Quantifying sensor contribution in vibration-based structural health monitoring using explainable multichannel convolutional neural networks

Deep learning models are increasingly used in vibration-based structural health monitoring (SHM) but operate as black boxes obscuring each sensor’s contribution to damage detection. This work introduces a compact and generalizable signal processing framework that integrates multi-channel time–frequency (TF) analysis with explainable artificial intelligence (XAI) to interpret model decisions and quantify sensor relevance. A triple-input convolutional neural network (CNN) is trained to classify damage states from TF spectrograms recorded by three accelerometers on a beam subjected to a moving-mass excitation. The gradient-weighted class activation mapping++ (Grad-CAM++) is adapted to produce class-wise saliency maps, from which a sensor contribution index (SCI) is derived to measure the relative informational value of each channel. The SCI revealed one sensor to be redundant, which implies that its removal preserves classification accuracy, while at the same time reduces instrumentation and computational effort. The saliency patterns align with structural dynamics parameters (e.g., eigenfrequency shifts), thus providing a physics-consistent validation of the learned TF features. Besides SHM, the proposed TF-XAI methodology offers insights in multi-sensor signal interpretation and optimized sensor placement.