Iterative Error Self-Correction for Robust Fault Diagnosis of Mechanical Equipment With Noisy Label

Abstract

Deep neural network (DNN) is an effective technology for machinery fault diagnosis. The good performance of DNN is based on the assumption that all labels are completely correct. However, mislabeled data is common in actual industrial applications, which will cause severe performance degradation. This paper explores the performance of DNN under noisy labels and the reasons for its performance degradation. Furthermore, a novel iterative error self-correction (IESC) algorithm based on the maximum-activation of softmax is proposed. During the training process, the label is dynamically optimized, and it is no longer fixed. IESC automatically models the distribution of correct labels, gradually identifies incorrect labels, and automatically corrects incorrect labels. In addition, a noise-tolerant loss is introduced to enhance the network’s noise robustness. Experiments on two real machinery fault diagnosis cases prove that our method has excellent label correction and fault diagnosis performance. It significantly improves the performance of DNNs and promotes its practical application potential.