基于GSL−YOLO模型的综放工作面混矸率检测方法

王福奇; 王志峰; 金建成; 井庆贺; 王耀辉; 王大龙; 汪义龙

doi:10.13272/j.issn.1671-251x.2024080011

基于GSL−YOLO模型的综放工作面混矸率检测方法

doi: 10.13272/j.issn.1671-251x.2024080011

1.
华能庆阳煤电有限责任公司，甘肃庆阳　745000
2.
中国矿业大学（北京）能源与矿业学院，北京　100083
3.
华亭煤业集团有限责任公司，甘肃平凉　744100
4.
扎赉诺尔煤业有限责任公司，内蒙古满洲里　021410
5.
华能煤炭技术研究有限公司，北京　100071
6.
华能云南滇东能源有限责任公司，云南曲靖　655500
7.
华能煤业有限公司陕西矿业分公司，陕西西安　710001

基金项目: 国家自然科学基金面上项目（52274207）；天地科技开采设计事业部科技创新基金项目（KJ-2021-KCMS-02）。

详细信息

作者简介:
王福奇（1978—），男，甘肃华亭人，高级工程师，研究方向为煤炭开采，E-mail：11185351260@qq.com

中图分类号: TD823.49
计量
- 文章访问数: 83
- HTML全文浏览量: 20
- PDF下载量: 19
- 被引次数: 0
出版历程
- 收稿日期: 2024-08-05
- 修回日期: 2024-09-22
- 网络出版日期: 2024-09-29

Detection method for gangue mixed ratio in fully mechanized caving faces based on the GSL-YOLO model

1.
Huaneng Qingyang Coal Power Co., Ltd., Qingyang 745000, China
2.
School of Energy and Mining Engineering, China University of Mining & Technology-Beijing, Beijing 100083, China
3.
Huating Coal Industry Group Co., Ltd., Pingliang 744100, China
4.
Zhalainuo'er Coal Industry Co., Ltd., Manzhouli 021410, China
5.
Huaneng Coal Technology Research Co., Ltd., Beijing 100071, China
6.
Huaneng Yunnan Diandong Energy Co., Ltd., Qujing 655500, China
7.
Shaanxi Mining Branch, Huaneng Coal Industry Co., Ltd., Xi'an 710001, China

