Deep learning-based open set multi-source domain adaptation with complementary transferability metric for mechanical fault diagnosis

Intelligent fault diagnosis aims to build robust mechanical condition recognition models with limited dataset. At this stage, fault diagnosis faces two practical challenges: (1) the variability of mechanical working conditions makes the collected data distribution inconsistent, which brings about the domain shift; (2) some unpredictable unknown fault modes that do not observe in the training dataset may occur in the testing scenario, leading to a category gap. In order to cope with these two entangled challenges, an open set multi-source domain adaptation approach is developed in this study. Specifically, a complementary transferability metric defined on multiple classifiers is introduced to quantify the similarity of each target sample to known classes to weight the adversarial mechanism. By applying an unknown mode detector, unknown faults can be automatically identified. Moreover, a multi-source mutual-supervised strategy is further adopted to mine relevant information between different sources to enhance the model performance. Extensive experiments are conducted on three rotating machinery datasets, and the results show that the proposed method is superior to traditional domain adaptation approaches in the mechanical diagnosis issues that new fault modes occur.