A hybrid semantic-based embedded zero-shot learning method for compound fault diagnosis of bearings

Abstract Due to the randomness and complexity of compound bearing faults, their diagnosis has remained a challenging problem. Current deep learning methods heavily depend on extensive compound failure samples for model training, yet obtaining adequate and diverse fault samples from practical industrial settings remains a significant challenging. To tackle the diversity and scarcity of compound fault samples, a hybrid semantics-driven zero-shot learning framework is presented to identify untrained compound faults using trained single fault samples. This framework constructs a novel semantic space by combining both human-defined and machine-extracted semantics, thereby generating single-fault representations that can be synthesized into compound-fault semantics. Fault features are extracted using Symmetrized Dot Patterns derived from raw vibration signals, which serve as input to the feature extractor. A semantic embedding module is then utilized to align fault semantics with extracted features, enabling fault diagnosis and classification. Experiments conducted on data captured from a custom-built bearing test rig validate the performance and advantage of the proposed framework, achieving a compound fault diagnosis accuracy of 75.94% and demonstrating excellent diagnostic performance.