摘要

摘要: 针对现有基于深度学习的综放工作面混矸率检测方法在井下低照度、高粉尘、煤矸堆叠等复杂条件下存在煤矸识别精度低、分割效果差、模型参数量和运算量大、未实现混矸率的实时检测等问题，提出了一种基于GSL−YOLO模型的混矸率检测方法。GSL−YOLO模型在YOLOv8−seg的基础上进行以下改进：在主干网络中引入全局注意力机制（GAM），通过减少信息弥散和放大全局交互表示提高模型特征提取能力；选用具有高效局部聚合网络的空间金字塔池化（SPPELAN）模块，提升模型处理不同尺寸目标时的检测性能；采用轻量级非对称多级压缩检测头（LADH），降低模型的训练难度，同时提高推理速度。提出了一种基于类别分割掩码的混矸率计算方法，该方法基于煤流图像处理结果中的分割掩码信息，计算其中矸石的像素面积与总像素面积的比值，作为瞬时混矸率。实验结果表明：① GSL−YOLO模型的mAP@0.5∶0.95达96.1%，比YOLOv8−seg模型提高了0.8%。② GSL−YOLO模型的参数量为2.9×10⁶个，浮点运算次数为11.4×10⁹，模型权重为6.0 MiB，比YOLOv8−seg模型分别降低了12.1%，5.8%，11.8%，实现了模型的轻量化。③ GSL−YOLO模型在测试集上的帧率为12帧/s，基本满足实时检测要求。④ 与YOLO系列模型相比，GSL−YOLO模型分割效果最好，检测精度最高，参数量和运算量较少，综合性能最佳。⑤ 基于截取的综放工作面后部刮板输送机上煤流视频中的3帧图像，计算了瞬时混矸率，结果表明，提出的混矸率计算方法基本实现了综放工作面混矸率的实时计算。
- 智能放煤 /
- 煤矸识别 /
- 混矸率检测 /
- YOLOv8−seg /
- 图像分割 /
- 全局注意力机制 /
- 非对称检测头
Abstract: Aiming to address the issues with current gangue mixed ratio detection methods in fully mechanized caving face based on deep learning, such as low accuracy of coal gangue identification, poor segmentation performance, large model parameters and computation load, and the inability to achieve real-time detection of gangue mixed ratio under complex conditions such as low lighting, high dust, and coal and angue stacking, the paper proposed a gangue mixed ratio detection method based on the GSL-YOLO model. The GSL-YOLO model introduced the following improvements to the YOLOv8-seg model: the incorporation of a global attention mechanism (GAM) in the backbone network to enhance feature extraction by reducing information dispersion and amplifying global interaction representation; the use of a spatial pyramid pooling with efficient local aggregation network (SPPELAN) module to improve detection performance for targets of varying sizes; and the adoption of a lightweight asymmetric dual-head (LADH) to reduce training difficulty while increasing inference speed. Additionally, a gangue mixed ratio calculation method based on category segmentation masks was proposed, which calculated the ratio of the pixel area of gangue to the total pixel area in the segmentation mask of coal flow images, serving as the instantaneous gangue mixed ratio. Experimental results showed that: ① The GSL-YOLO model achieved an mAP@0.5∶0.95 of 96.1%, which was 0.8% higher than the YOLOv8-seg model. ② The GSL-YOLO model had 2.9×10⁶ parameters, 11.4×10⁹ floating-point operations, and a model weight of 6.0 MiB, representing reductions of 12.1%, 5.8%, and 11.8% respectively compared to the YOLOv8-seg model, achieving model lightweighting. ③ The GSL-YOLO model achieved a frame rate of 12 frames per second on the test set, essentially meeting the requirements for real-time detection. ④ Compared with the YOLO series models, the GSL-YOLO model had the best segmentation effect, the highest detection accuracy, fewer parameters and computation load, and the best overall performance. ⑤ Based on three frames of images captured from the coal flow on the rear scraper conveyor of the fully mechanized caving face, the instantaneous gangue mixed ratio was calculated, and the results showed that the proposed method successfully realized real-time calculation of the gangue mixed ratio in fully mechanized caving face.
- intelligent coal drawing /
- coal and gangue identification /
- gangue mixed ratio detection /
- YOLOv8−seg /
- image segmentation /
- global attention mechanism /
- asymmetric detection head

HTML全文

图 1 GSL−YOLO模型结构

Figure 1. GSL-YOLO model structure

下载: 全尺寸图片幻灯片

图 2 GAM结构

Figure 2. Global attention mechanism（GAM） structure

下载: 全尺寸图片幻灯片

图 3 通道注意力子模块

Figure 3. Channel attention sub-module

下载: 全尺寸图片幻灯片

图 4 空间注意力子模块

Figure 4. Spatial attention sub-module

下载: 全尺寸图片幻灯片

图 5 SPPELAN模块

Figure 5. Spatial pyramid pooling with efficient local aggregation network （SPPELAN） module

下载: 全尺寸图片幻灯片

图 6 LADH结构

Figure 6. Lightweight asymmetric dual-head （LADH） structure

下载: 全尺寸图片幻灯片

图 7 模型改进前后热力图对比

Figure 7. Comparison of heat maps before and after model improvement

下载: 全尺寸图片幻灯片

图 8 煤矸图像增强效果

Figure 8. Images enhancement effect of coal and gangue

下载: 全尺寸图片幻灯片

图 9 模型分割效果对比

Figure 9. Comparison of model segmentation effects

下载: 全尺寸图片幻灯片

表 1 消融实验结果

Table 1. Ablation experimental results

YOLOv8− seg	GAM	SPPELAN	LADH	mAP@ 0.5∶0.95/%	参数量/ 10⁶个	浮点运算次数/10⁹	模型权重/ MiB
√	×	×	×	95.3	3.3	12.1	6.8
√	√	×	×	95.7	3.7	12.5	7.7
√	×	√	×	95.5	2.7	11.6	5.7
√	×	×	√	96.1	3.3	11.7	6.9
√	√	√	×	95.8	2.8	11.7	5.9
√	√	×	√	96.3	3.7	12.1	7.8
√	×	√	√	95.6	2.8	11.3	5.8
√	√	√	√	96.1	2.9	11.4	6.0

