基于TAUNet分割模型的爆堆块度空间分布研究

Spatial Distribution of Blast Reactor Block Based on TAUNet Segmentation Model

为了更好地满足矿区现场爆堆块度实时高精度检测的需要,提出了一种基于深度学习的爆堆块体分割模型TAUNet(Transformer Aspp UNet),该模型在UNet的编码器和解码器中融入Transformer的自我注意机制,利用其处理大型特征映射,改善全局信息的提取,恢复在编码器中跳过的粒度细节。在骨干网络特征提取阶段,加入了ASPP空洞卷积模块,增强了模型对块体局部特征的融合。在爆堆图像分割的基础上,采用爆堆分层的方法获取爆堆的块度空间分布信息。结果表明:(1)TAUNet分割模型具有良好分割性能,模型训练评价指标骰子系数、交并比、召回率分别达到97.12%、94.61%、96.2%,均优于主流的语义分割模型,对现场爆堆块体有着良好的分割效果;(2)通过爆堆分层的方法可知肇庆某矿山西采区315~300 m平台的爆堆块度空间分布,87.15%岩块粒径分布在0~0.6 m, 9.9%的岩块粒径分布于0.6~1.0 m,大于1.0 m的大块占2.95%。研究结果能够为爆破效果评价的精细化、智能化发展提供参考借鉴。

In order to better meet the need for real-time and high-precision detection of blast reactor block in mining sites, a blast reactor block segmentation model TAUNet(Transformer Aspp UNet) based on deep learning was proposed. The model integrated Transformer's self-attentive mechanism in the encoder and decoder of UNet, used it to handle large feature mappings, improved the extraction of global information and restored the granularity details skipped in the encoder. In the backbone network feature extraction stage, the ASPP null convolution module was incorporated to enhance the model for block local feature fusion. On the basis of the blast reactor image segmentation, the spatial distribution information of the blast reactor block was obtained by the method of layering blast reactor. The results show that the TAUNet segmentation model has a good segmentation performance, and the model training evaluation indexes, including dice coefficient, intersection over union, and recall rate reach 97.12%、94.61% and 96.2% respectively, which are all better than the mainstream semantic segmentation model. And the TAUNet model have good segmentation effects on the on-site blast reactor blocks. Through the method of layering blast reactor, the spatial distribution of blast reactor block of 315-300 m platform in the west mining area of a mine in Zhaoqing City is determined as 87.15% of the block size distribution in the 0-0.6 m, 9.9% of the block size distribution in the 0.6-1.0 m, and large blocks with size greater than 1.0 m accounted for 2.95%. The research results can provide a reference for the refined and intelligent development of blasting effect evaluation.