Deep learning-based crack detection with visual interpretability for rock mass engineering in complex environments

Accurate detection of rock cracks underpins structural health monitoring for rock mass engineering structures. However, existing detection methods suffer from the strong nonlinearity and multi-scale variability of underground crack morphologies, as well as severe interference from complex underground environments. They not only struggle to achieve highly reliable crack detection but also fail to effectively extract complex geometric features encoding crack engineering semantics. To address this challenge, a dedicated benchmark dataset of underground rock mass engineering cracks was established, and nine state-of-the-art object detection algorithms were systematically evaluated, thereby defining a comprehensive performance benchmark for underground crack detection tasks. To tackle the issues of insufficient discriminative power in crack feature extraction and uncertainty in recognition accuracy under complex underground conditions, this study designed a Crack Feature Extraction Module (CFEM) based on dynamic snake convolution, which can adaptively capture the geometric features of cracks with complex configurations, and embedded it into the YOLOv11-n backbone network to develop the specialized model CMYOLOv11 for underground rock mass engineering crack detection. Experimental results showed that CMYOLOv11 achieved 79.37% mAP@50, outperforming the baseline YOLOv11-n by 7.46 percentage points. t-SNE and UMAP visualizations demonstrate excellent feature clustering performance and significant inter-class separability, while Grad-CAM maps confirm more precise crack localization and enhanced model interpretability. Finally, the proposed model was further validated through field tests in two underground mine roadways, demonstrating its applicability under representative field conditions.