|
| Research on intelligent detection method of coal mine blast hole based on improved YOLOv7 |
|
Received:May 09, 2025
Revised:June 09, 2025
Accepted:June 10, 2025
Published Online:August 05, 2025
|
| View Full Text View/Add Comment Download reader |
| DOI： |
| KeyWord:intelligent identification of blast holes; improved YOLOv7 algorithm; deep learning; intelligent mining construction; small object detection |
| Author | Institution |
| Wang Maike |
School of Mechanics and Civil Engineering, China University of Mining and Technology Beijing |
| Jin Qingyu |
School of Mechanics and Civil Engineering, China University of Mining and Technology Beijing |
|
| Hits: 2013 |
| Download times: 314 |
| Abstract: |
| In the intelligent construction process of the mine drilling and blasting method, the accuracy of the mechanical arm aligning with the blast holes during the loading process affects the operation efficiency and the safety of the project. Currently, mechanical loading generally uses deep learning algorithms to optimize the positioning of blast holes. This paper reviews the application of representative target detection algorithms in China and points out that the existing blast hole detection methods have problems such as high missed detection rate and weak anti-interference ability. To optimize the algorithm and improve the model effect, the SSPD-YOLOv7 algorithm model, which is more suitable for mine blast hole recognition, is proposed by introducing the parameter-free SimAM attention mechanism, the SPD-Conv structure, and the K-means++ prior box optimization. Firstly, the parameter-free SimAM attention mechanism is added to the backbone network to generate 3D weights in the spatial dimension, enhancing the features of small blast hole targets and reducing the interference of the background where the blast holes are located. Secondly, the SPD-Conv layer is added after the traditional module pooling convolution to solve the problems of fine-grained loss and low feature learning efficiency caused by the stride convolution and pooling operations. Finally, the K-means++ clustering algorithm is used to re-determine the anchor boxes, making them more suitable for the size of the target boxes in the self-made dataset compared to the original YOLOv7, thereby enhancing the robustness of the model. Finally, through setting ablation experiments and comparison experiments, four different deep learning algorithms are compared and analyzed. Among them, YOLOx, due to the removal of anchor boxes, leads to a decrease in the detection ability of small targets, manifested as a decrease in recall rate and insufficient positioning accuracy; YOLOv5 is limited by the FPN and PANet structure, and its feature fusion mechanism is relatively simple, which easily leads to the loss of small target features in the complex mine environment; although YOLOv7 achieves a better balance between detection accuracy and speed, it still has the problem of insufficient capture of small target features. The experimental results show that the improved SSPD-YOLOv7 algorithm model has an average precision (AP) increased from 85.02% to 87.88%, a precision (Precision) increased from 83.29% to 88.70%, an increase of 5.41%, and a recall rate (Recall) slightly decreased, with the F1 score increased from 0.82 to 0.84. It can still meet the real-time recognition accuracy requirements of blast holes and successfully improve the detection ability of small-sized blast hole targets in complex environments such as low light and broken rock masses. |
| Close |
|
|
|