Physical knowledge-driven feature fusion and reconstruction network for fault diagnosis with incomplete multisource data

Adaptive exploration and utilization of the correlations are the crucial factors in determining the performance of fusion based intelligent diagnosis methods. However, subject to the impact of harsh operating environments in industrial applications, collected multisource data are inevitably suffer from the challenge of incompleteness, directly put these correlations disabled incomplete multisource data for diagnosis necessarily leads to unsatisfactory results. Therefore, a novel wavelet-driven multisource fusion and reconstruction network (WFRN) is proposed, it innovatively adopts a physically interpretable fault knowledge transfer strategy to overcome the incompleteness challenges. Specifically, the developed missing feature reconstruction module is capable of transferring complete fault knowledge to the reconstruction of incomplete information, and the coordinated representation-based reconstruction enables the missing fault feature completion without the restriction of missing modes. Furthermore, the completed information is fused to generate more discriminative fault representations by integrating both views of multi-sensor and multi-time series. Finally, experimental results on an aviation rotor fault simulator not only validate the feasibility and superiority of the proposed WFRN, but also demonstrate its strong adaptability in addressing all potential missing modes in widespread industrial applications.