下载: 导出CSV

表 2 对比实验结果

Table 2. Comparative experimental results

模型	mAP@ 0.5∶0.95/%	参数量/10⁶个	浮点运算次数/10⁹	模型权重/MiB
YOLOv8−seg	95.3	3.3	12.1	6.8
YOLOv3−tiny−seg	87.1	14.1	32.8	28.3
YOLOv5−seg	95.0	2.8	11.1	5.8
YOLOv6−seg	95.1	4.4	15.3	9.0
GSL−YOLO	96.1	2.9	11.4	6.0

下载: 导出CSV

表 3 混矸率计算结果

Table 3. Calculation results of gangue mixed ratio

图像	原图	分割结果	混矸率/%	帧率/ （帧·s⁻¹）
图像1			7	5.0
图像2			25	4.9
图像3			45	4.8

下载: 导出CSV

参考文献(21)

[1]	王国法,庞义辉,任怀伟,等. 智慧矿山系统工程及关键技术研究与实践[J]. 煤炭学报,2024,49(1):181-202. WANG Guofa,PANG Yihui,REN Huaiwei,et al. System engineering and key technologies research and practice of smart mine[J]. Journal of China Coal Society,2024,49(1):181-202.
[2]	王国法,孟令宇. 煤矿智能化及其技术装备发展[J]. 中国煤炭,2023,49(7):1-13. WANG Guofa,MENG Lingyu. Development of coal mine intelligence and its technical equipment[J]. China Coal,2023,49(7):1-13.
[3]	王国法,庞义辉,许永祥,等. 厚煤层智能绿色高效开采技术与装备研发进展[J]. 采矿与安全工程学报,2023,40(5):882-893. WANG Guofa,PANG Yihui,XU Yongxiang,et al. Development of intelligent green and efficient mining technology and equipment for thick coal seam[J]. Journal of Mining & Safety Engineering,2023,40(5):882-893.
[4]	王家臣. 我国放顶煤开采的工程实践与理论进展[J]. 煤炭学报,2018,43(1):43-51. WANG Jiachen. Engineering practice and theoretical progress of top-coal caving mining technology in China[J]. Journal of China Coal Society,2018,43(1):43-51.
[5]	王家臣. 我国综放开采40年及展望[J]. 煤炭学报,2023,48(1):83-99. WANG Jiachen. 40 years development and prospect of longwall top coal caving in China[J]. Journal of China Coal Society,2023,48(1):83-99.
[6]	王家臣,张锦旺. 综放开采顶煤放出规律的BBR研究[J]. 煤炭学报,2015,40(3):487-493. WANG Jiachen,ZHANG Jinwang. BBR study of top-coal drawing law in longwall top-coal caving mining[J]. Journal of China Coal Society,2015,40(3):487-493.
[7]	李良晖. 放顶煤工作面煤矸混合度自动识别研究进展[J]. 煤炭工程,2017,49(10):30-34. LI Lianghui. Research progress of automatic recognition of coal-gangue mixedness in longwall top-coal caving face[J]. Coal Engineering,2017,49(10):30-34.
[8]	李嘉豪,司垒,王忠宾,等. 综放工作面煤矸识别技术及其应用[J]. 仪器仪表学报,2024,45(1):1-15. LI Jiahao,SI Lei,WANG Zhongbin,et al. Coal gangue identification technology and its application in fully-mechanized coal mining face[J]. Chinese Journal of Scientific Instrument,2024,45(1):1-15.
[9]	李德永,王国法,郭永存,等. 基于CFS−YOLO算法的复杂工况环境下煤矸图像识别方法[J]. 煤炭科学技术,2024,52(6):226-237. LI Deyong,WANG Guofa,GUO Yongcun,et al. Image recognition method of coal gangue in complex working conditions based on CFS-YOLO algorithm[J]. Coal Science and Technology,2024,52(6):226-237.
[10]	何凯,程刚,王希,等. 基于CED−YOLOv5s模型的煤矸识别方法研究[J]. 工矿自动化,2024,50(2):49-56,82. HE Kai,CHENG Gang,WANG Xi,et al. Research on coal gangue recognition method based on CED-YOLOv5s model[J]. Journal of Mine Automation,2024,50(2):49-56,82.
