Online Dynamic Fault Diagnosis for Rotor System Based on Degradation Modeling and Knowledge-Enhanced Graph Convolutional Network
Abstract
The coupling analysis of model performance degradation and fault diagnosis under variable working states, known as dynamic fault diagnosis, is widely recognized as urgent need. The inadequate exploration of missing labeled samples and the intricate connection between degradation signals and fault signals has resulted in limited research on these issues. Therefore, it is necessary to design an effective method that can be used for degradation state identification and cross-domain dynamic fault diagnosis. This article proposes a novel strategy based on the optimized linear parameter-varying model and knowledge-enhanced graph convolutional adversarial network (KEGCAN). First, a degradation simulation model of the rotor system is established by analyzing the changes in contact stiffness and support stiffness caused by bolt loosening and bearing wear, respectively. Then, the proposed KEGCAN is developed based on prior knowledge from the time-frequency features of simulation data, taking advantage of the powerful capabilities of simulation model and graph convolutional network. Finally, high-precision cross-domain online dynamic fault diagnosis is achieved using both experimental data collected from the rotor system test rig and simulation data collected from the degradation simulation model. The results indicate that the proposed strategy has better diagnosis accuracy than the state-of-the-art algorithms.
