巷道顶板岩层裂隙分布特征是影响安全开采的重要因素,目前钻孔成像法是裂隙探测的主要方法,但钻孔成像法缺乏裂隙智能识别系统,传统人工观察法速度慢,只能定性分析裂隙特征。针对上述问题,在多个矿区进行钻孔成像,形成丰富的岩层裂隙识别数据库,并将数据库分为训练集和测试集。基于卷积神经网络搭建结构面识别网络,网络(1)训练时的像素类别分为背景像素、结构面像素和干扰像素;网络(2)训练时的像素类型分为背景像素和结构面像素。提出裂隙区划分方法,设计裂隙区内结构面数量自动计算程序,优化传统的霍夫变换算法,可实现裂隙区内多个结构面准确识别,并开发结构面参数自动输出程序。研究结果表明:将干扰像素划分为一种类别可以有效改善结构面识别效果。网络(1)对结构面的识别率为76.70%,通过开发的程序使得识别率达到90.20%。对于结构面的起点和终点位置预测误差主要集中在5mm内,倾向预测误差主要集中于20°范围内,倾角预测误差主要集中于10°范围内,裂隙宽度的平均误差为5.07mm。该研究工作促使钻孔成像技术由人工观察的定性分析向智能化的定量分析转变,为煤矿智能化发展提供了重要的地质保障技术。

The distribution characteristics of fractures in the rock stratum on the roadway roof are an im- portant factor affecting safe mining. At present, the primary method for detecting fractures is the borehole imaging technique which lacks an intelligent system for identifying fractures, and the conven- tional manual observation approach is slow and limited to qualitative analysis of fracture attributes. To solve problems, borehole imaging has been carried out in many mining areas, and a rich database of fracture identification has been formed. The database is divided into training set and test set, and a struc- tural plane recognition network is developed based on convolutional neural network. The pixel categories during network (1) training are divided into background pixels, structural plane pixels and interferingpixels. The pixel categories during network (2) training are divided into background pixels and structural plane pixels. The division method of the fracture area is proposed, and an automatic calculation program for the number of structural planes in the fracture area is devised The traditional Hough transform algo- rithm is optimized to achieve accurate identification of multiple fractures in the fracture area. The post- processing program for intelligent image recognition of structural planes is developed. The research results show that when the interfering pixels divided into one category, the identification of structural planes can be improved effectively. The recognition rate of structural planes by the network during trai- ning was 76.70% , which is improved to 90.20% with the implemented post-processing program. For structural planes, the start and end location errors mainly fall within 5 mm, while the dip direction error is mainly concentrated in the range of 20°, and the dip angle error is mainly concentrated in the range of 10°, and the average error of the fracture width is 5.07 mm. This research promotes the transformation of borehole imaging technology from qualitative manual analysis to intelligent quantitative analysis and provides an important geological guarantee technology for the intelligent development of coal mines.