
The Fig.3.7 contains pareto-fronts between tool life and forces; first subplot is pareto-
front between Tf-Ft, second subplot between Tf-Fa and third sub plot between Tf-Fr. From
the pareto fronts for minimum surface roughness and maximum tool life following region
range of forces were found to be favorable (approximately) for optimal solution in
45HRC machining.
Table.6.2 Tradeoffs among forces for Surface roughness and Tool life
Forces Surface Roughness(3.5-4m) Tool life (Tf>10 mins)
Tangential Forces 600-400 N <500N
Axial Forces 360-300N <160N
Radial forces 300-330N <350N
Since SPEA 2 is a combination of pareto envelope and nichepareto the rank and fitness
strength is calculated to both best elements in archive and individuals in population
through dominance level. The proposition rule of determining best solution is by
calculating effective fitness value which is sum of raw fitness and density distribution
about individuals. Table 3.6 holds optimal solution for AISI 4340 steel 35 HRC which
contains 300 archive elements. The matrix F contains the fitness strength of each
individual which calculated by summation of R (rank) and D (distance /density function)
matrix (in table 3.6), the distance D is calculated by K nearest cluster algorithm .The
matrix S is strength value which is evaluated by dominance count. The position matrix
represents process parameters and the cost matrix represents machining objectives.
The individual with least strength value represents the fittest individual i.e., the
individual in first front with strength fitness S 0.32 can be picked as best individual with
position (process parameters :cutting speed, feed rate and depth of cut) 259m/min,
0.17mm/rev, 1.07mm corresponding to machining objectives (Ra, Ft, Fa, Fr, Tf) 1.88m,
514N, 136N, 225N, 35.51 minutes.
In Fig.3.10 subplots have individual ranks and pareto fronts between surface roughness-
forces, the subplots 2, 3, 4 depicts the pareto front between Ra-Ft, Ra-Fa, Ra-Fr, which are
archive element containing only fittest individual. Similarly the pareto fronts between
tool life and forces are plotted in fig.3.11 the subplots 1,2,3 are pareto fronts for Tf- Ft,
Tf- Fa, Tf- Fr are represented. Though the solutions suggested by SPEA 2 is different
individuals from the NSGA II the force constraint obtained is same as NSGA II for
minimum surface roughness and maximum tool life
Table.6.3 Tradeoffs among forces for Surface roughness and Tool life
Forces Surface Roughness(1-2.5m) Tool life (Tf>35 mins)
Tangential Forces 780-590 N <500N
Axial Forces 300-200N <160N
Radial forces 300-330N <250N
In Fig.3.12 diversity among generation is plotted the average pareto distance is different
from the NSGA II and the spread is limited between 0.37-0.27 the spread is random
representing a good mix in individual among generations but the combination is restricted
between limited distance.
The results of 45 HRC steel is recorded in the Table 3.7 with same definition of elements.
From the table the strength fitness is least for third individual with F 0.31 the position
(process parameters: cutting speed, feed rate and depth of cut) of this individual is 174
m/min,0.23 mm/rev, 1.7 mm with machining performance (Ra, Ft, Fa, Fr, Tf) 3.93m,
674N, 389N, 355N, 14.19 minutes and next fittest individual is in consequent front with
strength fitness S 0.33 with process parameters168m/min,0.16mm/rev,1.02mm for
machining performance 3.67m,519.42N,384N,310N,19 minutes.
Pareto fronts for 45 HRC objectives are in Fig.3.13 and Fig.3.14,in fig 3.13 the sub plot
2,3,4 represents the pareto front between Ra-Ft, Ra-Fa, Ra-Fr, and subplots in Fig.3.14
shows subplots of Tf-Ft, Tf - Fa, Tf- Fr respectively for minimum surface roughness and
maximum tool life the force constraints should be as follows
Table.6.4 Tradeoffs among forces for Surface roughness and Tool life
Forces Surface Roughness(3.5-4m) Tool life (Tf>10 mins)
Tangential Forces 600-400 N <500N
Axial Forces 360-300N <160N
Radial forces 300-330N <350N
The PSO algorithm utilizes swarm movement to find optimal solution. The applied
swarm constitutes of particle structure with position, cost and their best position, best cost
associated with it a pseudo velocity and acceleration associated for each particle. Table
3.9 comprises of optimal solution for AISI 4340 steel where the position matrix holds
process parameters cost holds the machining performance for each position best cost ,best
position is defined for each position. The velocity matrix is which correspond to the
position matrix is utilized to move swarm in search space and changes at each transition.
