Adaptively Weighted Semantic Autoencoder for Zero-Shot Compound Fault Diagnosis

It is a great challenging task to diagnose compound faults of rolling bearings because of the complex coupling characteristics of the single faults. Compound fault samples are generally requisite to establish traditional compound fault diagnosis models. However, the infrequency of compound faults in bearings within industrial scenarios may result in a lack of available corresponding data for training the models. To address the above issue, this article proposes a new model named adaptively weighted semantic autoencoder (AWSAE) for bearing compound fault diagnosis based on zero-shot learning (ZSL). Specifically, the proposed AWSAE constructs compound fault semantics by weighted superposition of the semantics of the accessible single faults, in which the weights are adaptively determined by an attention mechanism. A projection matrix is established by the semantic autoencoder (SAE) that can effectively project the features of the tested compound fault samples to the corresponding semantics. Euclidean distances are then calculated in the semantic space to diagnose the types of the compound faults. In addition, to realize generalization zero-shot diagnosis, a prejudgment strategy is designed by integrating the idea of decoupling learning. The architecture of the proposed AWSAE model is simple because the feature extractor is used repeatedly in the prejudgment, the obtaining of adaptive weights, as well as the classification of all the health states. The proposed method is verified on two bearing datasets acquired from a rotor-bearing system and a wheel-rail system. The results show that the proposed AWSAE model performs well in identifying the bearing compound faults and is superior to the most advanced ZSL models.