A Novel Multidynamic Domain Adaptation Transfer Learning Method for Fault Diagnosis of Bearings With Insufficient Labeled Data

Recently, intelligent diagnosis methods for rotating machines have achieved prominent results. Existing intelligent methods rely on two conditions: 1) massive labeled data is necessary during training process; 2) the data of application scenarios and training data is under the same working condition. In some actual industrial scenarios, however, labeled samples are insufficient and working conditions are variable. To address this problem, we propose a novel intelligent fault diagnosis method named Multi Dynamic Domain adaptation Network (MDDAN) based on transfer learning, which can diagnose bearing fault with insufficient labeled data under varying working conditions. The crucial architecture of the proposed MDDAN is a feature extractor module and a multi adaptation module, which are designed to learn domain-invariant features with insufficient labeled data. Furthermore, the idea of adversarial training is introduced by the domain discriminators part of the multi adaptation module, which can improve the domain adaptation performance. To balance the contributions of global domain and sub domain discriminators, we add a dynamic adaptation strategy to domain adaptation module. Finally, Pareto-Efficient optimization is introduced to adaptively coordinate multi losses and metrics that further improves the stability and domain adaptation ability of MDDAN. The feasibility and effectiveness of MDDAN are verified on three datasets through a variety of scenarios transfer experiments.