Interpretable AI for vibration-based structural health monitoring: a comparative study of CNN and transformer architectures on a benchmark shear building

Abstract

The proliferation of Artificial Intelligence (AI) in Structural Health Monitoring (SHM) has catalyzed a paradigm shift from traditional, feature-based damage detection to end-to-end, data-driven methodologies. While Deep Learning (DL) models, particularly Convolutional Neural Networks (CNNs), have demonstrated remarkable efficacy, the advent of Transformer architectures presents a new frontier with unparalleled capabilities for sequence modeling. However, a direct comparative analysis of these architectures on a standardized experimental benchmark, coupled with a deep investigation into their decision-making processes, remains a critical research gap. This study addresses this void by conducting a comprehensive investigation using a publicly available experimental dataset from a six-storey laboratory shear building. We develop, train, and evaluate two distinct DL models: a lightweight one-dimensional CNN (Fast CNN) and a state-of-the-art Transformer-based model (Fast Transformer). Both models are tasked with directly classifying the structural state (undamaged vs. damaged) from raw accelerometer time-series data. Performance evaluation based on standard metrics reveals that both models achieve exceptional accuracy, with the Fast CNN reaching 99.44% and the Fast Transformer reaching 98.87% on validation datasets. This work’s core contribution lies in applying Explainable AI (XAI) techniques, including Integrated Gradients and saliency mapping, to deconstruct these models’ “black box” nature. Our analysis reveals a non-intuitive yet consistent finding: both the CNN and the Transformer primarily focus on the vibration signature of the base sensor (Sensor 1) to detect damage located at the fourth storey. This suggests the models have learned to identify damage through their influence on the structure’s global dynamic response as reflected at their boundary conditions. Furthermore, XAI reveals distinct operational strategies: the CNN acts as a highly localized feature detector, whereas the Transformer leverages its self-attention mechanism to weigh a broader spatiotemporal context. This paper provides a rigorous benchmark for modern DL architectures in vibration-based SHM and tells a technical story of how interpretable AI can uncover novel, physically meaningful damage detection strategies, enhancing trust and guiding future development of intelligent monitoring systems.