Data Generation and Prototype Learning Based Open-Set Fault Diagnosis for Rotating Machinery

Deep learning-based fault diagnosis method has achieved promising results under close-set assumption where the test data label set is the same as the training data label set. However, the real diagnosis scenario can be an open-set problem and detecting the unknown faults accurately is essential. Focusing on this issue, a data generation and prototype learning based open-set fault diagnosis method is proposed for rotating machinery in this paper. First, a conditional generative adversarial network (GAN) is developed for open-set sample simulation. Second, a prototype network is trained by known samples for known fault classification and the prototype features for known classes are acquired. Then, the simulated open-set data are input to the prototype network and the pseudo prototypes of the open space are calculated by K-means clustering. Finally, a distance-based strategy is designed for known class classification and unknown detection. Two case studies demonstrate that the proposed method is feasible for open-set fault diagnosis. Based on the synthetic open-set sample generation and prototype network, the proposed method can describe the open space reasonably and separate the known classes and unknown class effectively.