Tool wear monitoring based on multi-kernel Gaussian process regression and Stacked Multilayer Denoising AutoEncoders

Abstract

To improve product quality and realize accurate monitoring of cutting tool wear in the cutting process, the monitoring of cutting tool wear is presented based on the Multi-kernel Gaussian Process Regression (GPR) and the Stacked Multilayer Denoising AutoEncoders techniques. In this paper, a Multi-kernel GPR is constructed for the drawback of the original GPR in difficulty capturing the structure of non-homogeneous data from multiple data sources or different data types. The weight coefficients of the kernel function of the Multi-kernel GPR are set to 1 to eliminate the impact of the weight coefficients on the monitoring model and optimize its hyper-parameter using the Adaptive Moment Estimation algorithm (Adam). However, compressing the confidence interval (CI) of GPR to improve the GPR's predicted accuracy leads to incomplete coverage of the test points. As a novel feature representation learning approach, Stacked Multilayer Denoising AutoEncoders are trained locally to denoise corrupted versions of their inputs to yield significantly lower predicted error. The fused features learned by Stacking Multilayer Denoising AutoEncoders are able to zero in on more useful structure in force signal feature data, thereby improving predicted accuracy and maintaining CI stability. Twelve sets of cutting tests confirmed the reliability of the proposed tool wear monitoring technique. The experimental results show that Stacked Multilayer Denoising AutoEncoders can reduce the mean square error (MSE) of Multi-kernel GPR by>40% while maintaining CI stability. This research establishes the groundwork for monitoring tool wear in real industrial settings.