Nonstationary incipient fault detection based on hybrid supervised trend-period variational autoencoder and its application in thermal power generation

Abstract

Incipient fault detection has been considered as one of the most efficient approaches to reduce the risks of systematic failures. However, incipient fault signals are often obscured by nonstationary characteristics, such as trend features and periodic features. In this paper, a hybrid-supervised trend-period variational autoencoder (HSTPVAE) is proposed to achieve fault detection for incipient faults in nonstationary processes. The features of trend, period and residual are extracted from a novel trend-period variational autoencoder (TPVAE). Then, these features are optimized by a hybrid supervised strategy, which includes fault trend semi-supervised module and trend-period self-supervised module. The former enhances the distinctiveness between normal and fault trend features, while the latter prevents the overfitting issues. Finally, the effectiveness of the HSTPVAE is demonstrated on a numerical simulation process and real boiler combustion process of thermal power generation. The comparison with state-of-the-art (SOTA) methods proves that the proposed HSTPVAE method can fully utilize the trend and period features of nonstationary process and outperform other comparison methods in incipient fault detection.