Zero-Shot Compound Fault Diagnosis Based on Weighted Semantic Autoencoder

Abstract

Compound fault diagnosis of rolling bearings is always a challenging task because of complex coupling relationship of different single faults. Traditional diagnostic models require available compound fault samples for model training. However, due to the rarity of the bearing compound faults in industrial scenarios, there may be no compound fault data available to train the model. To this end, this paper proposes a new zero-shot compound fault diagnosis model based on weighted semantic autoencoder (WSAE). Specifically, the proposed WSAE considers the amplitude intensities of different single faults when constructing the semantics of compound faults, and establishes a projection matrix that can effectively project the features of the compound fault samples to the corresponding semantics. A pre-judge module is also designed to roughly classify the test samples into seen or unseen class in the beginning of the bearing fault diagnosis, which is trained via the data of bearings with healthy condition, single faults and fake compound faults. The proposed method is verified on a bearing dataset that were acquired from a self-built bearing test bench. The results show that the WSAE model can effectively identify the bearing compound faults and outperforms the state-of-the-art zero-shot learning-based models.