An Open Set Diagnosis Method for Rolling Bearing Faults Based on Prototype and Reconstructed Integrated Network

Abstract

To solve the diagnosis of misjudgment in the diagnosis of the unknown bearing fault, an intelligent open set fault diagnosis method is proposed for rolling bearings based on the integration of prototype and reconstructed networks. First, owing the prototype network, the model can learn more discriminative features to achieve accurate identification of the closed set of known fault classes. Second, we reconstruct the signal through the reconstruction network, which determines whether the fault is known or unknown according to the correlation between the reconstructed signal and the input signal. And the open set fault pattern recognition is realized, which avoids the problem of misjudging unknown faults as specific types of known defects. Finally, the model performance is evaluated from the perspectives of closed set and open set on the Western Reserve University (CWRU) and Northeastern University (NEU) datasets. Experiments show that the method proposed is superior to the comparison methods [one class SVM (OCSVM) and extreme value machine (EVM)], which has a high ability to identify unknown classes and reclassify known classes.