Joint distribution adaptation with diverse feature aggregation: A new transfer learning framework for bearing diagnosis across different machines

Abstract On account of lacking labeled samples for the bearing fault diagnosis in real engineering applications, transfer learning is widely investigated for transferring diagnosis information. A more challenging but realistic scenario called transfer across different machines (TDM) is investigated in this paper where previous approaches may degenerate greatly with more drastic domain shifts. A joint distribution adaptation-based transfer network with diverse feature aggregation (JDFA) is proposed, where the diverse feature aggregation module is added to enhance feature extraction capability across large domain gaps. Then the joint maximum mean discrepancy between source and target domain samples is adopted to reduce the distribution discrepancy automatically. Extensive TDM transfer learning experiments are conducted. The average accuracy reaches 99.178% that is much higher than state-of-the-art methods, demonstrating the proposed JDFA framework can effectively achieve superior diagnostic performance, and significantly promote fault diagnosis research under TDM scenario in view of applicability and practicability of algorithms.