Research On A Deep Learning-based Method For Identifying Roadway Roof Pallets

With the development of artificial intelligence technology,machine vision technology is widely used in intelligent applications in coal mines.As one of the important components in the roadway roof support operation,automatic equipment positioning,roadway deformation monitoring and anchor rod abnormality monitoring can be realized by identifying the characteristics of pallets.Among them,whether the pallet can be accurately identified is the key to the reliability of such applications.This paper optimizes the target detection network from two aspects of pallet identification accuracy and identification speed,and proposes a pallet identification method for roadway roof in underground coal mine based on improved Faster R-CNN.The main work is as follows:(1)The basic theories related to pallet identification,target detection and network lightweighting methods are studied,and the current research status of pallet identification,target detection methods and network lightweighting methods is summarized.The method of detecting the position of roadheading machine based on top plate pallet is introduced,and then the importance of pallet identification is induced.Next,the important and difficult points of pallet recognition are analyzed,the relevant theoretical knowledge involved in this topic is introduced,and finally the overall scheme of pallet recognition is designed.(2)In order to improve the recognition accuracy of pallets,three improvements are made to Faster R-CNN algorithm.First,Res Net50,which has better feature extraction effect,is used to replace VGG16 as the backbone network.Secondly,the size of anchor frame in regional suggestion network is optimized based on K-means++clustering algorithm.Finally,the attention mechanism module is introduced into the feature extraction network Res Net50 to enhance the network’s attention to the tray.(3)The Faster R-CNN algorithm has high detection accuracy but slow detection speed,so it needs to be studied in a lightweight way to improve the efficiency of pallet recognition.A lightweight feature extraction network is constructed by selecting Res Net18 as the backbone network and introducing depth-separable convolution.To achieve a balance between high accuracy and lightweight of the model,this paper adopts a knowledge distillation algorithm to distill the teacher network with high accuracy of tray recognition and instruct the student network with relatively simple structure,aiming to improve the speed of tray recognition while maintaining the accuracy of the model recognition.(4)The image was collected from the simulated roadway,and the target in the image was divided into two categories: complete pallet and incomplete pallet.The target position was marked,and the target was divided into training set,verification set and test set,and the establishment of the pallet image database was completed.Finally,the improved model is experimentally validated on the Pytorch deep learning framework,and the experimental results show that: after the improvement in(3),the average mean accuracy(m AP)of the improved Faster R-CNN network model for pallet recognition is improved by 7.59% compared with the original Faster R-CNN network;after the improvement in(4),the model occupies 59.47% less space and The m AP value is only lost by 3.58%,and the improved algorithm greatly improves the speed of pallet recognition while basically maintaining the original recognition accuracy.This paper researches the deep learning based coal mine underground roadway roof pallet identification method,and improves the algorithm accordingly in order to enhance the accuracy and speed of pallet identification,which realizes the high accuracy and high efficiency of coal mine underground roadway roof pallet identification and has certain theoretical significance to the development of relevant coal mine intelligent applications.