Unify the Diversity: Consistent-Inconsistent Semantic Collaboration Learning for Zero-Shot Industrial Fault Diagnosis

Zero-shot diagnosis (ZSD) uses expert-provided semantics to identify previously unseen industrial faults. It is first recognized that some fault semantics may change with varying working conditions, which are denoted as a semantic inconsistency issue here, while the remaining unchanged ones are denoted as consistent semantics. Changes in working conditions result in shifts of both data and semantic distributions, leading to mapping confusion between data and semantics in existing ZSD methods. For the semantic inconsistency issue, analyzing the two types of semantics to distill their respective diverse and common information is essential to clarify the mapping confusion. Therefore, we propose a consistent-inconsistent semantic collaboration learning (CISCL) method to identify the true link between data and semantics. First, in the training phase, we design consistent-inconsistent semantic supervision learning (CISSL) to refine the common feature of the data and align it with consistent semantics. Retaining consistent semantics, CISSL eliminates condition-related inconsistent semantics by adversarial techniques. This allows the model to diagnose unseen faults even under new conditions. Second, in the inference phase, we design a sequential reasoning mechanism to leverage inconsistent semantics while maintaining the foundational status of consistent semantics. When consistent semantics alone cannot pinpoint a specific fault, a rough target range is established, followed by a neighbor search within this range that incorporates inconsistent semantics. Our approach has been validated on a physical experimental three-phase motor rig and a thermal power plant process, and the validation confirms the effectiveness of our methodology.