Cutting anomaly detection in end-milling by multimodal variational autoencoder

Abstract

Anomaly detection for predictive maintenance in the cutting process is one of the challenging problems in shop-floor management. A modern machine learning approach, including deep learning, has been widely studied for the last decade. This study focuses on the multimodality of various cutting time-series data for extracting features of cutting status and proposes a multimodal variational autoencoder (MVAE) method. We collect a time series of vibration acceleration of a cutting tool and a main spindle motor load. Various cutting data is collected by conducting cutting experiments under diverse cutting conditions. Normal and abnormal data are collected, and only normal data is used to train MVAE. MVAE learns a so-called generative model, which is implicit but stochastic, capable of reproducing original time series data. Because MVAE is an unsupervised learning method, it does not require abnormal data during training. Therefore, it is considered suitable for tools management where it is difficult to collect abnormal data. Euclidean distance is employed to evaluate the normality of a given cutting status on the latent space acquired by MVAE. We demonstrate the applicability of the proposed MVAE method in anomaly detection for end-milling by comparing it with conventional machine learning methods such as autoencoder.

Keywords

• end-milling
• anomaly detection
• multimodal time series
• unsupervised learning
• variational autoencoder
• generative model