The Grid Index matrix contains the topology of swarm at each transition and grid sub
index contains the neighborhood topology of grid index. The table consists of 500 swarm
particles which are in repository element out of 1000 swarm.
In Fig.3.17 pareto front between the surface roughness and tangential force is graphed
which gives similar results as pareto front of Evolutionary Techniques.
Fig.3.18-3.19 are swam surfaces for surface roughness and tool life respectively which
shows the potential of swarm. Each particle on the swarm surface is defined by particle
position in Fig.3.18 the particles at the declination of surface are fittest particles which
converged to minimum surface roughness values and in Fig.3.19 the particles at the
projection of swarm surface represents the optimal tool life showing convergence at the
foot of swarm surface.
The solution for AISI 4340 45 HRC steel has same entities which are listed in table
3.10.The pareto-fronts and swarm surface are plotted in fig 3.21 and fig 3.22-3.23
respectively. The swarm surface of 45HRC showed similar trends as to 35 HRC steel the
45HRC steel shows similar trends in swarm surface. In Fig.3.21 the pareto front between
surface roughness and tangential force is drawn which gives force constraints similar to
NSGA II 45 HRC fronts .In fig 3.22 and fig 3.23 the swarm surface for surface roughness
and tool life is drawn.
In fig 3.24 the influence of depth of cut on force component on optimal particles is
plotted. The subplots 1, 2 and 3 show the force gradients with respect to depth of cut for
best particles.. From the table 3.10 optimal solution is determined by swarm leaders
among them the first leader corresponds to gird index 20510 with best position 136
m/min, 0.15mm/rev, 1mm with machining performance (Ra, Ft, Fa, Fr, Tf) 4.04 m, 522N,
134N, 331N, 29.32 minutes the subsequent solution is corresponding to particle with grid
index 31070 similarly other solutions are equally fit depending on the allowed tradeoffs
solutions can be chosen.3
The Fig.3.25 - Fig.3.30 are plots of solution spectrum obtained from applied global
search optimization technique. Each plot has solutions for five objectives i.e, Surface
roughness, tool life, radial force, axial force and tangential force. In fig 3.25 and fig 3.26
the solution space obtained from NSGA II for 35 HRC and 45 HRC steels is drawn.
Clearly distinction can be made the graphs that in 35 HRC steel radial forces are
dominating axial force while in 45 HRC steel axial force is dominating radial force. A
similar trend can be inferred from solutions of SPEA2 and PSO.
Fig 3.27 and Fig 3.28 are solution space obtained from the PSO technique for 35 HRC
and 45 HRC steel respectively and Fig.3.29 and Fig.3.30 are solution space obtained
from the SPEA2 for 35HRC and 45HRC respectively.
For 35 HRC the sequence of solution for each objectives are represented in following
order surface roughness, tool life, axial force, radial force and tangential force ,while for
35 HRC the solution for each objective is in order of surface roughness, tool life ,radial
force, axial force and tangential force.
The solution space of NSGA II (fig 3.25 & 3.26) has thousand populations which is a
combination of parents and offspring formed by tournament selection. The solution space
is distributed with constant amplitude in solution space. While the solution space of PSO
and SPEA2 (fig.3.27-fig.3.30) vary in amplitude in solution space at each generation.
The solution space of PSO (Fig 3.27 & fig 3.28) has five hundred best solutions which
are elite solutions. The solution trend is random when compared to NSGAII showing
varying amplitude across wavelength with immediate change in crest and troughs in local
and global minima and maxima.
However the solution space of SPEA2 has three hundred best elite solutions which are
highly disturbed amplitude when compared to NSGAII and PSO the jumps in maxima
and minima is uneven exhibiting unsaturation in local and global minima .
From the solution space analysis the NSGAII showed a better saturation in local global
maxima minima while PSO should a moderate saturation and SPEA2 exhibited lower
saturation levels in maxima and minima. The saturation levels also depicts tradeoffs in
objectives from which an following inference can be made.
The solution obtained from the NSGAII solution gave better tradeoffs among objectives
which can be observed from each front in solution space. While PSO gave a moderate
trade off with the solution space was favoring lower surface roughness by reducing tool
life for few swarm particles though the overall trade off was found to be good enough.
The solution space of SPEA2 gave unbalance tradeoff between objectives favoring
surface roughness and reducing tool life for many individuals. From above discussion it
can be concluded that the solution space provided by NSGAII is better than other two
while the solution space provided by the PSO is better than SPEA2.
