Objective Accurate identification of the coal-rock interface is a critical step in achieving intelligent coal mining and ensuring safe and efficient production. During the coal mining process, the position of the coal-rock interface is directly related to the rationality of the cutting path and the operational safety of equipment. Inaccurate interface identification can easily lead to cutting into rock, increased equipment wear, and even safety accidents. At the same time, it can also reduce coal recovery rates and production efficiency. Therefore, achieving high-precision identification of the coal-rock interface is of great engineering significance. To address issues such as the inconspicuous differences in coal-rock characteristics, the poor real-time performance of traditional detection methods, and their limited recognition accuracy, a coal-rock interface perception and identification method integrating deep learning and thermal imaging technology is proposed. By exploiting differences in thermal responses between coal and rock, high-precision coal-rock interface identification is achieved, providing reliable technical support for autonomous decision-making and precise control in intelligent coal mining equipment.

Methods Simulated coal-rock specimens were prepared (Fig.10), and an active infrared coal-rock sensing platform for the mining face was established based on the FLIR A655SC infrared thermal imager (Fig.13). A LabVIEW-based system and halogen lamps were used to regulate and excite the light source. Coal-rock thermal imaging and image acquisition were conducted under different excitation parameters. Composite data augmentation methods, such as flipping and noise addition, were applied to the collected data for expansion, thereby constructing a coal-rock thermal radiation image dataset (Fig.14). Meanwhile, an ITR-DeepLabV3+ network model (Fig.2) is proposed for the characteristics of coal-rock thermal radiation images. A lightweight MobileNetV4 is adopted as the backbone, reducing computational complexity while maintaining high segmentation accuracy. The convolutional block attention module (CBAM) is introduced to enhance feature representation and improve the model’s feature extraction capability. By combining depthwise separable convolution with the SE attention mechanism, a DS-ASPP module is proposed to reduce model complexity and strengthen multi-scale feature modeling. In addition, features with 4×, 16×, and 32× downsampling from the backbone are utilized, and a shallow multi-scale feature fusion enhancement module is constructed by integrating the coordinate attention (CA) mechanism, enabling effective complementarity of multi-level feature information.

Results and Discussions Ablation experiments were conducted on the proposed modules, and the results show that all the proposed modules effectively enhance model performance while balancing segmentation accuracy and model complexity (Tab.2, Tab.3, Tab.4). The proposed ITR-DeepLabV3+ model achieves an Intersection over Union (IoU) of 91.87%, a Mean Intersection over Union (MIoU) of 94.98%, a Mean Pixel Accuracy (MPA) of 97.34%, and an F1-score (F1) of 95.76%, with 4.872 1M parameters (Params) and FLOPs of 26.270 7G. Compared with the DeepLabV3+ model employing Xception as the backbone, the proposed model achieves improvements of 1.23%, 0.76%, 0.70%, and 0.67% in IoU, MIoU, MPA, and F1, respectively, while reducing the number of params from 29.024 3M to 4.872 1M and the FLOPs from 121.136 1G to 26.270 7G. Compared with the DeepLabV3+ model employing MobileNetV2 as the backbone, the proposed model achieves improvements of 3.86%, 2.38%, 1.97%, and 2.14% in IoU, MIoU, MPA, and F1, respectively, while reducing the number of params from 5.813 3M to 4.872 1M, with FLOPs remaining nearly unchanged. Meanwhile, comparative experiments with classical semantic segmentation models were conducted. The results indicate that, considering both accuracy and efficiency, the ITR-DeepLabV3+ model demonstrates superior overall performance compared to conventional network models (Tab.5).

Conclusions Active infrared thermography is employed to capture the coal wall to be cut, a coal-rock dataset is constructed and augmented, an ITR-DeepLabV3+ network model is proposed, and both ablation and comparative experiments are conducted. The experimental results demonstrate that the proposed modules yield significant gains in both accuracy and efficiency. Furthermore, the proposed ITR-DeepLabV3+ model outperforms classical networks, including PSP-Net, U-Net, and HR-Net, thereby enabling stable and reliable identification of the coal-rock interface. In conclusion, the proposed method integrating deep learning and active infrared thermography provides effective technical support for accurate coal-rock interface perception under complex operating conditions and is of significant importance for enhancing the safety and operational efficiency of coal mining.