INVESTIGATION OF BACK-PROPAGATION NETWORK STRUCTURES IN PREDICTING WEAR VALUE OF CBN-COATED TOOLS DURING HIGH-SPEED DRY TURNING OF SKD11 STEEL

Abstract

This paper presents the results of surveying the effect of backpropagation
network structures on the predictive quality of CBN-coated cutting tools in high-
speed dry turning of SKD11 steel on CNC machines. Based on the analysis of the
back-propagation network (BPN) characteristics and determining that the
number of hidden layers of this network structure is fixed, eighteen network
structures corresponding to the six ratios of the number of neurons between the
hidden layers are investigated and evaluated. The network training dataset is
collected from 280 high-speed turning experiments with four input variables and
one output variable. The network quality evaluation criteria include R2
, MSE, RMSE
and MAPE indexes. The survey results show that, in this particular case study, the
ratio of the number of neurons between hidden layer 1 and hidden layer 2 reaches
1:2, giving the best prediction quality. The 4-10-20-1 network configuration is the
model for the best quality. The research results can serve as a basis for selecting
the appropriate neural network (NN) configuration for models with large amounts
of data and many input variables. However, this study only examines the network
model with 2 hidden layers.