[11]	滕文想,王成,费树辉. 基于HGTC−YOLOv8n模型的煤矸识别算法研究[J]. 工矿自动化,2024,50(5):52-59. TENG Wenxiang,WANG Cheng,FEI Shuhui. Research on coal gangue recognition algorithm based on HGTC-YOLOv8n model[J]. Journal of Mine Automation,2024,50(5):52-59.
[12]	杨洋,李海雄,胡淼龙,等. 基于YOLOv5−SEDC模型的煤矸分割识别方法[J]. 工矿自动化,2024,50(8):120-126. YANG Yang,LI Haixiong,HU Miaolong,et al. Coal and gangue segmentation and recognition method based on YOLOv5-SEDC model[J]. Journal of Mine Automation,2024,50(8):120-126.
[13]	王家臣,潘卫东,张国英,等. 图像识别智能放煤技术原理与应用[J]. 煤炭学报,2022,47(1):87-101. WANG Jiachen,PAN Weidong,ZHANG Guoying,et al. Principles and applications of image-based recognition of withdrawn coal and intelligent control of draw opening in longwall top coal caving face[J]. Journal of China Coal Society,2022,47(1):87-101.
[14]	王家臣,李良晖,杨胜利. 不同照度下煤矸图像灰度及纹理特征提取的实验研究[J]. 煤炭学报,2018,43(11):3051-3061. WANG Jiachen,LI Lianghui,YANG Shengli. Experimental study on gray and texture features extraction of coal and gangue image under different illuminance[J]. Journal of China Coal Society,2018,43(11):3051-3061.
[15]	贺海涛,王佳豪,张海峰,等. 基于U−Net的放煤状态控制关键技术研究[J]. 煤炭科学技术,2022,50(增刊2):237-243. HE Haitao,WANG Jiahao,ZHANG Haifeng,et al. Calculation method of gangue content of coal gangue mixed image in fully-mechanized caving based on U-Net[J]. Coal Science and Technology,2022,50(S2):237-243.
[16]	单鹏飞,孙浩强,来兴平,等. 基于改进Faster R−CNN的综放煤矸混合放出状态识别方法[J]. 煤炭学报,2022,47(3):1382-1394. SHAN Pengfei,SUN Haoqiang,LAI Xingping,et al. Identification method on mixed and release state of coal-gangue masses of fully mechanized caving based on improved Faster R-CNN[J]. Journal of China Coal Society,2022,47(3):1382-1394.
[17]	REDMON J,DIVVALA S,GIRSHICK R,et al. You only look once:unified,real-time object detection[C]. IEEE Conference on Computer Vision and Pattern Recognition,Las Vegas,2016:779-788.
[18]	油亚鹏,马波,赵乐,等. 基于CA−YOLOv8的输送带大块煤检测方法 [J/OL]. 计算机辅助设计与图形学学报:1-12[2024-07-06]. http://kns.cnki.net/kcms/detail/11.2925.TP.20240204.1655.057.html. YOU Yapeng,MA Bo,ZHAO Le,et al. Large coal detection method for conveyor belt based on CA-YOLOv8[J/OL] Journal of Computer- Aided Design and Graphics:1-12[2024-07-06]. http://kns.cnki.net/kcms/detail/11.2925.TP.20240204.1655.057.html.
[19]	LIU Yichao,SHAO Zongru,HOFFMANN N. Global attention mechanism:retain information to enhance channel-spaial interactions[EB/OL]. [2024-06-20]. http:// arxiv.org/pdf/2112.05561.
[20]	WANG C Y,BOCHKOVSKIY A,LIAO H Y M. YOLOv7:trainable bag-of-freebies sets new state-of-the-art for real-time object detectors[C]. IEEE/CVF Conference on Computer Vision and Pattern Recognition,Vancouver,2023:7464-7475.
[21]	杨志渊,罗亮,吴天阳,等. 改进YOLOv8的轻量级光学遥感图像船舶目标检测算法[J]. 计算机工程与应用,2024,60(16):248-257. YANG Zhiyuan,LUO Liang,WU Tianyang,et al. Improved lightweight ship target detection algorithm for optical remote sensing images with YOLOv8[J]. Computer Engineering and Applications,2024,60(16):248-257.