煤矿井下钻孔裂隙识别技术

Crack identification technology of borehole in underground coal mine

为研究采区煤岩体中裂隙、断裂、破碎带等结构面的自动检测技术,解决现有人工智能技术中迭代次数大、检测框准确度低等问题,采用YOLOv5算法融合注意力机制、损失函数、多尺度检测的方法,对煤矿巷道上顶板5种不同地质钻孔进行裂隙检测试验。结果表明:将注意力机制SENet引入YOLOv5模型框架,避免了图像背景区域与裂隙区域相似度较高问题;采用有效交并比损失函数代替完全交并比损失函数,使得预测框能够更加有效拟合真实目标框;对YOLOv5模型增添3种不同尺寸的锚定框并添加160×160特征层,实现检测更小的目标。该方法与SSD、YOLOv5等检测算法在同样条件下相比,其检测精度分别提升了18.9%,2.1%,召回率提升了39.5%,1.6%,平均精度提升了28.1%,1.0%。改进后的模型将三尺度检测变为四尺度检测,提升了算法的多尺度目标检测性能,能够对钻孔裂隙进行高精度检测,满足钻孔裂隙实时检测需求。

In order to study the automatic detection technology of such structural planes as crack,fracture and fracture zone in coal rock mass in mining area,and solve the problems of large iteration times and low accuracy of detection frame in existing artificial intelligence technology,the YOLOv5 algorithm combining attention mechanism,loss function and multi-scale detection methods was used to conduct crack detection tests on 5 different geological boreholes in the roof of coal mine roadway.The results show that the attention mechanism SENet is introduced into the YOLOv5 model framework to avoid the problem of high similarity between the background region and the crack region,the effective crossover loss function is used to replace the complete crossover loss function,which makes the prediction frame fit the real target frame more effectively,and three different sizes of anchor frames are added to the YOLOv5 model and 160×160 feature layers are added to achieve smaller detection targets.Compared with detection algorithms such as SSD and YOLOv5 under the same conditions,the detection accuracy of this method is increased by 18.9%and 2.1%,the recall rate is increased by 39.5%and 1.6%,and the average accuracy is increased by 28.1%and 1.0%,respectively.The improved model changes the three-scale detection into four-scale detection,improves the multi-scale target detection performance of the algorithm,and can detect borehole cracks with high precision,meeting the real-time detection requirements of borehole cracks.