Low-Priori Intervention Transfer Learning Network for Large-Span Fault Diagnosis

Current domain adaptation (DA) methods can solve feature alignment and knowledge transfer problems between the source and target domains of transfer diagnostic domain; however, most of these methods only focus on aligning marginal distributions between two domains and ignore discriminative separable feature learning process. Therefore, in order to solve the problem of diagnostic accuracy constraints in large-span working scenarios, low-priori intervention transfer learning network (LPTLN) is proposed. Softmax and central loss are combined to preprocess and extract discriminative features that are more conducive to transfer diagnosis. Subsequently, a minimum number of target prior labels are embedded (only one sample is extracted for each class of faults) to perform supervised training. Improved enhanced joint DA mechanism (IEJDA) promotes more thorough joint distribution alignment (JDA); in IEJDA, to eliminate domain confusion, a new distribution discrepancy measure is formed by combining maximum mean discrepancy (MMD) and correlation alignment (CORAL) and embed prior labels to perform DA diagnosis. Using two sets of fault datasets, experimental results show that the proposed LPTLN has higher diagnostic performance and stronger robustness than other typical DA methods.