PET-AE: Physics-informed enhanced temporal autoencoder for incipient fault detection of shafting systems

Abstract

Incipient fault detection is crucial for improving the stable operation of shafting systems. Autoencoders (AEs) have gained popularity in the field of incipient fault detection. However, AE-based methods are weak in capturing temporal and periodic dependencies hidden in monitoring signals. This hinders the timely detection of incipient faults. To tackle these challenges, a physics-informed enhanced temporal autoencoder (PET-AE) is proposed for incipient fault detection of shafting systems. In this method, a Transformer autoencoder is constructed to reconstruct signals, where the differential Transformer encoder is used to mine temporal features from input signals. Moreover, a spectrum module is designed to capture global and local frequency information to enhance the periodic representations. Then, an enhanced memory module is employed to enlarge the distribution gap between normal samples and degradation samples. To verify the effectiveness of the proposed method, experimental studies are implemented on IMS bearing dataset and a self-built propulsive shafting system. Experimental results demonstrate that the proposed PET-AE has outstanding fault detection performance compared to other advanced detection methods.