Weighted domain separation based open set fault diagnosis

Cross domain fault diagnosis based on deep learning is of great significance for improving the reliability and safety of mechanical equipment. Generally, it assumes that the label sets of training data (source domain) and test data (target domain) are consistent. However, the test data usually contain unknown classes that are unseen in the training data due to unpredictable fault modes in real industrial scenarios. Therefore, the open set fault diagnosis (OSFD) where the training label set is a part of the test label set appeared. However, most previous studies directly aligned the source domain and target domain without considering the private features of each domain and required prior knowledge to set the threshold for unknown class detection. Thus, a weighted domain separation network (WDSN) is proposed. First, the unknown samples are detected by establishing the boundary between known class and unknown class by a binary classifier without setting a threshold. Then, the private features of each domain are separated to obtain the shared domain, thereby avoiding interference of unknown classes and noise during feature alignment. Results on two datasets demonstrate that the proposed method outperforms state-of-the-art methods and has more prospects for ensuring the reliability of mechanical equipment.