Zero-Shot Rolling Bearing Compound Fault Diagnosis Based on Envelope Spectrum Semantic Construction

Deep learning methods have achieved remarkable fault recognition accuracy with sufficient training data. However, in real industrial scenarios, the presence of compound faults poses challenges in obtaining ample data for training deep learning models. To address this issue, a zero-shot compound fault diagnosis approach based on envelope spectrum semantic construction is proposed, specifically tailored for addressing compound diagnosis in rolling bearings. The key idea involves training on data with single fault to construct a semantic space and a feature space, respectively. Then in the compound fault identification phase, a combination of these two spaces enables recognition of compound faults under zero-shot scenarios. Considering envelope spectrum has been proven as an effective technique for characterizing rolling bearing fault characteristics, this research preprocesses fault signals using envelope spectrum to enhance distinctive fault features. Moreover, the physical implications of signal envelope spectrum are harnessed to formulate semantics for both single and compound bearing faults. Experimental results demonstrate the proposed model's achievement of 87.83% accuracy in compound fault recognition, outperforming other models.