Loss of feature salience: An interpretable network diagnosis framework for multimodal industrial monitoring

Multimodal monitoring data from complex industrial systems contain rich information regarding system health conditions. However, extracting interpretable and localizable fault features from such data remains a critical challenge. This paper proposes an interpretable multimodal network diagnosis framework centered on the concept of Loss of Feature Salience . This concept hypothesizes that functional degradation of a physical component may be reflected in systematic changes in the topological salience of its corresponding node or nodes within association networks constructed from multimodal signals. To operationalize this concept, the framework incorporates three supporting technical components. First, each modal time series is transformed into a representation network using a sliding-window weighted limited-penetrable visibility graph to preserve local signal variations. Second, node-topology-based multilayer fusion is employed to integrate cross-modal evidence. Third, distribution-adaptive thresholding is used to determine the relative salience state of each node feature. By establishing a predefined spatial correspondence between each monitored physical component and a designated network node, changes in node-level feature salience can be traced back to the corresponding component, thereby supporting interpretable component-level fault localization. Using compensation-capacitor diagnosis in jointless track circuits as an engineering validation case, the proposed method correctly identified and localized all three documented fault cases (3/3) and produced no false alarms among the 93 healthy samples (0/93) across the three evaluated railway datasets. Post-hoc examination of the three documented fault cases identified a recurring empirical pattern, termed “Three Losses and Two Rises”: reduced salience in three functional features—degree centrality, betweenness centrality, and influence—accompanied by elevated values of two structural features—inter-layer coupling and clustering coefficient. These results demonstrate the potential of the proposed framework to provide component-level diagnostic evidence with explicit physical and topological meanings in the investigated track-circuit setting. More broadly, the framework offers a transferable analytical template for physics-guided and intrinsically interpretable diagnosis, although its feature semantics, parameter settings, and diagnostic patterns require domain-specific validation before application to other knowledge-intensive engineering systems.