基于双层路由注意力机制的煤粒粒度定量分析

Quantitative analysis of coal particle size based on bi-level routing attention mechanism

  • 摘要: 煤粒粒度分布特征与煤中甲烷气体传播规律的分析密切相关。目前,基于图像分割的煤粒粒度分析方法已成为获取煤粒粒度的主流方案之一,但存在上下文信息丢失、煤粒特征融合不当造成煤粒漏分割和过分割等问题。针对上述问题,设计了一种基于双层路由注意力机制(BRA)的煤粒粒度分析模型。在残差U型网络ResNet−UNet中嵌入BRA模块,得到B−ResUNet网络模型:为减少在煤粒分割过程中出现的漏分割问题,在ResNet−UNet网络的上采样前添加BRA模块,使网络根据上一层的特征调整当前特征层的重要性,增强特征的表达能力,提高长距离信息的传递能力;为减少在煤粒分割过程中出现的过分割问题,在ResNet−UNet网络的特征拼接模块后添加BRA模块,通过动态选择和聚合重要特征,实现更有效的特征融合。对分割出的煤粒进行特征信息提取,针对实验分析中采用的煤粒数据集的煤粒粒度与细胞大小相当,为精确表征煤粒粒度,采用等效圆粒径获取煤粒粒度及粒度分布。实验结果表明:① B−ResUNet网络模型的准确率、平均交并比、召回率较ResNet−UNet基础网络分别提高了0.6%,14.3%,35.9%,准确率达99.6%,平均交并比达92.6%,召回率达94.4%,B−ResUNet网络模型在煤样中具有较好的分割效果,能够检测出较为完整的颗粒结构。② 在上采样前和特征拼接后均引入BRA模块时,网络对煤粒的边缘区域给予了足够的关注,且对一些不太重要的区域减少了关注度,从而提高了网络的计算效率。③ 煤粒的粒度大小在1~2 mm内呈相对均衡的分布趋势,粒度在1~2 mm内的煤粒占比最大为99.04%,最小为90.59%,表明基于BRA的图像处理方法在粒度分析方面具有较高的准确性。

     

    Abstract: The distribution features of coal particle size are closely related to the analysis of methane gas propagation in coal. At present, the coal particle size analysis method based on image segmentation has become one of the mainstream solutions to obtain coal particle size. But there are problems such as loss of contextual information, improper fusion of coal particle features resulting in missed segmentation and over-segmentation of coal particles. In order to solve the above problems, a coal particle size analysis model based on bi-level routing attention (BRA) is designed. The BRA module is embedded in the residual U-shaped network ResNet-UNet to obtain the B-ResUNet network model. To reduce the problem of missed segmentation in coal particle segmentation, a BRA module is added before upsampling in the ResNet-UNet network. It allows the network to adjust the importance of the current feature layer based on the features of the previous layer, enhance the expression capability of features, and improve the transmission capability of long-distance information. To reduce the problem of over segmentation in coal particle segmentation, a BRA module is added after the feature concatenation module of the ResNet-UNet network. By dynamically selecting and aggregating important features, more effective feature fusion is achieved. The feature information from the segmented coal particles is extracted. The coal particle size of the coal particle dataset used in the experimental analysis is equivalent to the cell size. In order to accurately characterize the coal particle size, equivalent circular particle size is used to obtain the coal particle size and size distribution. The experimental results show the following points. ① The accuracy, average intersection to union ratio, and recall of the B-ResUNet network model have been improved by 06.%, 14.3%, and 35.9% compared to the ResNet-UNet basic network, with an accuracy of 99.6%, an average intersection to union ratio of 92.6%, and a recall of 94.4%. The B-ResUNet network model has good segmentation performance in coal samples and can detect relatively complete particle structures. ② When the BRA module is introduced before upsampling and after feature concatenation, the network pays sufficient attention to the edge areas of coal particles and reduces attention to some less important areas, thereby improving the computational efficiency of the network. ③ The particle size of coal particles shows a relatively balanced distribution trend within 1-2 mm, with the maximum proportion of coal particles within 1-2 mm being 99.04% and the minimum being 90.59%. It indicates that the image processing method based on BRA has high accuracy in particle size analysis.

     

/

返回文章
返回