Open-Set Fault Diagnosis based on Prototype Learning with Dual Category-Classifier

Fault diagnosis is one of the most essential parts of industry quality inspection. Recent years have seen great success in Deep-Learning (DL) based fault diagnosis methods. However, traditional supervised DL methods only aim at reducing the empirical risk but ignore the open space risk. Thus, when applied in the open-set scenario, these methods cannot recognize the unknown faults that have not occurred yet. To tackle this issue, this paper develops a novel dual category-classifier openset framework to reduce both the empirical classification risk and the open space risk. In this framework, a reciprocal point and a prototypical point are deployed for each known category in the training phase. The reciprocal points are optimized by the extra-class space to enlarge the between-class distance, while the prototypical points are optimized to compress the within-class distance. Moreover, a margin constraint term is added for further restricting the distribution range. Finally, a novel-categories detector named Kernel Null Foley-Sammon Transform (KNFST) is adopted to reject the unknown fault modes. Computational experiments conducted on CWRU dataset, and a practical bearing dataset shows that the proposed method can successfully deal with the open-set fault diagnosis problem and achieve a remarkable improvement compared with most state-of-the-art methods.