Integrating SHAP analysis in deep learning models for structural health monitoring

Structural Health Monitoring (SHM) is a critical discipline in modern engineering, providing the foundation for predictive maintenance and safety assurance across a range of infrastructures, from bridges and buildings to industrial machinery. Traditional SHM approaches, while effective, often face significant limitations related to the volume and complexity of vibration data, the computational overhead of dense sensor networks, and the opacity of black-box machine learning models. This thesis proposes a novel framework that integrates SHapley Additive exPlanations (SHAP) into deep learning architectures to enhance the interpretability, efficiency, and deployability of SHM systems in real-world, resource-constrained environments. The research is grounded in three case studies based on the benchmark Z24 bridge dataset, each addressing a distinct challenge in SHM: (1) leveraging compressed sensing and one-class classification neural networks (OCCNN) to reduce data transmission without sacrificing accuracy; (2) employing Hardware-Aware Neural Architecture Search (HW-NAS) to optimize model complexity for microcontroller deployment; and (3) implementing distributed, end-to-end sparse models suitable for edge computing environments. In all cases, SHAP analysis is applied to quantify the relative importance of input features—such as modal frequencies, temperature, and compressed sensor signals—thereby enabling engineers to validate model decisions and gain actionable insights into structural behavior. Experimental results demonstrate that the integration of SHAP not only preserves but enhances model transparency across all scenarios, revealing consistent patterns such as the dominant influence of specific sensors and the critical role of environmental parameters. Furthermore, models compressed to as few as 122 parameters maintained over 97% classification accuracy and were shown to be robust to simulated noise.