Scribd
Upload
12 views

Generalized Approach For Multi-Response Machining Process

This document discusses using machine learning models like Gaussian kernel regression (GKR), feed forward neural networks (FFNN), and generalized regression neural networks (GRNN) as surrogate models to optimize manufacturing processes. Surrogate models are used to approximate computationally expensive simulations and experiments when optimizing manufacturing process objective functions with multi-objective evolutionary algorithms. The document evaluates these surrogate models on optimization problems related to machining processes using metrics like mean squared error to determine the best performing model.

Uploaded by

Divyansh Verma
Copyright
© © All Rights Reserved
Available Formats
Download as TXT, PDF, TXT or read online on Scribd
12 views

Generalized Approach For Multi-Response Machining Process

This document discusses using machine learning models like Gaussian kernel regression (GKR), feed forward neural networks (FFNN), and generalized regression neural networks (GRNN) as surrogate models to optimize manufacturing processes. Surrogate models are used to approximate computationally expensive simulations and experiments when optimizing manufacturing process objective functions with multi-objective evolutionary algorithms. The document evaluates these surrogate models on optimization problems related to machining processes using metrics like mean squared error to determine the best performing model.

Uploaded by

Divyansh Verma
Copyright
© © All Rights Reserved
Available Formats
Download as TXT, PDF, TXT or read online on Scribd
You are on page 1/ 3

Page 1

Multi-Objective Evolutionary Algorithm (MOEA)


have been an evolving topic of research for solving real-world com
binatorial problems
Manufacturing Process Optimization Problem (MPOP)
When MOEAs are employed to solve the MPOPs, data-driven
surrogate models are generally used as the objective functions,
which potentially eradicate the need of performing computation
ally complex mathematical modeling, Finite Element Method
(FEM) approach, or expensive empirical experiments
Surrogate modeling could also be classified as the
black-box modeling when there would be little or no information
available about the machining process under consideration
Page 2
These are, Response Surface Methods
(RSM), Gaussian process (GP), Redial Basis Functions (RBF), Support
Vector Machines (SVM), Multi-Layer Perceptron (MLP), Gaussian
Kernel Regression (GKR), Artificial Neural Network (ANN), General
ized Regression Neural Network (GRNN) etc.
Page 3
The most crucial vari
ables are determined as the cutting speed, feed rate, axial depth,
and radial depth.
Page 4
feed forward neu
ral network (FFNN), GKR, and GRNN.
FFNN is an apt tool for surrogate modeling because it is natu
rally proficient in approximating outputs from the arbitrary input
parameters
GRNN
GRNN estimates the responses using weighted average of the
outputs of the training data. The weights are computed using the
euclidean distance between the training and testing data, which
are also termed as activation function
The response equation of GRNN is,
y ¼Piyie#d2id2

# # 2r2
Pie#d2id2

# # 2r2ð3Þ
where d22

i ¼ ðÞ x # xi Tðx # xiÞ is the euclidean distance between x and


xi, e#d2d2

ð Þi22r

ð Þd2

ð Þi22r

# # ð Þ is the activation function and r is the spread constant,


which could be adjusted to hold the optimal value using GRNN
training so that the mean square error (MSE) value converges to
zero
Most of the research works are based on the empirical real
world data; however, the systematic MPOP test data or pub
lished data are not yet available in cumulative forms for the
algorithmic testing.
For that matter, three different machine-learning
models based on MLP/Feed Forward ANN, GKR, and GRNN are
employed and the best performing surrogate is picked.
GKR
GKR is based on the data mapping from the low dimensional
space to high dimensional space.
Page 5
The linear
regression learner used, is based on the SVM regression. In this
approach, the input parameters x are mapped onto an m
dimensional attribute
The abovementioned regression model can be transformed into
an optimization problem using,
min Z ¼ 12 kx2k þ C Xn
k¼1ðci þ c# i Þ
f x ð Þ ¼ ; x Xm
j¼1xjgjð Þ þ x b
Rð Þ ¼ x 1n Xn
i¼1Lðyi; dðxi; xÞÞ
Page 6
MSE is used as the performance measure for the
abovementioned surrogate models.
This procedure could group
the members in new population POP using different ranking sys
tem (Fi where i=1, 2,. . ., n). The population for the next iteration is then
obtained based on this ranking system.
Page 7
The reference points are obtained using the procedure defined in
ref. [53], which arranges the reference points on the normalized
hyper-plane.
The normalization procedure proposed by Deb and Jain [44]
helps identify the set of ideal points Zmax by solving a set of linear
equations, which is computationally expensive.
Niches are matched for each reference point based on the asso
ciated population member. Niche preservation is performed to
select the desired candidates from Fl
If the niche count is zero, the member is
chosen based on the smallest perpendicular distance to the refer
ence line else if the niche count is one or greater, a random mem
ber is selected from Fl front
Page 8
These are based
on the turning, grinding, heat exchanger tube, milling, abrasive
water jet machining, dry turning, drilling, welding, and emulsifica
tion processes.
For the purpose of comparison,
another well-versed many-objective algorithm, MOEA/D [54] is
considered in this study. The difficulty of the data-driven
surrogate-assisted optimization algorithms could amplify insanely
with the number of process parameters and performance charac
teristics for the MPOPs.
Page 9
Population
size, number of generations and initial population generation tech
niques differ depending on the size of the data and number of the
objectives.
Page 11
The objective of this statistical test is to measure the perfor
mances of the EAs in the state space. For that matter, the euclidean
distance between multi-objective solutions (BFS and EA solutions)
are computed using,
dist z ðÞ ¼ 1;
z2ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiff
iffiffiffiffiffiffi

qXm
i¼1ðffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffif
fiffiffiffiffiffiffi

Þ f ið Þ # z1 f ið Þ z2 2
where z1 and z2 are the solutions and fi(z1) and fi(z2) are the fitness
values for ith objective.
Page 12
These types of optimization problems are based on the lim
ited off-line data.
The proposed data-driven technique exploited the GKR
based surrogate for the objective approximation. In order to verify
the proposed methodology, a comprehensive number of data are
considered from the past literature.
Page 13
Generic objective approximation and data-driven many/multi
objective optimization approaches are rarely practiced together
for the manufacturing process optimization, which is done in
this work
The initial population generation for the data-driven EA main
tains the variation in the population. For that matter, LHS tech
nique is mostly exploited

You might also like