Abstract
In the transition from pilot to mass production, certain fault classes lack available samples, creating a zero-shot condition that challenges the training of fault diagnosis models. Zero-shot learning (ZSL) maps features to shared attribute vectors formed by engineering knowledge information, offering a potential solution. In manufacturing processes, disturbances propagate across operations, causing multiple attributes to co-occur. However, the internal co-occurrence relationships among attributes are often overlooked, which may lead to discrepancies in attribute prediction. To address this issue, we propose a zero shot fault diagnosis method for manufacturing processes that leverages attribute co-occurrence relationships to identify unseen faults. To capture this relationship, mutual information is first employed to quantify attribute co-occurrence and construct an adjacency matrix of attribute relations. In parallel, distributional features are extracted using multi-delay ordinal patterns, and a distance-based loss function is designed to align these features with the attributes. The adjacency matrix and distributional features are then input into a graph convolutional network, with attribute relationships embedded into the distributional features to maintain consistency. Finally, distributional features support zero-shot fault diagnosis via mapping and similarity measures. Experimental validation on the Tennessee Eastman Process (TEP) and fuel rods manufacturing process demonstrates the effectiveness of the proposed method.

This publication has 37 references indexed in Scilit: