Explainable Deep Learning for Structural Health Monitoring in Smart City Infrastructure: A Convolutional Autoencoder Approach with SHAP Analysis

Abstract

Anomaly detection by means of deep learning models has demonstrated strong performance in vibration-based structural health monitoring; however, its widespread adoption in smart city infrastructure management is constrained by the black-box nature of the models, flagging structural anomalies without providing a physically interpretable explanation to the engineers required to act on the detected anomaly. Although recent studies that applied one-class deep learning methods to the Z24 bridge benchmark established strong detection baselines, none have provided systematic SHAP-based spatial feature attribution across all measurement setups of the Z24 Progressive Damage Test (PDT) dataset to interpret the physical meaning of unsupervised anomaly scores.

This thesis demonstrates that a pipeline integrating a phase-invariant Conv1D autoencoder with a SHAP explainability framework enables reliable anomaly detection on the Z24 bridge and provides a quantifiable empirical settlement threshold below which any structural attribution is unreliable for progressive pier settlement. The autoencoder was trained on a log-PSD representation of healthy ambient vibration data from five reference channels and validated against 4 settlement scenarios out of the 17 PDT scenarios. SHAP feature attributions were computed using a validated XGBoost surrogate trained on 92 physically named spectral features, and TreeSHAP was applied to attribute each anomaly score to specific vibration characteristics.

The AUROC exceeded 0.95 across all nine spatial measurement zones (mean = 0.987), and the SHAP attribution identified the inter-sensor transmissibility ratio R3V/R2V peak frequency as the dominant feature, with the lateral mid-span channel R2L accumulating the highest importance across all independently trained models, and most importantly, reliable SHAP explanations were only achievable from 80 mm settlement onward, below which surrogate fidelity collapsed. These findings reframe the objective of deep learning SHM systems from maximising detection accuracy alone to simultaneously enabling physically interpretable explanations.