煤炭是我国主要能源，随着工业互联网、计算机视觉、物联网、信息处理等技术的发展，国家七部委将智能化矿井建设列为国家发展战略，同时也是煤炭工业数字化转型发展的必然选择。透明地质技术保障是以矿产地质信息资料为基础，实现地质信息清晰化，是智能矿山建设的重要组成部分。当前，煤矿钻孔探测视频多采用人工直接观测方法判断钻孔裂隙，存在效率低下、任务量繁重、漏判误判等诸多问题；对于地质钻孔的岩性分析也是通过专业人员对自然伽马曲线判读进行标注划分，专业要求性高。本文对于钻孔探测信息处理做了两部分研究工作：利用深度学习技术实现钻孔裂隙智能检测，借助计算机对自然伽马数据分析并融合深度数据进行地层划分。

本文利用数字图像处理技术对煤巷上顶板5个不同地质钻孔视频生成数据图片进行位置变化、图像增强处理完成样本数据扩充，人工标注了1098张煤矿钻孔裂隙图片作为数据集，采用YOLOv5算法实现钻孔裂隙自动检测。针对误检问题，在YOLOv5模型基础上添加了SENet注意力机制，提升了模型对非裂隙区域的区分能力，使得检测精度较原模型提升了1.8%，平均精度提升了0.9%；针对检测框回归问题，使用有效交并比损失函数替换YOLOv5中完全交并比损失函数，改善了检测框位置精度情况，检测精度较原模型提升了1.3%，平均精度提升了0.4%；针对漏检问题，对YOLOv5模型锚定框参数优化以及添加检测层，提升了模型对小裂隙区域的检测，检测精度较原模型提升了1.3%，平均精度提升了0.2%。最后结合SENet注意力机制、有效交并比损失函数、添加检测层三点共同优化的模型比YOLOv5原模型精度提升了2.1%，召回率提升了1.6%，平均精度提升了1.0%，比SSD算法模型精度提升了18.9%，召回率提升了39.5%，平均精度提升了28.2%。表明基于YOLOv5优化改进模型能够对钻孔裂隙进行很好检测，满足钻孔裂隙检测需求。

针对自然伽马数据在采集过程中受煤矿井下环境以及工程操作影响带来噪声的问题，通过采用高斯滤波算法对自然伽马数据进行滤波平滑处理去噪，利用数理统计极值方差法对煤巷上顶板采集的自然伽马数据划分，划分了煤层、泥岩层、粉砂岩层。在利用深度学习识别岩性方面，通过对5563张煤巷上顶板地层图片数据集训练，得到能够识别煤系地层的YOLOv5算法模型。通过深度数据校准视频数据与自然伽马数据位置，结合视频识别结果与自然伽马数据划分情况综合判断了煤巷上顶板深度为51.12米钻孔煤层、泥岩层、粉砂岩层情况，为煤矿岩性划分提供了一定方法和思路。

Coal is the main energy source in China. With the development of industrial Internet, computer vision, internet of things, information processing and other technologies, seven national ministries and commissions have listed the construction of intelligent mines as a national development strategy, and it is also an inevitable choice for the digital transformation and development of the coal industry. The transparent geological technology guarantee is based on the mineral geological information data and realizes the clear geological information, which is the important part of the intelligent mine. At present, mostly use manual direct observation of coal mine borehole detection video to judge borehole fractures, which has many problems such as low efficiency, heavy workload, missed judgment and misjudgment, etc. The lithology stratification of geological borehole is also marked and divided by professionals on the interpretation of the natural gamma curve, and the professional requirements are high. This paper has done two parts of research work on information processing of borehole detection：uses deep learning technology to realize intelligent detection of borehole fractures, and uses computer to analyze natural gamma data and integrate depth data for stratigraphic division.

In this paper, the digital image processing technology is used to change the position and enhance the image of the video generated data pictures of five different geological boreholes in the upper roof of coal roadway to complete the sample data expansion. Manually annotated 1098 pictures of coal mine boreholes fractures as the dataset, and uses the YOLOv5 algorithm to realize automatic identification of borehole fractures. For the false detection problem, the SENet attention mechanism is added to the YOLOv5 model to improve the ability of the model to distinguish non fractured areas, which improves the detection accuracy by 1.8% compared with the original model, and the average accuracy is increased by 0.9%. For the detection frame regression problem, the effective intersection over union loss function is used to replace the YOLOv5's complete intersection over union loss function, which improves the position accuracy of detection frame. Compared with the complete intersection over union loss function, the detection accuracy is increased by 1.3% and the average accuracy is increased by 0.4%. For the missed detection problem, the YOLOv5 model anchor frame parameters are optimized and the detection layer is added to improve the model's detection of small fractured areas. Compared with the original model, the detection accuracy increased by 1.3%, and the average accuracy increased by 0.2%. Finally, combined with the SENet attention mechanism, the effective intersection over union loss function, and the addition of the detection layer, the three-point joint optimization model improves the accuracy of the original YOLOv5 model by 2.1%, the recall rate increases by 1.6%, and the average accuracy increases by 1.0%. The model precision increased by 18.9%, the recall rate increased by 39.5%, and the average precision increased by 28.2% compared with SSD algorithm model. It shows that the optimized and improved model based on YOLOv5 can well identify borehole fractures and meet the needs of borehole fractures detection.

Aiming at the problem of noise caused by the underground environment of coal mines and engineering operations during the acquisition process of natural gamma data, the Gaussian filtering algorithm is used to filter and smooth the natural gamma data for denoising. The natural gamma data collected from the roof of the coal roadway is divided by the extreme value variance method of mathematical statistics, and the coal seam, mudstone layer and siltstone layer are divided. In terms of using deep learning to identify lithology, the YOLOv5 algorithm model that can identify coal measure strata is obtained by training 5563 pictures of the roof strata of the coal roadway. Through the depth data to calibrate the position of video data and Natural Gamma data, combined with the video recognition results and the division of Natural Gamma data, the situation of borehole coal seam, mudstone layer and silty sand layer with a depth of 51.12m on the upper roof of coal roadway is comprehensively judged, which provides a certain method and idea for the lithology division of coal mine.