Well-informed neural network: an approach for the prediction of the width of flank wear land in turning processes

Precise information on the current state of the tool wear is essential in machining processes in order to produce workpieces with an adequate surface quality as well as to use the full tool capacity and save valuable resources. In this work, different artificial neural networks are developed and compared to a regression model to predict the width of flank wear land by using measured cutting force as input data. In particular, a “well-informed” neural network approach is introduced. This is inspired by physics-informed neural networks, in which differential equations are taken into account, but uses empirical knowledge instead. Turning experiments with three different feeds were conducted and tool wear was measured at several process times until tool failure. Measured data for two of the feeds were used for training and data for the third feed were used for testing. As a result in the test scenario, the well-informed neural network with pre-knowledge based on Kienzle’s cutting force equation yielded the highest accuracy in tool wear prediction, outperforming both a regression approach with no artificial neural network extension and an artificial neural network with no pre-knowledge. By changing the datasets used for training and testing, the results also reveal a better extrapolation capability compared to the artificial neural network without pre-knowledge. Supplementary 1