Prediction performance analysis of neural network models for an electrical discharge turning process

In many of the modern-day manufacturing industries, electrical discharge machining (EDM) now appears as an effective non-traditional material removal process for generating intricate shape geometries on various hard-to-cut work materials to meet the ever-increasing demands of higher dimensional accuracy and better surface quality. Development of an appropriate prediction model for any of the EDM processes is quite difficult due to complex material removal mechanism, and dynamic interactions between the input parameters and responses. To address the problem, this paper proposes development and deployment of five neural network models, i.e. feed forward neural network, convolutional neural network, recurrent neural network, general regression neural network and long short term memory-based recurrent neural network as effective prediction tools for an electrical discharge turning (EDT) process. The EDT is a variant of EDM process involving removal of material from cylindrical workpieces. The input parameters for the considered EDT process are magnetic field, pulse current, pulse duration and angular velocity, whereas, the responses are material removal rate and overcut. Several statistical error metrics, like R-squared (R²), adjusted R-squared (R²_adj), root mean square error and relative root mean square error are employed to compare the prediction accuracy of all the investigated neural network models. Based on a past experimental dataset, it is observed that long short term memory-based recurrent neural network provides more accurate prediction of both the responses under consideration. On the other hand, general regression neural network is noticed to be extremely robust having highly repetitive prediction performance.