CiiT International Journal of Artificial Intelligent Systems and Machine Learning, Vol 5, No 7, July 2013
Implementation of Radial Basis Function Neural
Network for Estimation of Strain of Blade
R.I. Rajidap Neshtar and S. Purushothaman
Abstract---This paper presents estimation of stress and strain of a
Rapid prototype product using artificial neural network (ANN).
Radial basis function network is used to train the ANN topology. 3D
model of blade is developed by using PROE. The model is analyzed
using ANSYS to find the Von Mises stress and equivalent strain. The
algorithm is trained using 15 values in the input layer of the ANN
topology and two values in the output layer: stress and strain that are
to be estimated during the testing stage of RBF algorithm.
Keywords---Radial Basis Function Network, Finite Element
Method, Structural Analysis, and Blade.
I. INTRODUCTION
R APID prototyping plays an important role in
manufacturing sample products for quick approval of the
customers. Colors can be added to the products for aesthetic
appearance to impress the customers. Different functionalities
can be provided in the RP product and put into use to verify
the usability. RP products are manufactured with layered
Manufacturing. It is the process of deposition of material layer
by layer. Continuous solid wire is melted, deposited on a non-
sticky platen. The melt solidifies. During this process, the
platen is moved in x, y directions as per a sequence and a
sheet of solidified layered material is formed with a certain
thickness. This is called first layer of deposition and the platen
is lowered to receive the next layer of deposition. The process
is repeated until, the entire height of the RP product is
achieved. The material is not deposited during layering
whenever holes, curves, slopes are to be provided in the
product.
The sequence of RP manufacturing is presented as follows:
Step 1: Created computer aided drafting and design of the RP
product.
Step 2: Analyse the model using Finite element software like
ANSYS, NASTRAN, if the RP product is the end
product. Otherwise, if the product is only for a display or
customer approval process, no analysis is required for
small sized RP product.
Step 3: Transfer the final RP model to RP machine. The software
in the RP machine will convert the model suitably for
layered manufacturing.
Manuscript received on July 05, 2013, review completed on July 12, 2013
and revised on July 18, 2013.
R.I.Rajidap Neshtar, Research Scholar, Department of Mechanical
Engineering, Vinayaka Missions University, Salem, India. E-Mail:
rajidap@gmail.com
Dr.S.Purushothaman, Professor, Department of Mechanical Engineering,
PET Engineering College, Tirunelveli District-627117, India. E-Mail:
drsppuru@gmail.com
Digital Object Identifier: AIML072013005.
0974-9667/CIIT–IJ-4551/05/$20/$100 © 2013 CiiT
315
In step 2, an alternative procedure is proposed. After
analyzing the RP product with FEM software, the outputs of
the FEM software along with the inputs of FEM are used as
training patterns for the proposed artificial neural network
algorithms. The algorithms learn the input output relationship
of the values used an input and obtained as outputs for
analyzing the RP product. This type of analysis can be
extended to all types of RP products as well as other
engineering products.
1 Companies working on unique products and companies
which modify the size of a particular RP product can avail the
proposed.
2. Such companies need not obtain costly FEM software;
instead they can use the proposed approach to save the
investment of the FEM software for obtaining the stress, strain
values of the RP product.
II. REVIEW OF LITERATURE
Wenbin et al, 2002, applied optimization algorithms like
genetic algorithm for the outputs of FEM. Attention is devoted
to examining the effects of critical geometric features on the
stress distribution at the interface between the blade and disk
using a feature-based geometry modeling tool and the
optimization techniques. Various aspects of this problem are
presented: (1) geometry representation using ICAD and
transfer of the geometry to a finite element analysis code, (2)
application of boundary conditions/loads and retrieval of
analysis results, (3) exploration of various optimization
methods and strategies including gradient-based and modern
stochastic methods. A product model from Rolls-Royce is
used as a base design in the optimization.
Manevitz et al, 2005, Basic learning algorithms and the
neural network model are applied to the problem of mesh
adaptation for the finite-element method for solving time-
dependent partial differential equations. Time series prediction
via the neural network methodology is used to predict the
areas of ‘‘interest’’ in order to obtain an effective mesh
refinement at the appropriate times. This allows for increased
numerical accuracy with the same computational resources as
compared with more ‘‘traditional’’ methods.
REN Xiao-qiang et al, 2006, studied the nonlinear problems
of mechanics using ANN. They simulated two simple
elastoplastic problems. Lagrange Polynomial neural network
optimization results for elastoplastic problem are found to be
comparable to traditional Hopfield neural network
optimization model.
Published by the Coimbatore Institute of Information Technology
CiiT International Journal of Artificial Intelligent Systems and Machine Learning, Vol 5, No 7, July 2013
Hacib et al 2007, used radial basis function neural network
and finite element method (FEM) to solve the inverse problem
of parameter identification. The methodology use a large
