Zero-shot gear fault diagnosis based on attribute consistency embedding

In industrial scenarios, as a core component of mechanical transmission systems, gears are prone to various faults, such as wear and fatigue after long-term operation. Diagnosing its operating status is of vital importance for ensuring production safety. However, there are many difficulties in actually obtaining gear fault samples, such as the difficulty in collecting rare fault samples and the high cost of collecting some faults. To alleviate this problem, this paper proposes a zero-shot gear fault diagnosis method based on Attribute Consistency Embedding (ACE). The method constructs semantic links between fault samples and attributes by designing semantic vectors of length-width-depth attributes with gear surface faults. The consistent mapping of cross-modal semantics has been achieved. The embeddedreconstruction structure is adopted to jointly map the fault samples and attribute vectors to a unified semantic embedding space. The correlation-consistent embedding matrix with an orthogonal constraint optimization strategy is introduced to strengthen the clustering of similar faults and the separability from dissimilar faults. Experiments were conducted on the dataset collected on the gear test bench. The results show that the average diagnostic accuracy rate of the method proposed in this paper in the unseen fault category reaches 93.65%. The fault features and attribute vectors can be precisely aligned in the semantic embedding space through the correlation-consistent embedding matrix and feature visualization. Enable the samples of the same type of faults to be highly clustered in the feature space and the faults of different types to be significantly distinguishable. The Grad-CAM heat map further indicates that the key fault characteristics such as high-frequency shock in the vibration signal can be effectively retained through the embedding-reconstruction structure. It is conducive to providing a visual verification basis for the fault diagnosis model.