基于改进U-net网络模型的综采工作面煤岩识别方法

Coal-rock recognition method of fully-mechanized coal mining face based on improved U-net network model

煤岩识别是实现工作面智能化开采的核心技术,也是煤炭开采领域的技术难题。针对当前综采工作面煤岩识别精度低的问题,提出了一种基于改进U-net网络模型的煤岩图像识别方法。该方法通过使用深度分离卷积代替传统卷积减少了网络模型的参数,提高了语义分割的效率;添加Res2net模块来提高编码器提取特征的能力,同时加入条件随机场对分割图像进行后处理,提高了网络模型在分割煤岩图像交界区域的精确性。为了获取更加丰富的煤岩分布图像,研制了不同特性的煤岩试样,搭建了采煤机煤岩截割试验台。通过煤岩截割试验获取了煤岩分布图像数据,并对其进行切分、缩放、旋转、裁剪、加噪声等操作,生成了包含8000个样本的煤岩图像语义分割数据集,采用自适应学习算法对模型进行训练,给出了模型训练过程中准确率和损失函数的变化规律。选取像素准确度和交并比对语义分割结果进行评估,结果表明,改进U-net网络模型的像素准确度和交并比的平均值分别为95.81%和91.13%,所占内存为35 M,测试用时为36.45 ms/张,与其他网络模型相比,该方法在煤岩图像分割中具有明显的优越性。在井下现场试验中,通过构建综采工作面煤岩图像语义分割数据集对改进U-net网络模型进行训练和测试,最后实现了综采工作面的煤岩识别,验证了该方法的可行性和实用性。

Coal-rock identification is the core technology to realize an intelligent coal mining,which has become a technical problem in the field of coal mining.In view of the low accuracy of coal-rock identification in the fully mechanized face,a coal-rock image recognition method based on improved U-net network model is proposed.This method reduces the parameters of the network model by using deep separation convolution instead of traditional convolution,and improves the efficiency of semantic segmentation.The Res2net module is also used to improve the ability of encoder to extract features.At the same time,a conditional random field is added to post-process the segmented image to improve the accuracy of the network model in the segmentation area.In order to obtain more abundant coal-rock distribution images,coal-rock samples with different characteristics are developed,and the coal-rock cutting experimental platform of shearer is built.The coal-rock images are obtained through the coal-rock cutting experiments,and some operations such as segmentation,scale,rotation,shear and adding noise are simultaneously performed to produce a rich data set for coal-rock image semantic segmentation,containing 8000 images.The adaptive learning algorithm is used to train the model,and the change rules of accuracy and loss function in the process of model training are given.The pixel accuracy and intersection over union are selected to evaluate the semantic segmentation results.The results show that the pixel accuracy and intersection over union of improved U-net network model are 95.81%and 91.13%,respectively.The model only occupies 35 M of memory,and the test time is 36.45 ms/sheet which is superior to other network models in the aspect of coal-rock segmentation.In the underground field experiment,the improved U-net network model is trained and tested by constructing the semantic segmentation data set of coal and rock images collected from the fully mechanized coal mining face.Finally,the coal-rock identification of fully mechanized mining face is realized,which verifies the feasibility and practicability of this proposed method.