number of variations of magnetic relative permeability and
electric conductivity in a material under test by FEM. The
obtained results are used to generate a set of vectors for the
training of a RBF neural network. The obtained neural
network is used to identify electromagnetic parameters of a
group of new materials that not belonging to the original
dataset. Performance of the RBF neural network is compared
with the most commonly used multilayer perceptron network
model.
Scott et al, 2007, studied the accuracy of a finite element
model for design and analysis of a metal forging operation.
They found that, the method is limited by material model’s
ability to predict deformation behavior over a wide range of
operating conditions. Current rheological models prove
deficient in several respects due to the difficulty in
establishing complicated relations between many parameters.
ANN has been suggested as an effective means to overcome
these difficulties. A robust ANN with the ability to determine
flow stresses based on strain, strain rate, and temperature is
developed and linked with finite element code.
Toraño et al, 2008, used neural networks to estimate stress –
strain of long mining wall. This knowledge and the detailed
structural and constructive characteristics of the support
systems allow the simulation of the behavior of the roof
supports through finite element method.
Mohsen Ostad Shabani et al, 2011, used ANN and FEM for
estimating the yield stress, Ultimate Tensile Strength,
maximum force and elongation percentage of solidification in
A356 alloy. The prediction of ANN model with one output
was found to be in good agreement with experimental data.
The results show that the prediction of neural network
modelling with four outputs cannot really give a good
performance and show the best relationship between each
individual output and its inputs.
Mohsen Ostad Shabani et al, 2012, developed ANN model
to predict the hardness, yield stress, ultimate tensile strength
and elongation percentage. The prediction of ANN model was
found to be in good agreement with experimental data. It is
concluded that considerable savings in terms of cost and time
could be obtained by using neural network model.
Hyuntae et al, 2013, developed noncommercial ANN
simulator with graphical user interface (GUI) to assist in rapid
data modeling and analysis in the engineering diffraction field.
This software has been used to train and simulate the finite
element modeling (FEM) data for a fiber composite system,
both forward and inverse. The forward neural network
simulation precisely reduplicates FEM results several orders
of magnitude faster than the slow original FEM. The simulator
GUI also reveals that output node size for materials parameter
and input normalization method for strain data are critical train
conditions in inverse network.
0974-9667/CIIT–IJ-4551/05/$20/$100 © 2013 CiiT
316
III. METHODOLOGY
ANSYS 14 software is used for analyzing the rapid
prototypes. A blade is considered for analyzing the estimation
accuracy of RBF. The analysis conditions of the blade is
presented. Load deformation response characteristics of the
blade are determined using a finite element computer analysis.
Results of this analysis is used as training patterns for the
ANN algorithm.
Artificial neural network algorithm is used to supplement
the estimation of stress and strain values of the RP models.
The result of the analysis of blade is obtained both in graphical
and in numerical values. The numerical values of stress and
strain are used to train the artificial neural network (ANN)
topology by using RBF.
The ANN is trained using 15 values in the input layer two
values in the output layer: stress and strain that are to be
estimated during the testing stage of ANN algorithms. The
number of nodes in the hidden layer varies depending upon
the weight updating equations. Exact number of nodes, is
fixed based on the trial and error method, in which the
accuracy of estimation by the RBF is used as the criteria for
the performance of ANN algorithm.
The training of patterns used for the ANN are chosen from
the stress and strain data generated using ANSYS software.
During the training process, mesh node numbers are presented
in the input layer of the ANN and correspondingly, stress and
strain values are presented in the output layer of the ANN.
Depending upon the type of values present in the patterns, the
learning capability of the ANN algorithms varies.
A. Radial basis Function Neural Network (RBF)
Radial Basis Functions (RBFs), [John Moody and Christian
J Darken, 1989, Craddock R.J., and Warwick K., 1996],
assures approximation to a function ‘f’ given input data. Input
point is presented through a set of basis functions and the
approximation is produced with RBF centers. Then the result
is multiplied with a coefficient to sum them linearly. The
approximation is stored in the coefficients given function‘t’ to
obtain centers of the RBF. RBFs have the following
mathematical representation:
Where,
c is a vector containing the coefficients of the RBF,
CR is a vector containing the center of the RBF, and
a is the basis function or activation function of the
network.
Training RBF with strain values
Step 1: Apply Radial Basis Function.
No. of Input = 15
No. of Patterns = 500
No. of Centres = 10
Calculate RBF as
RBF = exp (-X)
Calculate Matrix as
G = RBF
A = GT * G
Published by the Coimbatore Institute of Information Technology
CiiT International Journal of Artificial Intelligent Systems and Machine Learning, Vol 5, No 7, July 2013 317

