This paper addresses the development of a back-propagation neural network model for flow stress prediction based on plane strain compression (PSC) test data. Basic concepts of the neural network modelling are given, followed by discussions on training data requirements and other critical issues in neural network modelling. Original training data have been obtained via many PSC tests for a low carbon steel (C430). Data preprocessing is very important in neural network modelling, especially when the data are from industrial processes where various disturbances are very likely. A two-stage data preprocessing procedure was proposed to deal with the PSC data: data rationalising and data filtering. The quality of the training data is significantly improved after the data preprocessing. The developed BP neural network model had been implemented on a Pentium-based personal computer. Simulation results show that the average output prediction error by BP network is less than 4% of the prediction range. The training error gradually decreases with increasing hidden neurons. However, increasing hidden neurons do impose a danger of over-training, with the validation error increasing instead of decreasing. Compromising between the training error and validation error, we suggest that a BP neural network with a single hidden layer and 10-20 hidden neurons should be sufficient for flow stress modelling.
[1]
John G. Lenard,et al.
A comparative study of artificial neural networks for the prediction of constitutive behaviour of HSLA and carbon steels
,
1996
.
[2]
N. J. Silk,et al.
Interpretation of hot plane strain compression testing of aluminium specimens
,
1999
.
[3]
C. M. Sellars,et al.
Constitutive equations for high temperature flow stress of aluminium alloys
,
1997
.
[4]
Kurt Hornik,et al.
Multilayer feedforward networks are universal approximators
,
1989,
Neural Networks.
[5]
C. M. Sellars,et al.
Modelling Microstructure and Its Effects during Multipass Hot Rolling
,
1992
.