From Seen to Unseen: Harnessing Temporal Dependencies and Graph Structures for Zero-Sample Fault Diagnosis in Industrial Systems

As industrial systems become increasingly complex and automated, fault diagnosis based on operational sequential data has attracted significant attention. However, in real-world production, the intricate spatiotemporal relationships within industrial systems often result in diverse and unseen fault types. Moreover, economic and safety constraints frequently limit the availability of sufficient fault samples. Therefore, it is crucial to develop methods that effectively extract the coupled dimensional and temporal features from multidimensional industrial time-series data and diagnose unseen faults using knowledge of common seen faults. To address these challenges, we propose an innovative fault diagnosis method that integrates zero-shot learning (ZSL) with graph structures to extend diagnostic capabilities from seen to unseen faults. To analyze complex variable relationships and temporal characteristics, long short-term memory (LSTM) is used to extract temporal features across dimensions, which serve as graph nodes. Simultaneously, mutual information neural estimation (MINE) is used to learn interdimensional relationships, forming the graph edges. These relationships are transmitted and integrated through the graph structure. A public attribute matrix is established to bridge seen and unseen faults, which is then projected onto an embedding vector space to infer unseen fault types based on embedding similarity. This method, termed LSTM-MINE graph network (LMG)-ZSL, relies solely on training with seen fault samples while effectively detecting unseen faults. Experiments conducted on a simulation platform and two real industrial test platforms demonstrate that LMG-ZSL outperforms other advanced methods, achieving higher accuracy and lower standard deviation in identifying unseen faults and fault attribute prediction. Stability tests further confirm its superior performance and robustness.