A deep capsule neural network with stochastic delta rule for bearing fault diagnosis on raw vibration signals

摘要

In recent years, deep learning techniques are explored unceasingly for machinery fault diagnosis. The vibration signal of faulty rotating machines contains distinct periodic impacts, and hence is the ideal candidate for the model input. However, there are still three challenges in deep learning on raw vibration signals: (1) The shifts of the fault impacts among the input samples prone to cause misdiagnosis; (2) The working load is always changing; (3) The background noise such as the vibration from non-goal machines is inevitable. Therefore, a novel method called deep capsule network with stochastic delta rule (DCN-SDR) is proposed for rolling bearing fault diagnosis. DCN-SDR takes raw temporal signal as input and achieves very high accuracy under different working loads. Moreover, the model performs outstandingly under noisy environment via a regularization method based on SDR. The network visualization is demonstrated and analyzed. Comparing with the state-of-the-art methods, superiority of the proposed method is verified. (C) 2019 Elsevier Ltd. All rights reserved.