The solutions obtained by fittest individuals in NSGAII and PSO keeps the flank wear of
tool under working limit. The tool used for current hard turning was multi-layer coated
carbide inserts and the cutting speed suggested by both the optimization techniques
suggested cutting speed below 200 m/min and feed rate, depth of cut in between LFLD
(low federate low depth of cut) and HLHD (high feed high depth of cut) at this condition
the flank wear is less than 0.15 mm for tool life greater than 40 minutes [2]. This cutting
condition keeps flank wear under appreciable level, restricting sharp rise in cutting forces
due to high flank wear rate [3]. If the flank wear is inhibited above this then machining
chatter due to excessive forces is controlled and better surface finish is obtained.
In contrast SPEA2 suggests cutting speed close to 260m/min (250-260 m/min) and feed
rate, depth of cut near (HLHD condition) for which flank wear about 0.2mm for time
cutting greater than 35 minutes.[2]. At this flank wear the forces tend to increase sharply
which may increase machining vibrations resulting in poor surface finish. The tradeoff
between the surface roughness and tool life should be picked wisely for a successful hard
turning.
Intelligent learning systems were applied to recognize machining pattern sequence which
is discernible in the machining statistics. In fig 4.1 the adopted strategy for chapter 4 is
illustrated and accordingly the results were discussed. Learned networks were utilized to
predict machining performance on experimental runs and obtained results were analyzed
for statistical error. An extensive mathematical framework is discussed for the all the
applied learning techniques.
The applied network architect for both steels is illustrated in fig 4.4. A multi-layer feed
forward perceptron type network was used to for both steels which require many
exemplars for mapping multiple vectors. From the regression model thousand machining
data were generated for each vector model for better interpolation of target vectors in
learning space. These data sets were randomly permutated and split for training, testing
and validating samples. Out of the thousand sets 700 samples were utilized for training
150 sample for training and 150 for validation.
The description of applied neural network for AISI 4340 steel 35 HRC is listed in table
4.2.4 (a) and the calibrated weights and bias for the targets is given in table 4.2.4 (b). The
network performance plot in Fig.4.5 shows drop in mean square error from 104 to 10-3 in
1000 epoch while learning (training, testing and validating stage). The error gradient
Fig.4.6 between the targets and output converged to 10-2 at 1000 epoch from gradient of
105. For each objective an error histogram and regression plots were drawn in (Fig. 3.7-
Fig. 3.16) in each error plot the mean error was close enough to the zero error line
depicting that the error minimized to its least possible value. Linear regression fit was
obtained between targets and output while learning each objective with regression co
efficient close to 1.
Same network was utilized for leaning machine statistics of 45HRC steel. The network
description and calibrated weights and bias are listed in Table 4.2.5(c) and 4.2.5(d).The
mean square error while learning converged to similar order (Fig.5.17) to that of 35HRC.
While in training state the error gradient gave sharp slopes though at final epoch the
gradient converged to 1.125 with zero validation fails. The error histogram and regression
plot for each vector is in (Fig.4.19-Fig.4.28).The error for each machining vector was in
order of 10-2.
In this network two adaptive layers with learnable parameters were used viz., antecedent
and consequent layer. Three different clustering techniques were applied on membership
function i.e., Grid portioning clustering, Subtractive clustering and Fuzzy C- mean
clustering. Each cluster techniques had different weights (connections) and the fuzzy
structure of each technique varied. The analytical description of hybrid neuro-fuzzy
techniques is discussed in brief. For each machining vector fuzzy structure learns
machining data sequentially. From the RSM model thousand machining data for learning
were generated. These vectors were randomly permutated and split into training (850
samples), testing (150 samples) and validated on complete machining vector.
At each epoch the error between targets and output while learning was calculated and
error gradients were minimized through hybrid leaning and back propagation algorithm.
In table 4.3.4 (b) statistical errors in training, testing and validating for each objective in
35HRC is listed and mean error for each vector were as low of order 10-5 these statistical
results are plotted in Fig.4.36-Fig.4.60.
Same architect and fuzzy structure is utilized for learning machining pattern for 45 HRC.
The statistical error in training, testing and validation for each objective is tabulated in
table 4.3.4 (c) the error for was as low as low as 10-5 and plots for each objective are
illustrated in between Fig.4.61-Fig.4.85.
The developed structure for subtractive clustering is illustrated in fig.4.86, fuzzy structure
implemented is tabulated in table 4.3.16 (a) and in table 4.3.16 (b) &4.3.16 (C) statistical
errors in training, testing and validating for objective in 35 HRC and 45HRC respectively
is listed and these statistical results are plotted for each objective.
