Explainable AI-Driven Optimal Feature Selection for the Identification of Structural Damage

Abstract

The existing scholarly investigations into intelligent structural damage recognition predominantly emphasize the enhancement of the precision and efficacy of damage detection. Nonetheless, the opaque “black box” characteristic inherent to deep learning frameworks constrains users’ comprehension of the underlying decision-making mechanisms, which significantly obstructs their practical progression and execution. Consequently, this manuscript employs the interpretative framework known as Shapley Additive exPlanation (SHAP) to elucidate and scrutinize the attributes of a convolutional neural network–based intelligent structural damage recognition model, while also proposing a methodology for the optimization of features pertinent to structural damage recognition. In particular, this inquiry clarifies the foundational principles that govern the output results of damage assessment and identifies the prospective optimal characteristics of structural damage identification signals. In assessing the contribution of various features to the results of damage recognition and the interrelations among these features, both global and local perspectives of the damage signal were taken into account. The interpretation and analysis of damage recognition signal characteristics can facilitate the selection of structural damage recognition features, thereby aiding deep learning models in the extraction of high-dimensional features and markedly enhancing the recognition accuracy of structural damage identification. The efficacy of the proposed algorithm was corroborated through two experimental scenarios, with results indicating that the accuracy of the structural damage identification algorithm delineated in this study surpassed 95%. This research offers thorough guidance for the implementation of SHAP analysis within intelligent structural damage models, and the findings hold significant implications for augmenting the interpretability of intelligent damage identification algorithms.