Autoencoders for learning latent information of cutting tool health

This article presents an unsupervised method for monitoring the health of cutting tools in precision machining processes. The method is developed and validated using the IEEE PHM 2010 data, which comprises three sets of cutting tool run-to-failure tests. Despite the uniformity in machining operations and settings used across the three sets of data, significant variations in tool wear levels were observed. These differences prompted the development of a method capable of generalization across different tools. Inspired by the physics of the machining process, the method computes health indicators of tool condition from the measurements of cutting forces, vibration, and acoustic emission. The health indicators derived from the cutting forces are found to have a strong association with tool wear values across tools, with high mutual information. Subsequently, an optimized Autoencoder network is employed to learn an efficient representation of the information contained in these health indicators. This encoded information is then explored in a Gaussian Mixture Model (GMM) to categorize the tool condition. The industrial implication here is the ability of the Autoencoder network to codify latent knowledge of tool health into explicit information that embodies practical shopfloor scenarios. Results demonstrate that the Autoencoder learned meaningful representations of tool health indicators that exhibit a common pattern across tools, regardless of variations in tool wear profiles, sensor measurements, and the concomitant health indicators. The study reveals that tool health can be effectively inferred from a mixture of Gaussian distributions. The findings highlight the significance of indicators and emphasize the importance of unsupervised algorithms, such as Autoencoders and GMM, in assessing the tool condition.