The developed structure for FCM clustering is shown in fig.4.137 and fuzzy structure
implemented is tabulated in table 4.3.27 (a). In table 4.3.27 (b) & 4.3.27 (C) statistical
errors in training, testing and validating for both steels is listed. The learning error in
FCM was further reduced to order of 10-8 in both the steels though not for all vectors The
mean error in all vectors was of similar order (10-6) showing similar results as that of
subtractive clustering. The plots for each objectives and their corresponding figures for
35HRC and 45HRC steel obtained from ANFIS FCM are shown in Fig.4.138-Fig.4.187.
The developed predictive model was tested on Experimental statistics and a statistical
analysis of the errors in predictions was made for each model. Comparison graphs for
each objective in each technique were plotted (fig.4.188-fig.4.219). The Mean errors and
their standard deviations of each objectives in both the steel are listed in table (6.5) and
the means squared error (MSE), root mean squared error (RMSE) is tabulated in table.
Table.6.5 Mean error and Standard Deviation between Experimental and Predicted
statistics
Machining Neural ANFIS ANFIS ANFIS Neural ANFIS ANFIS ANFIS
Objective Network Grid Sub FCM Network Grid Sub FCM
Though the statistical errors in neural network was relatively in comparison to ANFIS
models the relative change in error at each prediction points in neural network is higher
than the ANFIS models and the curve traced by neural network in both the steels did not
match well when compared to curved traced by ANFIS models. This observation can be
inferred by analyzing corresponding comparison graphs for each objectives in specific
techniques. The error plots for each technique are drawn in Fig.4.188-4.219.
The results obtained from the prediction models were accurate enough to predict around
the experimental statistics, though the relative degree of accuracy varied for different
learning technique. The mean error of neural network was less than the neuro-fuzzy but
the prediction curve could not trace well with the experimental curve in converse the mean
error of ANFIS was comparatively larger but the prediction curve traced well with the
experimental curve.
6.3 Synergies of CI
6.3.1 EA-NN
The mean square error for network converged to order of 10-3 just in 300 epoch exhibiting
good convergence in network error. The plot for network performance is in Fig.5.3.The
training state plots is illustrated in Fig.5.4 the error gradient dropped to 0.194 which is
lesser than the neural network. Likewise the error plots and regression for each vector is
in Fig.5.5-Fig.5.15.
Same architect is utilized for 45 HRC steel the learning performance was better than the
exclusive neural network with mean square error as low as 10-4 in training, testing,
validation stages. The performance plot is drawn in Fig.5.16 and the training state plots
are in Fig.5.17, error gradients were also found to be minimized to gradient as low as 10 -,
error and regression plots for each vector are in Fig.5.16-Fig.5.41.
6.3.2 SI-NN
Applied architect for PSO-NN combination is illustrated in Fig.5.29. The best solutions
in archive elements of PSO were utilized for network learning. The archive had 500
elements out of which 350 samples were utilized for training, 75 samples for testing and
75 samples for validation. The network description and calibrated weights are tabulated
in table 5.3.1 (a) and table.5.3.1 (b). Two hidden layers were defined for network to work
accurately since the examples provided by PSO were less compared to NSGA II. For
each input and output vectors 20 weights were utilized for mapping and 10 bias were
added at each hidden layer.
The network performance of combined SI-NN was better than NN with mean square
error dropping as low as 10-4 in all the learning state and error gradient going as low as
10-4 at final epoch. The performance plot of network is illustrated in Fig.5.30 and training
state plots are in Fig.5.31.The overall network performance was better than the exclusive
neural network.
To enhance ANFIS, GA and PSO are applied to optimize membership function, FCM
structure from previously developed model was utilized for combination. The applied
strategy utilized for ANFIS-GA is explained in Fig.5.42, the architect utilized is same as
that of ANFIS-FCM. Thousand learning samples were generated from RSM model these
exemplars were randomly permuted and used for leaning. Out of thousand exemplars 700
were split for training sets and 300 for testing. Since fuzzy maps only one vector per
training the multi objective NSGA was reduced to single objective GA with no change in
genetic and selection operators.
From the error table 5.3.1 (a) the mean error for each vector of 35HRC dropped to order
of 10-13 which is way better compared to exclusively applied ANFIS (10 -6) demonstrating
excellent improvement in pattern learning ability. The error plots for each vector are
drawn in Fig.5.44-Fig.5.64.Similarly for 45 HRC from the Table.5.3.7 mean error for
each vector dropped to order of 10-13 exhibiting better leaning than exclusive ANFIS.
