Vibration Monitoring and Semisupervised Multivariate Invertible Deep Probabilistic Learning for Gearbox Faults Identification

The identification of faults for gearbox plays a significant role in maintaining high reliability of industrial equipment. Multivariate invertible deep probabilistic learning (MIDPL), as a generative model, assesses the distribution characteristics of the hidden condition using observations. It is a high-dimensional-to-high-dimensional transformation and is limited by poor classification performance. In this article, semisupervised MIDPL (S-MIDPL) fault identification method is proposed to overcome the drawback of MIDPL. First, the dimension of the observation is reduced. Second, supervised learning is performed based on the reduced dimension to improve the classification performance. Fault identification validations on a gearbox test bench and wind turbine gearbox are conducted to investigate the effectiveness of the proposed method. Comparisons with other state-of-the-art classification methods have also been conducted. The results indicate that S-MIDPL has a significant advantage over MIDPL and a variational inference-based semisupervised approach.