Robust and interpretable sensor optimisation for bridge structural health monitoring using POD–QR screening and explainable deep learning with grey relational fusion

Real-time health monitoring of bridges relies on sensor networks to capture structural response data, where the selection of an optimal sensor configuration is crucial for achieving reliable damage detection while minimising instrumentation and operational costs. This study proposes a unified sensor optimisation strategy for vibration-based damage classification that incorporates proper orthogonal decomposition (POD) and QR pivoting for sensor independence, masking-based deep learning sensitivity, and integrated gradient-based saliency analysis for interpretability. A truss bridge model subjected to moving train loads is simulated under multiple damage scenarios to evaluate the proposed methodology. A hybrid long short-term memory-gated recurrent unit model is developed to perform damage detection from multichannel acceleration time-series data. The importance of sensors is assessed using three independent yet complementary methods, and the resulting indicators are normalised and integrated through weighted grey relational analysis (WGRA) to obtain a robust and unified sensor ranking. The WGRA-guided incremental sensor-subset evaluation strategy is employed to identify compact sensor configurations whose classification accuracy remains comparable to the full-sensor baseline. The proposed method is also validated through experimentally collected multi-sensor acceleration data of a real-life truss bridge (Hell Bridge test arena) under progressive damage scenarios. The results demonstrate that near full-sensor classification performance can be achieved with sensor reductions of 65 and 80% in the numerical and experimental studies, respectively, highlighting the practical potential of the proposed framework for cost-effective structural health monitoring.