Broad auto-encoder for machinery intelligent fault diagnosis with incremental fault samples and fault modes

Abstract

Intelligent fault diagnosis (IFD) has been a widely concerned topic in the field of prognostics and health management. Existing machinery IFD approaches are generally developed based on the one-time learning manner. Therefore, they are powerless to deal with the data stream issue in which new fault samples and fault modes will be progressively collected for model training. To overcome this drawback, this paper proposes a broad auto-encoder (BAE) with incremental learning capabilities for on-line IFD of machinery. The BAE is constructed by stacking a series of auto-encoders in the width direction. Then, the output weight matrix of the BAE is calculated by the ridge regression algorithm. After that, the capabilities of sample-incremental learning and class-incremental learning are developed, so that the BAE can easily update itself to accommodate the new fault samples and fault modes without model retraining. With the two incremental learning capabilities, the BAE can be first trained using limited historical fault samples, and then incrementally learn new diagnosis knowledge from the newly coming fault samples and fault modes. In this way, the BAE will be more and more powerful over time. Finally, the proposed BAE is applied to diagnose faults for high-speed train wheelset bearings and disc components. The results show that the proposed BAE offers an efficient solution for machinery IFD to deal with the continuous data stream issue.