Hybrid Deep Learning Architectures for Structural Health Monitoring via PZT Signals: A Multi-Scale Attention AI with Uncertainty Quantification

Abstract

To keep pace with rapid developments in infrastructure, an AI-based damage detection tool is crucial for ensuring safety. However, the rapid deterioration of civil infrastructure necessitates intelligent, autonomous structural health monitoring (SHM) systems capable of real-time damage detection, localization, and quantification. In this article, a novel method for smart structural damage identification is proposed. a novel hybrid architecture integrating electromechanical impedance ( ) sensing with multi-scale attention-based deep learning. Unlike conventional approaches reliant on hand-crafted features, our method employs a Squeeze-and-Excitation Residual Convolutional Network ( ) coupled with multi-head self-attention mechanisms for automatic feature extraction from raw PZT-EMI signals. The model was designed and trained using an effective approach for optimal feature learning from the raw PZT-EMI response signals. Further, Bayesian optimization has increased the network’s reliability.  It substantially increases the network performance and knowledge acquisition process. To validate the proposed approach experimentally, PZT-EMI signals have been obtained from a beam with various structural conditions. The model demonstrates 100% classification accuracy across four damage severity levels (healthy, 25%, 50%, 75% stiffness reduction), with an RMSE of 0.137 cm for damage localization. Comparative study against state-of-the-art methodologies, such as GNN reviles the superior efficiency of the introduced approach. This research establishes a new paradigm for intelligent SHM, bridging the gap between data-driven pattern recognition and physics-aware structural assessment