X-Bolt: an explainable vision-based framework for bolt loosening detection using a novel visual target

Bolt loosening detection plays an important role in structure maintenance, but traditional detection methods prove inadequate for large-scale infrastructure inspection. Recent advances in computer vision and deep learning (DL) provide more accurate and rapid detection capabilities, but the inherent issue of the model’s black-box nature cannot be ignored. This article proposes an explainable bolt loosening detection framework, named X-Bolt. This framework features a novel visual target establishing physical–visual correlations with bolt states, coupled with BoltYOLO, a customized neural network achieving accurate and real-time detection. Through gradient-weighted class activation mapping-based multiperspective interpretation, the alignment between the model’s decision-making rationale and the visual target is made explicit, and the feature learning procedure of the model, particularly the attention allocation behavior, is effectively visualized. In summary, X-Bolt provides an efficient and interpretable solution for vision-based bolt loosening detection, while providing a generalizable framework for similar inspection tasks involving DL, thereby enhancing trustworthiness and deployability in artificial intelligence-based structural health monitoring.