Canonical polyadic decomposition and deep learning for machine fault detection

Abstract

Acoustic monitoring for machine fault detection is a recentand expanding research path that has already provided promisingresults for industries. However, it is impossible to collectenough data to learn all types of faults from a machine. Thus,new algorithms, trained using data from healthy conditionsonly, were developed to perform unsupervised anomaly detection.A key issue in the development of these algorithms isthe noise in the signals, as it impacts the anomaly detectionperformance. In this work, we propose a powerful data-drivenand quasi non-parametric denoising strategy for spectral databased on a tensor decomposition: the Non-negative CanonicalPolyadic (CP) decomposition. This method is particularlyadapted for machine emitting stationary sound. We demonstratein a case study, the Malfunctioning Industrial MachineInvestigation and Inspection (MIMII) baseline, how the use ofour denoising strategy leads to a sensible improvement of theunsupervised anomaly detection. Such approaches are capableto make sound-based monitoring of industrial processes morereliable.