Zero-Sample Fault Diagnosis for Bearings Using an Hierachical Constrast Learning Approach

Fault diagnosis in rotating machinery is critical for ensuring reliable operation and reducing maintenance costs. Bearings, being essential components, require particular attention. Traditional diagnostic methods depend on labeled training data, limiting their ability to detect previously unseen faults. This limitation has driven the development of zero-sample fault diagnosis. Existing zero-sample approaches depend on manual semantic frameworks, which require significant expert input. Automated methods, however, often create semantic redundancy, reducing accuracy. To address these challenges, this paper introduces a novel hierarchical automatic semantic construction framework based on contrastive learning (HCL). The proposed method extracts fault features using contrastive learning. The semantics are organized into a hierarchical tree structure based on the contrast measurement of fault categories. Next, two autoencoders and a semantic classifier are employed to align extracted fault features with their corresponding semantic vectors, ensuring accurate fault recognition. Experimental results on a bearing dataset evaluated the effectiveness of the method.