Non-intrusive anomaly detection for carving machine systems based on CAE-GMHMM under multiple working conditions

This paper is concerned with a non-intrusive anomaly detection method for carving machine systems with variant working conditions, and a novel unsupervised detection framework that integrates convolutional autoencoder (CAE) and Gaussian mixture hidden Markov model (GMHMM) is proposed. Firstly, the built-in sensor information under normal conditions is recorded, and a 1D convolutional autoencoder is employed to compress high-dimensional time series, thereby transforming the anomaly detection problem in high-dimensional space into a density estimation problem in a latent low-dimensional space. Then, two separate estimation networks are utilized to predict the mixture memberships and state transition probabilities for each sample, enabling GMHMM to handle low-dimensional representations and multi-condition information. Furthermore, a cost function comprising CAE reconstruction and GMHMM probability assessment is constructed for the low-dimensional representation generation and subsequent density estimation in an end-to-end fashion, and the joint optimization effectively enhances the anomaly detection performance. Finally, experiments are carried out on a self-developed multi-axis carving machine platform to validate the effectiveness and superiority of the proposed method.