Gaussian mixture model for tool condition monitoring

Abstract

This article presents an unsupervised method for monitoring tool condition in precision machining processes. The method utilizes cutting force measurements to infer the tool condition. It computes the tool health indicators by analyzing force signals in the Fourier space. The method calculates signal energy, magnitude, and variance of regions surrounding the tooth passing frequency (TPF) of the force spectra as health indicators of tool wear. Then, a Gaussian Mixture Model (GMM), as an unsupervised machine learning (ML) algorithm, utilizes these indicators to estimate the tool condition. The method is validated using two sets of tool run-to-failure tests conducted on different machines with varying capabilities to develop a generic approach that is applicable across machines and cutting parameter settings. The developed approach is also validated using the IEEE PHM 2010 data. Results show that the derived indicators are highly informative of the tool condition, with a high Pearson correlation coefficient of 0.94 with tool wear measurements, regardless of differences in machine or cutting parameter settings. Results also reveal that tool conditions can be clustered into a mixture of Gaussian distributions by the same health indicators. Classification accuracy of 0.965 is achieved on the in-house machining data and 0.96 with the IEEE PHM 2010 data. Results also demonstrate that the GMM effectively predicts the evolution of tool life with the longest duration in the initial usage period, decreases during the uniform wear stage, and reaches a minimum during the accelerating wear stage towards the end of life. These tool life stage durations are measured with respect to the number of machining test runs. The study emphasizes the significance of indicators derived from the physics of the process and highlights the importance of an unsupervised and explainable monitoring system for assessing the tool condition.