Damage Detection Method Of Feeding Belt Based On Target Localization Enhanced YOLOv3

Belt conveyor is an important facility for conveying materials in many industrial fields.Damage detection of feeding belt is of great significance to the stable operation of belt conveyor.However,the commonly used feeding belt damage detection system mostly uses leakage detection device,which is affected by the environment and prone to misdetection.Convolutional neural network has the advantages of high precision and high reliability in the field of target detection,which provides a new solution for the detection of feeding belt damage.Therefore,this thesis proposes a feeding belt damage detection method based on target location enhanced YOLOv3,and uses adjacent frame analysis to realize the feeding belt damage alarm mechanism.The main research achievements and innovations of this thesis are as follows:(1)The image enhancement algorithm of feeding belt based on image component fusion is proposed;Aiming at the problems of overexposure of damaged area,uneven background light and too many shallow scratches interfering with target detection in the image of feeding belt collected,an image enhancement algorithm of feeding belt based on image component fusion is proposed in this thesis.After the damage feature of belt image was enhanced by line laser assisted,the original image was processed by an improved multi-scale Retinex algorithm for high-light suppression at the damaged area,the histogram equalization algorithm of edge reservation was used to eliminate the illumination imbalance of background,and the image noise was reduced by bilateral filtering.Finally,the results of the three image enhancement methods are fused to the HSV color model to obtain the belt image for target detection.(2)A belt damage detection method based on the enhanced YOLOv3 target location is proposed;The accuracy of traditional target detection algorithm can not meet the needs of belt damage alarm.In this thesis,an enhanced YOLOv3 network for target location is constructed to improve the target location accuracy.Firstly,Dense Block is added into the feature extraction network to improve the feature extraction capability of the network.Secondly,a branch with smaller receptive field is added to the multi-scale prediction network to solve the problem of detecting small-scale damage.Finally,the normal distribution is introduced for the boundary box prediction part,and the predicted boundary box coordinate is changed into the predicted boundary box coordinate distribution,so the accuracy and reliability of boundary box detection are improved.(3)Belt damage alarm algorithm based on adjacent frame analysis is proposed;In this thesis,a belt damage alarm algorithm based on adjacent frame analysis is proposed,which uses parameters such as the size and confidence of the prediction frame,and analyzes the position of the prediction frame of adjacent frame damage as the basis for judging the belt damage alarm.Finally,the reference index of damage severity is given based on various damage characteristics.The experimental results show that the proposed method can effectively detect and alarm the damage of the feeding belt.The m AP of the damage target detection reaches 88.7%,the Io U and the detection confidence are higher than other comparison algorithms,and the final damage alarm algorithm also reaches the comprehensive accuracy of95.88%.