Intelligent Identification of Overburden Fractures in Physical Simulation Experiments Based on Improved U-Net Deep Learning and Multimodal Data Fusion

Abstract To enhance the intelligence level of overburden fracture recognition in physical similarity simulation experiments, this study proposes an automatic fracture identification method based on an improved U-Net deep learning model integrated with multimodal data fusion. A comprehensive intelligent recognition system was developed, incorporating real-time inference, visual interaction, and 3D reconstruction, enabling dynamic tracking and accurate extraction of fracture features from high-resolution images. The method integrates RGB imagery, FMI resistivity data, and multispectral images, and employs physically driven normalization strategies such as Z-score standardization, linear scaling, and band-wise normalization. To further improve recognition accuracy under complex backgrounds, a sliding-window strategy and attention gating mechanism are introduced. Structurally, the network employs ResNet-34 as the encoder backbone, coupled with a progressive decoder and dual attention modules (CBAM), effectively enhancing multi-scale semantic information fusion. Experimental results demonstrate that the system achieves a real-time processing speed of 200 frames per second and a Dice coefficient of 0.91, significantly outperforming traditional manual interpretation and static segmentation methods. In addition, 3D point cloud reconstruction and geometric quantification of fractures expand the model’s applicability in scenarios such as strata pressure analysis, gas drainage, and ground pressure monitoring. This research provides both theoretical foundations and engineering pathways for the digital identification of overburden fractures and the intelligent upgrade of physical simulation platforms, promoting a transition in mining engineering from manual, experience-based decisions to intelligent and quantitative control, with promising prospects for broader adoption.