The developed predictive model was tested on Experimental statistics and a statistical
analysis of the errors in predictions was made for each model. Comparison graphs for
each objective in each technique were plotted (fig.4.188-fig.4.219).The Mean errors and
their standard deviations of each objectives in both the steel are listed in table 6.7 and the
means squared error(MSE), root mean squared error (RMSE) is tabulated in table 6.8
Table.6.7 Mean error and Standard Deviation between Experimental and Predicted
statistics
Machinin NSG PSO-NN ANFIS- ANFIS- NSGA PSO-NN ANFIS ANFIS-
g A-NN GA PSO -NN -GA PSO
Objective
Error Mean Error Standard Deviation
Measure
Comparison Errors for 35 HRC AISI 4340 Steel
Ra(m) -0.221 1.749 -0.1572 0.1578 0.979 0.732 0.4507 0.4494
Ft (N) 42.67 -53.25 17.6 17.13 86.21 104.6 57.06 57.16
Fa (N) 156.2 138.22 136.152 135.9 49.56 40.44 51.87 52
Fr (N) -0.723 40.08 11.065 10.89 14.77 64.63 20.54 70.79
Comparison Errors for 45 HRC AISI 4340 Steel
Ra(m) -0.355 0.0371 0.0372 0.378 0.4507 0.2796 0.2798
Ft (N) -57.23 -21.84 -27.34 128.36 57.06 107.08 106.19
Fa (N) -13.57 -45.95 -46.83 38.046 51.87 41.46 41.55
Fr (N) -11.74 -0.91 -0.747 35.99 20.54 36.86 36.77
Table.6.8 Mean Square Error and Root Mean Square Error between Experimental and
Predicted statistics
From table 6.7 the values of mean error and standard deviations of ANFIS combinations
were less than the Neural Network synergies. Similarly (table 6.8) the mean square error
and root mean square error for ANFIS synergies were lesser compared to Neural
Network Synergies. The amplitude of error between experimental and predicted statistics
was shorter compared to Neural Network.
The ANFIS combination performed better than the NN combination this enhancement is
possibly due to the optimization of adaptive layers. Though the statistical error is not very
large when compared in magnitude the curve traced by ANFIS synergies were found to
better than the curve traced by NN synergies. The error plots for each technique are
illustrated in Fig.5.126-Fig.5.153.
In brief the combinations of ANFIS performed better than the combinations of Neural
Network. However the developed prediction techniques were accurate enough to learn
experimental machining statistics.
CHAPTER 7
CONCLUSION
On an overview the work dealt with optimizing machining performance in hard turning
of AISI 4340 steel and building accurate predictive models by applying intelligence
learning techniques.
From the results of optimization the best tradeoffs among machining objectives was
obtained by NSGA II solution space, followed by PSO while solution space of
SPEA2 was biased toward surface roughness recommending lower tool life in most
of its best solution.
The difference in tradeoffs recommended by algorithms is due to the degree of
elitism demonstrated by algorithms Elitism is similar to a pseudo memory that is
associated to algorithm by which it recognizes the best individual so that it doesnt
search for same individual in again in consecutive reducing exploration time for
searching best individual.
The elitism in NSGA II was implemented by tournament selection where
consecutive best individuals selected at each tournament are preserved by
replacement of chromosome in intermediate solution. Elitism was performed on
single set of population i.e. both fittest individuals and intermediate chromosomes
were members of chromosome. Though achieving elitism by single set of
chromosome slowed down the convergence while recursion. This drawback was
overcome by providing randomness in genetic operator. The mutation operator acted
as agent of entropy in solution space when elite individuals enveloped to local
minimums.
While in PSO the elitism was carried out by introducing pbest and gbest population
sets where the local minimum were stored in pbest and global minimum in gbest.
The repository element in PSO was elite members which contained gbest
populations. The elite mechanism in PSO was dynamic compared to NSGA II in
PSO, the gbest and pbest attractor in neighborhood topology acted as initiator for
gathering elite members. The problem of local minimum was overcome by the
swarm movement equation Unlike NSGA, PSO doesnt have genetic operator so
when the size of elite members exceeded inertia damping co efficient controlled it,
for better convergence a pseudo mutation on particles were introduced.
Similarity in solution space in NSGA II and PSO could be justified by finding
analogies between both algorithms. The best parents in NSGA acted as pseudo
particle attractor in recognizing elite members which is similar to the gbest and pbest
attractor in grid topology and the genetic operator is similar to that of swarm
movement operator Since the elite members in both the algorithms were same (i.e.,
500) which justly supports analogy.
But when solution space of SPEA 2 is compared the fittest individuals is quite
different from NSGA and PSO algorithm it shows inclination towards one vector
weakening other vector even though elitism is fairly applied. This behavior is drawn
from the niche behavior of fit individuals in archive (elite members).
used to optimize the weights in hidden layer and reduce the error between targets and
outputs the prediction results obtained from the NSGA-NN were better in
comparison to PSO-NN.
