![](https://csdnimg.cn/release/download_crawler_static/22381612/bg1.jpg)
1
Proceedings of the ASME 2019
International Design Engineering Technical Conferences
and Computers and Information in Engineering Conference
IDETC/CIE2019
August 18-21, 2019, Anaheim, CA, USA
IDETC2019-97510
CONSTRUCTION METHOD OF GENERALIZED REGRESSION NEURAL NETWORK
SURROGATE MODEL BASED ON KRIGING FITTING DEVIATION
Heju Tian, Haiyan Li
1
, Yunbao Huang
Provincial Key Laboratory of Computer Integrated
Manufacturing, Guangdong University of Technology
Guangzhou, Guangdong, China
Jingliang Lin
Provincial Key Laboratory of Computer Integrated
Manufacturing, Guangdong University of Technology
Guangzhou, Guangdong, China
ABSTRACT
In the process of simulation optimization of complex
electromechanical products, the complexity model is usually
reduced by constructing a surrogate model. The generalized
regression neural network (GRNN) surrogate model is one of the
most frequently used methods in the process of establishing a
surrogate model. GRNN has strong nonlinear mapping ability
and flexible network structure, as well as a high degree of fault
tolerance and robustness, so it is suitable for solving nonlinear
problems. but the fluctuation deviation in the construction
process seriously affects the accuracy of the surrogate model.
Aiming at the fluctuation deviation generated in the construction
of GRNN neural network, a method of constructing Generalized
regression neural network surrogate model based on Kriging
fitting deviation is proposed. The method first calculates the
fluctuation deviation value of the Generalized regression neural
network surrogate model, and the fluctuation deviation value is
approximated into a multivariate normal distribution form. Then
the multivariate normal distribution deviation is fitted in
Kriging, Finally the corresponding value of the multivariate
normal distribution deviation is subtracted from the Generalized
regression neural network surrogate model. Experimental
results show that this method can effectively reduce the
fluctuation deviation and improve the accuracy of the
Generalized regression neural network surrogate model.
Keywords: Generalized regression neural network,
fluctuation deviation, surrogate model, Kriging interpolation
1. INTRODUCTION
For complex electromechanical products, it is necessary to
improve the design efficiency continuously. Meanwhile it should
ensure the quality of the products. Product quality requires the
optimization of product performance constantly. The process of
product performance optimization is an expensive simulation
1
Contact author: cathylhy@gdut.edu.cn
optimization process involving multiple disciplines [1].
Although the computing performance of computers had been
improved greatly, the complexity of analytical calculations in
engineering is increasing constantly.
The surrogate model replaces the computationally intensive
complex simulation model with some mathematically simple
analytical models. The surrogate model can reduce the
computational complexity greatly and shorten the product
development cycle. The surrogate model needs some
experimental data at the first in the project, then uses the
experimental data and selects the appropriate surrogate model to
construct the interrelationships contained in the project, and then
uses the surrogate model to predict in the project.
There are many surrogate models on the current. These
surrogate models are being refined constantly. The most popular
surrogate models are polynomial response surface method,
Kriging interpolation [2], Gradient enhanced Kriging (GEK) [3],
Radial basis function method (RBF) [4],[5], Support Vector
Regression Method (SVR), Artificial Neural Network (ANN)
[6], and Generalized Regression Neural Network (GRNN).
Generalized Regression Neural Network (GRNN) was put
forward by American scholar Specht in 1991 [7], which is a kind
of radial basis neural network. GRNN has the strong nonlinear
mapping ability and flexible network structure, also it has the
high fault tolerance and robustness, it is suitable for solving the
nonlinear problems. GRNN has a stronger advantage than RBF
network in approximating ability and learning speed. Finally, the
network converges on the optimal regression surface with more
sample size accumulation. Meanwhile, the prediction effect is
better when the sample data is less [8]. In addition, the network
can handle the unstable data. Therefore, GRNN has been used in
various fields widely such as signal processing, structural
analysis, education, energy, food, medicine, finance, and
biology. Related to this article, Nobuo NAMURA, Koji