Actual bearing compound fault diagnosis based on active learning and decoupling attentional residual network

Existing deep learning methods commonly requires massive labeled data for compound fault diagnosis, which is difficult and time-consuming to collect in the real application. This paper presents a novel decoupling attentional residual network for compound fault diagnosis. Original signal is processed by the short-time Fourier transform and its output is fed into the subsequent network. Then, attention modules are introduced into the model to selectively emphasize certain features. Additionally, a multi-label decoupling classifier is designed to accurately decouple and identify the compound faults. Besides, active learning approach is introduced to achieve the same results using few compound faults samples. The proposed method was validated on our bearing dataset, which shows that it can reach 100% overall accuracy. Moreover, the proposed method achieves the same diagnosis performance by utilizing only 150 labeled compound fault samples as using all 1200 labeled samples, which greatly reduces the labeling workload of domain experts.

Automated summary

This paper introduces a novel deep learning method for compound fault diagnosis, achieving high accuracy through a decoupling attentional residual network, multi-label decoupling classifier, and active learning approach. The method can reach the same accuracy with a small number of compound fault samples as using a large number of samples, reducing the labeling workload for domain experts.