Adversarial Embedding Fusion Network for Multi-sensor Fusion Fault Diagnosis of Wheelset Bearings

Wheelset bearing health conditions can be accurately recognized by using multi-sensor signals. Existing intelligent fault diagnosis methods, however, are short of an explicit learning mechanism to reduce the distribution differences between the features of different sensor signals caused by irrelevant noise, and effectively learn the joint representation of multi-sensor information. To overcome the above problem, a multi-sensor information fusion intelligent fault diagnosis method named adversarial embedding fusion network (AEFN) is proposed for wheelset bearings diagnosis. In the proposed AEFN, multiple generators first extract features from multi-sensor signals while embedding them in latent space. Then, multiple discriminators are constructed to form adversarial learning between it and generators, guiding generators to learn the joint representation and suppress irrelevant information. Finally, the fused features are used to recognize the health conditions of wheelset bearings through a classifier. The proposed AEFN is evaluated using multisensor monitoring data of wheelset bearings from an equally scaled train bogie testbed, and compared with some state-of-the- art wheelset bearing fault diagnosis methods based on multisensor information fusion. Experimental results demonstrate the effectiveness and superiority of the proposed AEFN.