Self-learning transferable neural network for intelligent fault diagnosis of rotating machinery with unlabeled and imbalanced data

Abstract

As a promising tool for intelligent diagnosis of rotating machinery with unlabeled data, transfer learning (TL) has attracted considerable attentions from academia and industry. However, mechanical data in real-case have obviously unlabeled and imbalanced characteristics, which are not simultaneously concerned enough by existing intelligent TL fault diagnosis methods. Specially, the highly imbalanced mechanical data lead to the skewed marginal distribution and classifier, which makes a huge challenge for the TL-based intelligent fault diagnosis method. Self-updating knowledge via tradeoff of believing, doubting and rectifying is a main learning process of human being, from which the previous experience is of great importance to enhance the learning ability. Inspired by the learning strategy of human being, a self-learning transferable neural network (STNN) is proposed for the intelligent machinery fault diagnosis with unlabeled and imbalanced data in this study. Three novel loss terms are constructed into STNN for realizing the self-believing, doubting and rectifying in the prediction of health conditions. First one is the self-believing loss term, which uses the conditional distribution adaptation to align the learned cross-domain features proactively. Second one is the self-doubting loss term that provides the ability of freeing from the false experience for STNN. Third one is the self-rectifying loss term, in which the information entropy is employed to regulate the learning process of STNN. Two experimental cases of rotating machinery show that the effectiveness and superiority of the proposed STNN in enhancing the intelligent fault diagnosis of rotating machinery with unlabeled and imbalanced data.