Research On Real-time Detection Technology Of Conveyor Belt Surface Damage Based On Improved YOLOv5

Conveyor belt is an indispensable equipment in mineral collection and production.In long time operation,due to the bump between the surface of the conveyor belt and the ore,the conveyor belt is prone to tearing,breaking and other damage,which causes significant harm to property and safety.Therefore,the timely detection of conveyor belt damage is an extremely important task.At present,most of the main conveyor belt damage detection relies on regular manual inspection of the conveyor belt under the condition of machine idle,this method has the problems of low efficiency,accuracy and stability can not be guaranteed,high testing costs.In this paper,we propose a two-stage conveyor belt surface damage detection method based on target detection and image segmentation based on deep learning technology,which firstly detects and locates the damage area by target detection algorithm,and then segments the damage area by image segmentation algorithm,and calculates the damage area to achieve quantitative analysis of conveyor belt surface damage.The quantitative analysis of the damage on the conveyor belt surface is achieved by calculating the area of the damaged area,and the specific research is as follows:In order to address the current situation that deep learning algorithms need a large amount of training data to ensure their accuracy and generalization,and there is a lack of high-quality conveyor belt damage data sets,this paper proposes a data enhancement method combining GAN and Copy-Pasting strategies to address the problem of insufficient number of conveyor belt damage data samples in addition to labeling the conveyor belt image samples collected in the field,which greatly enriches the number of samples and morphology and lays the foundation for the subsequent increase of model detection accuracy and generalization performance.To address the limitations of the computational performance of the inspection equipment in the actual production environment and the requirements for inspection accuracy and speed,this paper improves the target inspection model YOLOv5 using two methods,model pruning and knowledge distillation,to inspect the damaged parts of the conveyor belt surface.First,model pruning is used to reduce the number of model parameters and speed up the detection speed of the model,and then knowledge distillation is used to effectively improve the detection accuracy of the model without increasing the model computation.After the experiments,it is proved that this method can achieve 97.33% accuracy in the detection of conveyor belt surface damage,while the detection speed reaches 200 FPS,which meets the requirements of accuracy and real-time.Since current conveyor belt damage detection methods are usually limited to the qualitative detection of the damage area and cannot accurately analyze the damage area quantitatively,this paper proposes a segmentation algorithm based on improved BiSeNet to further segment the shape and size of the damage area at the pixel level and calculate the area of the damage area based on the target detection of the damage area of the conveyor belt.The quantitative analysis of conveyor belt surface damage is realized.The overall process of conveyor belt damage detection is realized by setting the safety threshold of the damage area according to the actual demand and alarming the parts larger than the threshold.