Coal-rock interface perception and accurate recognition in heading face under coal dust environment based on machine vision

Abstract

The coal-rock identification technology in the roadway excavation process is the core of the automatic adjustment of roadheader’s cutting head, and it is also one of the key problems restricting the development of intelligent mines. In view of the current mining imbalance, the excavation face lacks a mature and effective coal-rock identification scheme, and the existing image based coal-rock identification models have problems such as poor segmentation accuracy and inability to flexibly deploy, a coal-rock cutting interface perception and precise recognition method based on image segmentation is proposed in the heading face. This method combines the actual cutting situation of the excavation working face and uses the MobileNetV2 feature extraction network as the backbone network of DeepLabV3+, so that the model can better balance the segmentation accuracy and model complexity. The channel attention (SE) operation is performed on the advanced features output by the Atrous Spatial Pyramid Pooling module, and channel weights are assigned to strengthen the training of key feature information. The channel spatial attention (CA) mechanism is introduced into the shallow features output by the backbone network to weight the low-level representation information in the shallow feature map, thus designing a coal-rock cutting interface identification model that integrates the double attention mechanism in DeepLabV3+. At the same time, an experimental platform for coal-rock identification in a dusty environment is built to simulate the coal and rock cutting surface formed by the roadheader after cutting, and the coal-rock cutting interface acquisition system in the process of roadway excavation is developed. Taking the actual mine excavating face as the engineering background, the recognition accuracy and practical applicability of the coal-rock identification model are verified. The research results show that the average intersection ratio and average pixel accuracy of the SE-CA-DeepLabV3+ network model are 97.15% and 98.51%, respectively, which have better segmentation performance than other network models. The established model is used to verify the original coal and rock images from the heading face of the experimental mine in northern Shaanxi, the average error is 0.7%, and the number of transmission frames per second is 43fps, which meets the deployment conditions of downhole field applications.

Keywords

• coal-rock cutting interface
• se-ca-deeplabv3+
• roadway excavation
• coal-rock identification platform
• image segmentation