CDAAN: A Component Domain Adaptation Adversarial Network With Antinegative Transfer Property for Mechanical Fault Diagnosis

Due to the differences such as equipment system, working conditions, fault occurrence and spread, Intelligent Bearing Fault Diagnosis (IBFD) is required to have higher Domain Adaptation (DA) capabilities. Mainstream DA methods take the signal ontology as the analysis object, which is prone to Negative Transfer (NT) due to the non-uniformity and inconsistency of fault information distribution. To overcome the aforementioned problem, this paper proposes a Component Domain Adaptation Adversarial Network (CDAAN), which uses appropriate components to replace original signals to improve the data quality. First, the wavelet transform is employed to decompose the signal into several components to separate purer fault information. Second, a component matching method based on transferability evaluation is proposed, comprehensively considering knowledge validity, distribution discrepancy and cross-domain generalization ability, to achieve mutual matching of components and recombine them into new inputs. Finally, the adversarial learning strategy is used to optimize the network parameters, and the outputs based on each component combination are weighted and summed as the final result. To test the performance of the proposed CDAAN for mechanical fault diagnosis in practice, two cases, namely traditional bearing test and cryogenic test of oxidant pump in aerospace engine, are investigated. The results show that the proposed CDAAN has better DA capability than comparison methods and effectively reduces the NT risk through reasonable component matching.