The ANFIS combination Technique exhibited better learning trends than the ANFIS
applied exclusively. Both the combinations ANFIS-GA and ANFIS-PSO gave better
results in comparison to experimental statistics. The accuracy of both techniques was
similar with minor difference in learning error. In comparison to neural network
synergies the ANFIS synergies gave better results on experimental statistics and
illustrated enhanced learning.
To summarize the combined predictive models performed better in comparison to the
exclusive techniques and in optimization techniques the NSGA II and PSO algorithm
gave relatively good tradeoffs in MOOPS when compared to SPEA2 algorithm.
Future Scope
The unexplored Meta heuristic techniques can be applied for much better tradeoffs in
MOOPS and further enhancement in learning technique is possible by applying further
introducing stronger coupling among the prediction and optimization techniques.
The applied strategy for current machining system can be generalized to other
conventional and non-conventional machining systems.
REFERENCES
[1] Andries P. Engelbrecht Computational Intelligence An Introduction John Wiley &
Sons Ltd ISBN 978-0-470-03561-0.
[2] Satish Chinchanikar, S.K. Choudhury, Effect of work material hardness and cutting
parameters on performance of coated carbide tool when turning hardened steel: An
optimization approach Measurement 46 (2013) pp. 15721584.
[3] Wear behaviours of single-layer and multi-layer coated carbide inserts in high speed
machining of hardened AISI 4340 steel Journal of Mechanical Science and
Technology 27 (5) (2013) 1451~1459
[4] Satish Chinchanikar, S.K. Choudhury, Hard turning using HiPIMS-coated carbide
tools: Wear behaviour under dry and minimum quantity lubrication (MQL)
Measurement 55 (2014) 536548.
[5] Satish Chinchanikar, S.K. Choudhury Investigations on machinability aspects of
hardened AISI 4340 steel at different levels of hardness using coated carbide tools
Int. Journal of Refractory Metals and Hard Materials 38 (2013) 124133.
[6] Gaurav Bartarya S.K. Choudhury State of the art in hard turning International Journal
of Machine Tools & Manufacture 53 (2012) 114.
[7] Satish Chinchanikar, S.K. Choudhury Machining of hardened steel Experimental
investigations performance cooling techniques: A review International Journal of
Machine Tools & Manufacture 89 (2015) 95109.
[8] Satish Chinchanikar, S.K. Choudhury Cutting force modeling considering tool wear
effect during turning of hardened AISI 4340 alloy steel using multi-layer
TiCN/Al2O3/TiN-coated carbide tools Int. J. Adv. Manuf. Technol. (2016) 83:1749
1762 DOI 10.1007/s00170-015-7662-5
[9] Predictive modeling for flank wear progression of coated carbide tool in turning
hardened steel under practical machining conditions Int. J Adv. Manuf. Technol.
(2015) 76:11851201 DOI 10.1007/s00170-014-6285-6
[10] T. G. Ansalam Raj & V. N. Narayanan Namboothiri, An Improved genetic
algorithm for the prediction of surface finish in dry turning of SS 420 materials Int J
Adv Manuf Technol (2010) 47:313324 DOI 10.1007/s00170-009-2197-2.
[11] Hesam Shahali, M Reza Soleymani Yazdi, Aminollah Mohammadi and Ehsan
Iimanian, Optimization of surface roughness and thickness of white layer in wire
electrical discharge machining of DIN 1.4542 stainless steel using micro genetic
algorithm and signal to noise ratio techniques Proc IMechE Part B: J Engineering
Manufacture 226(5) 803812_ IMechE 2012.
[12] A. Garg & L. Rachmawati & K. Tai Classification-driven model selection
approach of genetic programming in modelling of turning process Int J Adv Manuf
Technol (2013) 69, pp. 11371151.
[13] Khaider Bouacha, Mohamed Athmane Yallese, Samir Khamel, Salim Belhadi
Analysis and optimization of hard turning operation using cubic boron nitride tool
Int. Journal of Refractory Metals and Hard Materials 45 (2014), pp. 160178.
[14] Yiit Karpat & Turul zel, Multi-objective optimization for turning processes
using neural network modelling and dynamic-neighbourhood particle swarm
optimization Int J Adv Manuf Technol (2007) 35, pp. 234247.
[15] M. Al-Ahmari Prediction and optimisation models for turning Operations ISSN:
0020-7543, pp. 1366-588
[16] Adel T. Abbas Karim Hamza Mohamed F. Aly and Essam A. Al-Bahkali,
Multiobjective Optimization of Turning Cutting Parameters for J-Steel Material
Advances in Materials Science and Engineering Volume 2016, Article ID 6429160, 8
pages Hindawi Publishing.
