A Stable Lifting Convolutional Autoencoder for Anomaly Detection of Turbopump Bearings of Liquid Rocket Engine

Abstract

Anomaly detection, which could not only identify potential risks early but also offer the first time for remaining useful lift prediction, plays a crucial role in assuring the safe operation of major equipment, including liquid rocket engines. However, due to the complicated modulation phenomena caused by speed variations, existing anomaly detection techniques for vibration signals with stationary speeds would fail on varying-speed signals. Simultaneously, the turbopump bearings run under transient rotating speeds. Thus, motivated by the outstanding performance of redundant second generation wavelet transform in non-stationary feature extraction, a stable lifting convolutional autoencoder (LiftingCAE) is presented. First, a lifting decomposition-based encoder is introduced to layer-by-layer decompose the components of various scales. Then, stable loss is suggested to extract latent features by minimizing the interference information whereas maximizing the health state-dependent features. Finally, the decoder based on lifting reconstruction is utilized to model health data through fusing the features of different scales. The proposed LiftingCAE was validated by vibration signals collected on a turbopump bearing test rig working in a cryogenic environment, and was compared to some state-of-the-art methods. The results show the effectiveness and superiority of LiftingCAE in detecting turbopump bearing anomalies.