Comparative analysis of explainable artificial intelligence methods in sensor fault diagnosis for vibration-based structural health monitoring

Reliable structural health monitoring (SHM) depends on accurate analysis of sensor data as well as on fault-free operation of sensors collecting the data, which is usually ensured via sensor fault diagnosis (FD). Artificial intelligence (AI) models have shown promising poten-tial for sensor FD, mainly exploiting the pattern-learning capabilities in the sensor data. However, the logic behind AI predictions is often opaque, limiting practitioner trust and adoption in autonomous SHM systems. To counteract the aforementioned limitation, ex-plainable artificial intelligence (XAI) has emerged, aiming to improve transparency of AI models by indicating which aspects of input sensor data influence AI-derived predictions. Nevertheless, XAI applications in SHM remain rare, and most studies focus on the feasibility of applying XAI for SHM rather than practical aspects, such as the reliability and robust-ness of XAI explanations. This study investigates explanation robustness in XAI-based sen-sor FD for vibration-based SHM through a comparative analysis of different XAI methods. In particular, AI models are trained on sensor data, and synthetic faults are injected in the sensor data, simulating realistic sensor faults. Explanation robustness is analyzed using at-tribution consistency (Spearman rank correlation coefficient), and attribution stability (co-efficient of variation) across multiple fault scenarios, comparing gradient-based and per-turbation-based XAI methods. Validation tests are presented using acceleration response data, collected from a pedestrian bridge. The validation results show that gradient-based methods demonstrate higher explanation robustness than perturbation-based methods when explaining vibration-based sensor fault patterns. Future work will extend the comparative analysis to distinguish between anomalies caused by sensor faults and anomalies arising from changes in structural dynamic behavior.