Edge-based explainable multisensor fault diagnosis for structural health monitoring

Structural health monitoring (SHM) systems depend on reliable sensor measurements, rendering sensor fault diagnosis essential for maintaining data quality and supporting decision-making. Deep learning enables automated sensor fault diagnosis, while explainable artificial intelligence (XAI) improves the transparency of model predictions. In parallel, edge computing processes monitoring data closer to sensor networks and reduces dependence on centralized systems. However, most edge-based SHM studies focus on prediction performance and computational feasibility, while agreement between predictions and explanations remains insufficiently examined. This study proposes an integrated edge-based XAI-enabled evaluation methodology for multisensor fault diagnosis in SHM. Sensor faults are formulated as classes combining the faulty sensor and fault type (sensor-fault classes), enabling sensor fault detection, isolation, and identification within a single classification task. An established gated recurrent unit-based sequence-to-label classifier is used to classify multisensor time-series windows into “no-fault” or “sensor-fault” classes. Gradient × input generates time-step-level attributions, which are aggregated into sensor-level attributions. The main contribution is a task-specific quantitative evaluation of whether the dominant attributed sensor agrees with the sensor encoded in the predicted class. Prediction performance, explanation faithfulness, and sensor-level prediction–explanation agreement are evaluated for an independently trained edge deployment and compared with a reference centralized deployment under controlled implementation conditions. Local explanation stability, attribution-method dependence, and the influence of chronological response portions are additionally examined for the centralized deployment, whereas computational feasibility is assessed only for the edge deployment. The results show broadly comparable prediction performance and aggregate sensor-level attribution behavior between the two deployments, although neither achieves full prediction–explanation agreement. Supplementary analyses indicate moderate local faithfulness and high stability of the sensor-attribution distributions, while dominant-sensor selection varies among attribution methods and classification performance is sensitive to chronological response portions for specific fault classes. The measured mean runtime indicates that, on average, prediction and explanation generation are completed on the next unit of computing (NUC)-based edge-computing node within the selected window update interval. The findings provide proof-of-concept evidence for the investigated edge-based XAI-enabled multisensor fault diagnosis, establishing a basis for further development of decentralized SHM systems with larger sensor networks, different structures, and varying operating and environmental conditions.