[17] ZhenhuaWang,Juntang Yuan, Zengbin Yin and Chao Li, Study on high-speed
cutting parameters optimization of AlMn1Cu based on neural network and genetic
algorithm Advances in Mechanical Engineering 2016, Vol. 8(4), pp. 112 .
[18] Yunguang Zhou & Yadong Gong& Zongxiao Zhu & Qi Gao & Xuelong Wen
Modelling and optimisation of surface roughness from micro grinding of nickel-based
single crystal super alloy using the response surface methodology and genetic
algorithm Int J Adv Manuf Technol (2016) 85:2607 2622
[19] Shahram Saeidi & Maghsud Solimanpur & Iraj Mahdavi & Nikbakhsh Javadian
A multi-objective genetic algorithm for solving cell formation problem using a fuzzy
goal programming approach Int J Adv Manuf Technol (2014) 70, pp. 16351652.
[38] Chen, Z. and Li, Y. (2008). An improved particle swarm algorithm and its
application in grinding process optimization, Proceedings of the 27th Chinese Control
Conference. 2-5
[39] Rajkamal Shukla, Dinesh Singh Experimentation investigation of abrasive water
jet machining parameters using Taguchi and Evolutionary optimization techniques
Swarm and Evolutionary Computation journal .via Taguchi method-based response
surface analysis. Measurement 45 (2012) 785794
[40] Ilhan Asiltrk a, Mehmet unkas, Modeling and prediction of surface roughness
in turning operations using artificial neural network and multiple regression method
Expert Systems with Applications 38 (2011) 58265832
[41] N Senthilkumar, T Tamizharasan, Flank wear and surface roughness prediction
in hard turning Via artificial neural network and multiple regressions.
[42] Miron Zapciu Jean-Yves Knevez Alain Grard Olivier Cahuc Claudiu
Florinel Bisu Dynamic characterization of machining systems Int. J. Adv. Manuf.
Technol. (2011) 57:7383DOI 10.1007/s00170-011-3277-7.
[43] Dilbag Singh & P. Venkateswara Rao, Flank wear prediction of ceramic tools in
hard turning Int J Adv Manuf Technol (2010) 50:479493 DOI 10.1007/s00170-010-
2550-5
[44] Yahya Isik, Investigating the machinability of tool steels in turning operations
Materials and Design 28 (2007) pp. 14171424.
[45] Hamdi Aouici, Mohamed Athmane Yallese, Kamel Chaoui, Tarek Mabrouki,
Jean-Franois Rigal. Analysis of surface roughness and cutting force components in
hard turning with CBN tool: Prediction model and cutting conditions optimization
Measurement 45 (2012), pp. 344353.
[46] Suha K. Shihab, Zahid A. Khan, Aas Mohammad and Arshad Noor Siddiquee,
Investigations on the Effect of CNC Dry Hard Turning Process Parameters on Surface
[47] Waleed Bin Rashid & Saurav Goel & J. Paulo Davim & Shrikrishna N.Joshi
Parametric design optimization of hard turning of AISI 4340 steel (69 HRC) Int J
Adv Manuf Technol (2016) 82, pp. 451462.
[75] Yosra JARRAYA, Souhir BOUAZIZ, Adel M. ALIMI Ajith ABRAHAM Fuzzy
Modeling System based on Hybrid Evolutionary Approach Machine Intelligence
Research Labs, WA, USA.
[76] J. Kennedy, R. Eberhart, Particle Swarm Optimization, From Proc. IEEE Intl.
Conf. on Neural Networks (Perth, Australia), IEEE Service Centre, Piscataway, NJ,
IV:19421948, 1995.