Calculate mentioned in the x-axis. Figure 5 presents the strain patterns

B = A-1
Calculate
E = B * GT
Step 2: Calculate the Final Weight. F=E*D
Step 3: Store the Final Weights in a File.
The final updated weights are saved for estimating the stress
and strain.
IV. RESULTS AND DISCUSSIONS
Static analysis calculates the effects of steady loading
conditions on a structure. Static analysis, however, includes
steady inertia loads (such as gravity and rotational velocity),
and time-varying loads that can be approximated as static
equivalent loads. Static analysis is used to determine the
displacements, stresses, strains, and forces in structures or
components caused by loads that do not induce significant
inertia and damping effects. Steady loading and response
conditions are assumed; that is, the loads and the structure’s
response are assumed to vary slowly with respect to time.
used for training RBF.
The node number is given as one of the inputs at the input
layer of the ANN topology. Based on the requirements of the
outputs, number of input parameters can be increased. Table 1
presents the parameter used in the input and output layer of
RBF. Table 2 presents the numbers of nodes used in the input
layer, hidden layer and output layer during training.
ANSYS 14 software is used for analyzing the blade. The
numerical values of strain are used to train the artificial neural
network (ANN) topology by using RBF. The training of
patterns used for the ANN algorithms are chosen from the
strain data generated using ANSYS program. During the
training process, node numbers are presented in the input layer
of the ANN and correspondingly, strain values are presented
in the output layer of the ANN. Depending upon the type of
values present in the patterns, the learning capability of the
ANN algorithms varies.
Figure 6 present performance of RBF in estimating the
strain for different number of centers.

Bounding Box
Length X 7.0667e-002 m
Length Y 1.8172e-002 m
Length Z 7.0667e-002 m
Properties
Volume 5.311e-006 m³
Mass 4.1691e-002 kg
Centroid X 7.9537e-018 m
Centroid Y -7.5e-004 m
-1.3338e-018
Centroid Z
m
Moment of Inertia 1.1968e-005
Ip1 kg·m²
Moment of Inertia 2.3542e-005
Ip2 kg·m²
Moment of Inertia 1.1968e-005
Ip3 kg·m²
Statistics Fig 1 Blade
Nodes 3556
Elements 1603
Mesh Metric None

Structural Steel > Constants

Density 7850 kg m^-3
Coefficient of Thermal
1.2e-005 C^-1
Expansion
434 J kg^-1
Specific Heat
C^-1
60.5 W m^-1
Thermal Conductivity
C^-1
1.7e-007 ohm
Resistivity
m

Figure 1 presents a blade under analysis. Figure 2 presents

the distribution of Von-Mises stress on the blade. Figure 3
presents the distribution of strain on the blade. Figure 4 shows Fig 2 Equivalent Stress
the amount of strain presented at various nodes that are

0974-9667/CIIT–IJ-4551/05/$20/$100 © 2013 CiiT Published by the Coimbatore Institute of Information Technology

CiiT International Journal of Artificial Intelligent Systems and Machine Learning, Vol 5, No 7, July 2013 318

Fig 3 Equivalent Strain

Fig 6 Strain Estimaion by RBF for Different Centers

0.015
Deformation

0.01 TABLE 1
TRAINING PARAMETERS FOR RBF
0.005 Inputs to the ANN topology Target outputs for the ANN
topology
0 1. Diameter / Length of the model 1. Equivalent von-Mises
2. Thickness of the model stress in (Pa)
1189

1981

2773
3169
397

1585

2377
793
1

3. Width of the model 2. Equivalent strain (m/m)

4. Number of holes
5. Number of ribs
Node numbers 6. Maximum to minimum
diameter / Length
7. Uniformly distributed load
Fig 4 Strain Value in the Blade (1=’yes’ , 0=’no’)
8. Uniformly distributed load
value
9. Point load (1=’yes’ , 0=’no’)
10. Point load value
11. Yield strength
12. Ultimate strength
13. Young’s modulus
14. Poisson ratio
15. Node number

TABLE 2
ANN TOPOLOGY
Nodes in Centers in Nodes in the
the input the Hidden output layer
layer layer
RBF 15 3 2
Fig 5 Strain Values used for Training RBF

V. CONCLUSION
In this paper, ANSYS 14 software is used for analyzing the
alloy RP model. The experimental simulation and the
conditions under which the experiments were simulated are
mentioned. Load deformation response characteristics of blade
model is determined using a finite element computer analysis.
Artificial neural network algorithms have been used to
supplement the estimation of strain values of the proposed
blade model. The result of the analysis of the model is
obtained both in graphical and in numerical values. The
numerical values of strain is used to train the artificial neural
network (ANN) topology by using RBF.

