Interpretable research on the health monitoring network of prefabricated building beam–column joints

Abstract

As neural network models increasingly apply in structural health monitoring (SHM), their strong data processing and generalization abilities have been demonstrated in damage identification, localization, performance prediction, and early warning. Consequently, enhancing model interpretability and addressing the “black-box” nature to improve user trust have become key research priorities. To address these challenges, this study initially establishes a single-layer assembled steel frame structure, where accelerometer signals from four positions on beam components are acquired through controlled small hammer impact tests conducted under 10 damage conditions at the beam–column joints. Subsequent to data collection, a health monitoring model is devised that leverages one-dimensional convolutional neural networks that are trained to effectively discriminate between accelerometer signals under the 10 damage conditions while exhibiting robustness against varying noise levels. To shed light on the network’s decision-making process, gradient-weighted class activation mapping (Grad-CAM) is employed to elucidate the network’s degree of attention to different parts of the input data during the learning phase. Furthermore, the input signals were decomposed into single-modal subsequences through singular spectrum analysis, with Grad-CAM heatmaps illustrating the attention distribution within these subsequences, thereby visualizing the network’s learning process. Finally, a comparative analysis was conducted between the proposed visualized neural network model and traditional spectral analysis methods in terms of their advantages and limitations for classifying SHM signals. The study revealed the inherent periodicity of the structural vibration acceleration signals and identified the model’s periodic high-activation behavior during classification. This indicates that the model is capable of automatically recognizing the internal periodic patterns of the signals, thereby enhancing its credibility.