Unsupervised anomaly detection of machines operating under time-varying conditions: DCD-VAE enabled feature disentanglement of operating conditions and states

Anomaly detection (AD) plays a key role in condition monitoring (CM) to ensure the machine system's operating reliability and safety. When machinery operates under time-varying operating conditions (TVOCs), interference from varying operating conditions (OCs) exacerbates the difficulty of AD. To address this issue, a Disentangled Representation Learning(DRL) approach is proposed to dissociate the features linked with OCs and operating states (OSs). Expanding on the pre-existing Variational Autoencoder (VAE), Distribution Constraint Decomposition (DCD) is proposed as a regularization approach, which implements a loose-tight constraint depending on Kullback-Leibler(KL) divergence to enforce prior constraints on the latent features. As a result, DCD-VAE, which enables the selective allocation of different types of information, achieving disentanglement between OCs’ information and the OSs’ information, is proposed in this paper. An anomaly indicator(ANI) constructed based on the OSs features enables AD. Simulation and experiments validate the substantial advantage of the proposed approach over comparable methods, facilitating the timely and precise identification of mechanical faults.