Research on the identification method of non-coal foreign object ofbelt conveyor based on deep learning

Gong-kuang zidonghua

GAO Yan1¹  JIN Baoquan2¹  ZHANG Hongjuan1²  HU Jinghao1² 

[1] ;1.College of Electrical and Power Engineering, Taiyuan University of Technology, Taiyuan 030024, China;

关键词: belt conveyor;    non-coal foreign object identification;    target detection;    deep learning;    yolov3;    focal loss function;   

DOI  :  10.13272/j.issn.1671-251x.2021020041

来源: DOAJ

【 摘 要 】

In order to solve the problems of single identification target and lack of positioning ability of the existing image identification methods of foreign objects, an identification method of non-coal foreign object of belt conveyor based on deep learning is proposed.This method uses the target detection algorithm YOLOv3 as the basic framework, and uses the Focal Loss function to replace the cross entropy loss function in the original model to improve the YOLOv3 model. By adjusting the optimal hyperparameters (weight parameter α and focus parameter γ) to balance the ratio between samples, the method solves the non-coal foreign object sample imbalance problem. Therefore, the model focuses more on learning complex target sample characteristics during training and improves the model forecast performance. A foreign object dataset is built and the classification performance and speed are tested by the foreign object dataset.The results show that the Focal Loss function performs better than the cross entropy loss function in the foreign object dataset, and the accuracy is increased by 5% when γ=2 and α=075. Therefore, the optimal hyperparameter is γ=2 and α=075.The improved YOLOv3 model's identification accuracy of the three non-coal foreign objects of bolts, angle ironsand nuts increases by about 47%, 35% and 68% respectively, and the recall rate increases by about 66%, 35% and 60% respectively. Under the 2080Ti platform, the predicted type of each image is consistent with the actual type, and the confidence level is above 94%.

【 授权许可】

Unknown