Conferences
APPENDIX: B CERTIFICATES
Appendix
ANFIS Algorithm
//Load data
//Shuffle data
S=randompermutation([data])
[Inputs, Targets]=[Inputs(S,:),Targets(S,:)]
Do Training objectives
Switch case
Case 1
Params
No of mf nmfs : 5
Rule_n=prod(nmfs)
Fis.name=anfis
Fis.and method=prod
Fis.implication method=prod
Fis.agrregation=max
Case 2
Cluster method
Influence Radius
Case 3
No of clusters (ncluster)
Maximum no of iteration
Fis = generate fis(Train input, Train target ,ncluster, FCM options,Optimization method)
Training params [max epoch, error goal,initial step size, step size decreament rate, step
size increament rate]
//Statistical Analysis
ANFIS-GA/PSO Algorithm
ANFIS GA/PSO
//Load data
//Shuffle data
S=randompermutation([data])
[Inputs, Targets]=[Inputs(S,:),Targets(S,:)]
Switch case
Case 1
Case 2
Switch case
Case 1
Params
No of mf nmfs : 5
Rule_n=prod(nmfs)
Fis.name=anfis
Fis.and method=prod
Fis.implication method=prod
Fis.agrregation=max
Case 2
Cluster method
Influence Radius
Case 3
No of clusters (ncluster)
Maximum no of iteration
P=x*p0
X=data.Train inputs
t=data.Train targets
Y=evaluate(x,fis)
Error :[t]-[y]
MSE : mean(Error)
P=[]
N=size[fis.input]
For i [1-N]
Nmfs =size[fis.Input[i].mf]
For j [1-nmfs]
P=[p, fis.Input[j].mf.params]
Noutput=size[fis.output]
For i [1-noutputs]
Nmfs =size[fis.output[i].mfs.params]
For j [1-nmfs]
P=[p, fis.Input[j].mf.params]
P0=[]
X= data.Train Inputs
Y=data.Train Targets
Nvar=size(p0)
Max it=50
N=25 // no of pop
Nc=2*round(pc*N/2)
Nm =round(pm*N)
=0.7
//Initialize population
Pop.position=[]
Pop.cost=[]
For i [1-N]
If it>1
//Evalute
//sort Population
Pop=pop[sort order]
//Best cost
Best pop=pop[1]
Worst pop=pop[end]
Main Loop
* best cos t
p exp( )
worst cos t
p
P N
p
n 1
//Mutate
Pop=[pop,popc,popm)
Parametrs :(,range)
:random(-,1+,N)
y1 * p1 (1 )* p 2
y 2 * p 2 (1 )* p1
Y1=min[max(y1,range)]
Y2=min[max(y2,range)]
Parameters : , range
Rmin=min(range)
Rmax=max(range)
dr=Rmax-Rmin
: *dr
y=p1+*random(Nm)
y=min(max(y,range))
nvar =size(p0,1)
range(min, max)
max it =50
For i :[1-N]
Pop[i].position=random[Range,N]
Pop[i].velocity : zeros[N]
Pop : domination(pop,N)
For i:[1-N]
For j:[i+1-N]
True(pop[j] Is dominated)
Else if dominates(pop[j],pop[i])
True(pop[i] Is dominated)
b=dominates(pop[i],pop[j])
b= all(x<y)&&any(x<y)
For i : [1-N]
//Select Leader
leader=Select leader(rep, )
New pop=mutatute(pop,pm,Range)
Determine domination(pop)
If dominates(New.pop.position, pop.position)
Pop[i].Position = max(particle[i].Position,Rmin);
pop[i].Position = min(particle[i].Position,Rmax);
if particle(i).Cost<particle(i).Best.Cost
pop(i).Best.Position=particle(i).Position;
pop(i).Best.Cost=particle(i).Cost;
if pop(i).Best.Cost<BestSol.Cost
BestSol=pop(i).Best
Neural network
//Load data
Hidden layers
//network type
nnetparamsInfo.hiddensize
nnetparamsInfo.hidden layers
nnetparamsInfo.nntype
nnetparamsInfo.Train func
net.divide params.testratio=0.15
//performance function
Net.perform fucn=MSE
//test network
Y=net(x)
E=[t]-[y]
EANNHRC(x)
//Input layer 1
X1_step1_xoffset=[Input]
X1_step1_gain=[e]
X1_step1_ymin=-1
//Layer 1
//Layer2
//output layer
Y1_step1_ymin=-1
Y1_step1_gain=[]
Y1_step1_xoffset=[outputs]
//simulation
For i:[1-size(output)]
//Input1
Xp1=mapminmax_apply([x1,i],x1_step1_gain, x1_step1_xoffset,x1_step1_ymin)
//Layer 1
A1=tansig_apply((b1,1,q)+IW1*xp1)
//layer 2
A2=[b1,1,q]+lw1*a1
//output 2
Y(1,t,s)=mapminmax(a2,y1_step1_gain,y1_step1_xoffset,y1_step1_ymin)
Mapminmax_apply(x,settings_gain,settings_xoffset,setting_ymin)
Y=bsxfun(@minus,x,setting_x_offset)
Y=bsxfun(@times,y,setting_gain)
Y=bsx(@plus,y,setting_ymin)
Tansig_apply(x)
Mapminmax_reverse(y,setting_gain,setting_xoffset,setting_ymin)
X=bsxfun(@minus,y,setting_x_offset)
x=bsxfun(@times,x,setting_gain)
x=bsx(@plus,x,setting_ymin)