基于深度学习的矿用输送带损伤检测方法  

Damage detection method for mine conveyor belt based on deep learning

作　　者：张梦超 周满山[1,2] 张媛 于岩[1] 李虎 ZHANG Mengchao;ZHOU Manshan;ZHANG Yuan;YU Yan;LI Hu(College of Mechanical and Electronic Engineering, Shandong University of Science and Technology,Qingdao 266590, China;Libo Heavy Industries Science and Technology Co., Ltd., Tai'an 271000, China;CCTEG Changzhou Research Institute, Changzhou 213015, China)

机构地区：[1]山东科技大学机械电子工程学院,山东青岛266590 [2]力博重工科技股份有限公司,山东泰安271000 [3]中煤科工集团常州研究院有限公司,江苏常州213015

出　　处：《工矿自动化》2021年第6期51-56,共6页Journal Of Mine Automation

基　　金：国家重点研发计划资助项目(2018YFC0604702);山东省自然科学基金资助项目(ZR2018BEE014)。

摘　　要：针对目前输送带损伤检测方法缺乏对输送带撕裂以外其他损伤类型研究的问题,提出一种基于深度学习的矿用输送带损伤检测方法,通过Yolov4-tiny目标检测网络对输送带损伤类型进行分类。Yolov4-tiny目标检测网络以CSPDarknet53-tiny作为主干特征提取网络,借鉴Resnet残差思想,使用残差块防止深层网络中高层语义特征丢失,同时采用特征金字塔网络实现高低层语义信息融合,达到提高检测精度的目的;将CSPDarknet53-tiny中的2个有效特征层输入预测网络Yolo Head,通过得分排序和非极大值抑制算法对预测框进行筛选,从而预测输送带损伤类型。实验结果表明,Yolov4-tiny目标检测网络在输送带损伤数据集上对表面划伤、撕裂、表面破损和击穿4种损伤类型检测的平均精度分别为99.36%,94.85%,89.30%,86.76%,平均精度均值达92.57%;与Faster-RCNN,RFBnet,M2det,SSD,Yolov3,EfficientDet和Yolov4目标检测网络相比,Yolov4-tiny目标检测网络在数据集上取得了最快的检测速度,帧速率达101帧/s,实现了较好的速度与精度的平衡,且占用计算资源相对较少;通过对数据集外新鲜样本的检测,验证了本文方法具有较好的泛化能力。In order to solve the problem that the current conveyor belt damage detection methods lack research on damage types other than conveyor belt tear,a damage detection method for mine conveyor belt based on deep learning is proposed.And the conveyor belt damage types are classified by the Yolov4-tiny target detection network.The Yolov4-tiny target detection network uses CSPDarknet53-tiny as the backbone feature extraction network,draws on the Resnet residual idea,uses residual blocks to prevent the loss of high-level semantic features in the deep network.At the same time,the method uses feature pyramid network to obtain the fusion of high-level and low-level semantic information to achieve the purpose of improving detection precision.The two effective feature layers in CSPDarknet53-tiny are input into the prediction network Yolo Head,and the prediction frames are filtered by the score ranking and non-maximum suppression algorithm to predict the types of conveyor belt damage.The experimental results show that the average precision of the Yolov4-tiny target detection network on the conveyor belt damage data set for the four damage types of surface scratches,tears,surface damage and breakdown is 99.36%,94.85%,89.30%,and 86.76%respectively,and the mean average precision is 92.57%.Compared with Faster-RCNN,RFBnet,M2det,SSD,Yolov3,EfficientDet and Yolov4 target detection networks,the Yolov4-tiny target detection network achieves the fastest detection speed on the data set with a frame rate of 101 frames/s.The network achieves better balance between speed and precision,and occupies relatively less computing resources.The detection of fresh samples outside the data set verifies that the method in this paper has good generalization ability.

关 键 词：带式输送机 损伤检测 深度学习 目标检测 表面划伤 撕裂 表面破损 击穿 

分 类 号：TD528[矿业工程—矿山机电] TD634