Multiscale One-Class Classification Network for Machine Health Monitoring

Abstract

Currently, most machine health monitoring methods require labeled abnormal or faulty samples to build a supervised classification model. However, it is difficult and expensive to collect enough faulty samples for accurate and reliable fault detection and diagnosis. To address this issue, we formulate machine health monitoring as a one-class classification problem and propose a new unsupervised health monitoring model named multiscale one-class classification network (MOCN) only relying on normal data without any label information. It consists of two sequential learning modules: a multiscale feature representation learning module to extract effective features from raw one-dimensional vibration signals and a one-classifier training module to build a hypersphere model to contain all normal samples with a minimum radius. Unlike traditional two-step one-class classification methods, both modules in our proposed MOCN are jointly optimized, and therefore it can simultaneously learn compact representations and train a deep one-class classifier. Our proposed MOCN method is evaluated through experiments on three datasets, and experimental results have proved the superiority of our proposed MOCN model compared with several existing methods.