With the rapid development of artificial intelligence (AI) technologies, deep learning-based structural health monitoring (DeepSHM) methods have gained significant attention. However, their black box nature often limits interpretability and trust. The field of Explainable AI (XAI) aims to address this by enhancing model transparency and reliability through human-comprehensible explanations. This study investigates the use of XAI algorithms in interpreting a 1D convolutional neural network (1D CNN) developed for Lamb wave monitoring of bolt-loosening detection in multi-bolted double-layer aluminum plates under varying temperatures. Four existing XAI algorithms were employed, including Sensitivity Analysis, Deep Taylor, Gradient-weighted Class Activation Mapping (Grad CAM) and Guided Grad CAM. In addition, this paper introduces two new XAI methods, Smooth Simple Taylor and Deep Grad CAM as an enhancement of the Simple Taylor and Grad CAM methods, respectively. These six XAI algorithms were used to establish the relation between the 1D CNN model parameters and the input vector. The results were evaluated for their effectiveness in comparison to the physical insights of the input vector using two proposed methods, namely the Correlation Coefficient with Residual Signal and the Residual Signal Weighted Importance Score Ratio. The results of the evaluation methods, in conjunction with Infidelity, Sense sum, and Sanity check, were utilized to rank the performance of the six XAI algorithms. The rankings were consistent in both simulation and experiment data sets, and the newly proposed XAI algorithm, Smooth Simple Taylor, appeared to be the best in both data sets. Overall, this research establishes a novel approach to using XAI algorithms to enhance the explainability of AI in practical engineering applications.

With the improvements in computational power and advances in chip and sensor technology, the applications of machine learning (ML) technologies in structural health monitoring (SHM) are increasing rapidly. Compared with traditional methods, deep learning based SHM (Deep SHM) methods are more efficient and have a higher accuracy. However, due to the black box nature of deep learning, the trained models are usually difficult to interpret, which blocks their practical application. Therefore, it is of great importance to develop explainable artificial intelligence (XAI) methods to understand the internal decision-making mechanisms of damage classification in Deep SHM. In this paper, a novel XAI algorithm named Deep Gradient-weighted Class Activation Mapping (Deep Grad CAM) is proposed by combining the existing method Grad CAM with the convolutional neural network (CNN) deconvolution mechanism. In this paper, Deep Grad CAM is used to interpret a one-dimensional convolutional neural network trained to detect bolt loosening based on guided wave propagation. The interpretation performance of Deep Grad CAM is compared with Grad CAM, and their performances are quantified using Infidelity. The results show that the Infidelity of Deep Grad CAM is much smaller than that of Grad CAM, indicating significant improvements in explanation accuracy and reliability.

With the increasing application of artificial intelligence (AI) techniques in the field of structural health monitoring (SHM), there is a growing interest in explaining the decision-making of the black-box models in deep learning-based SHM methods. In this work, we take explainability a step further by using it to improve the performance of AI models. In this work, the results of explainable artificial intelligence (XAI) algorithms are used to reduce the input size of a one-dimensional convolutional neural network (1D-CNN), hence simplifying the CNN structure. To select the most accurate XAI algorithm for this purpose, we propose a new evaluation method, feature sensitivity (FS). Utilizing XAI and FS, a reduced dimension 1D-CNN regression model (FS-X1D-CNN) is proposed to locate and predict the torque of loose bolts in a 16-bolt connected aluminum plate under varying temperature conditions. The results were compared with 1D CNN with raw input vector (RI-1D-CNN) and deep autoencoders-1D-CNN (DAE-1D-CNN). It is shown that FS-X1D-CNN achieves the highest prediction accuracy with 5.95 mm in localization and 0.54 Nm in torque prediction, and converges 10 times faster than RI-1D-CNN and 15 times faster than DAE-1D-CNN, while only using a single lamb wave signal path.