0974-9667/CIIT–IJ-4551/05/$20/$100 © 2013 CiiT Published by the Coimbatore Institute of Information Technology

CiiT International Journal of Artificial Intelligent Systems and Machine Learning, Vol 5, No 7, July 2013 319

The algorithm is trained using 15 values in the input layer

of the ANN topology and two values in the output layer: strain
that are to be estimated during the testing stage of RBF
algorithm. The number of centers in the hidden layer for RBF
varies depending upon the estimation accuracy.

REFERENCES
[1] Craddock R.J., and Warwick K., 1996, Multi-Layer Radial Basis
Function Networks. An Extension to the Radial Basis Function, IEEE
International Conference on Neural Networks, Vol.2, No.1, pp.700-705.
[2] Hacib T., Mekideche MR., and Ferkha N., Computational Investigation
on the Use of FEM and RBF Neural Network in the Inverse
Electromagnetic Problem of Parameter Identification, IAENG
International Journal of Computer Science, 33:2.
[3] Hyuntae Na, Seung-Yub Lee, Ersan Üstündag, Sarah L. Ross, Halil
Ceylan, and Kasthurirangan Gopalakrishnan,Development of a Neural
Network Simulator for Studying the Constitutive Behavior of Structural
Composite Materials, ISRN Materials Science, Vol.2013, Article ID
147086, 10 pages.
[4] Javier Toraño, Isidro Diego, Mario Menéndez and Malcolm Gent, 2008,
Finite element method (FEM) – Fuzzy logic (Soft Computing) – virtual
reality model approach in a coalface long wall mining simulation,
Automation in Construction, Vol.17, pp.413–424.
[5] John Moody and Christian J Darken, 1989, Fast Learning in Networks of
Locally-Tuned Processing Units, Neural Computation, Vol.1, No.2,
pp.281-294.
[6] Larry Manevitz, Akram Bitar and Dan Givoli, 2005, Neural network
time series forecasting of finite-element mesh adaptation,
Neurocomputing, Vol.63, pp.447–463.
[7] Mohsen Ostad Shabani and Ali Mazahery, 2011, The ANN application
in FEM modeling of mechanical properties of Al–Si alloy, Applied
Mathematical Modelling, Vol.35, pp.5707–5713.
[8] Mohsen Ostad Shabani, Ali Mazahery, Mohammad Reza Rahimipour
and Mansour Razavi, 2012, FEM and ANN investigation of A356
composites reinforced with B4C particulates, Journal of King Saud
University – Engineering Sciences, Vol.24, pp.107–113.
[9] REN Xiao-qiang, CHEN Wu-jun, DONG Shi-lin and WANG Feng,
2006, Neural network method for solving elastoplastic finite element
problems, Journal of Zhejiang University SCIENCE A, Vol.7, No.3,
pp.378-382.
[10] Scott Kessler B., Sherif El-Gizawy A., and Douglas E. Smith, 2007,
Incorporating Neural Network Material Models Within Finite Element
Analysis for Rheological Behavior Prediction, Journal of Pressure
Vessel Technology, Vol. 129, 58-65.
[11] Wenbin Song, Andy Keane, Janet Rees, Atul Bhaskar, and Steven
Bagnall, 2002, Turbine blade fir-tree root design optimisation using
intelligent CAD and finite element analysis, Computers and Structures,
Vol.80, pp.1853–1867.

R.I. Rajidap Neshtar completed his B.Tech in

Mechanical Engineering and M.E in CAD/CAM from
Karunya University, Coimbatore in 2008. He has 5
years of teaching experience. Presently he is working as
an Assistant Professor in Bethlahem Institute of
Engineering, India. He has attended various Faculty
development programs and Soft skill trainings and
attended various national and international conferences.

Dr.S. Purushothaman completed his PhD from Indian

Institute of Technology Madras, India in 1995. He has
124 publications to his credit. He has 19 years of
teaching experience. Presently he is working as
Professor in PET college of Engineering, India..

0974-9667/CIIT–IJ-4551/05/$20/$100 © 2013 CiiT Published by the Coimbatore Institute of Information Technology