Video Monitoring System For Non-coal Foreign Matter Of Belt Conveyor Based On Deep Learning

As a key component of belt conveyors,coal conveyor belts are often damaged and torn due to sharp non-coal foreign objects(such as broken angle irons,broken anchors,etc.)entering the conveyor belt,which affects the safe production of coal mines and causes high economic losses.In response to this problem,this paper designs a belt conveyor non-coal foreign object detection video system based on deep learning.The system is based on the YOLOv3 target detection framework and the Jetson NANO development board,with the goal of the detection accuracy and real-time speed,using Focal Loss function and Cut Mix data augmentation to improve the model,to achieve the high-precision detection,real-time,and miniaturization requirements of the system,and solve the problem from the source of the damage and tear of the conveyor belt.Taking into account the requirements of non-coal foreign matter detection accuracy and speed performance,this article describes the principles of common object detection methods,analyzes the advantages and disadvantages of the two types of algorithm models,sets the algorithm model evaluation indicators,builds an image data set based on non-coal foreign matter,and uses VOC2007 The data set and the self-made non-coal foreign matter data set to analyze the performance of the algorithm model and choose.The results show that the YOLOv3 algorithm model has better comprehensive performance.This algorithm model is selected as the basic model.Aiming at the insufficient detection precision rate and recall rate of the YOLOv3 algorithm model on the non-coal foreign matter data set,the loss function improvement strategy and the data enhancement method improvement strategy are proposed.In the aspect of loss function improvement,analyze the realization principle of Focal Loss function,improve the category prediction loss and confidence probability prediction loss in the original loss function,and introduce the grid search method for hyperparameter setting;in terms of data augmentation method improvement,analyze and compare the realization principle of the two data augmentation methods of Mixup and Cut Mix.Through the design of comparative experiments,according to the improvement of foreign body detection precision rate and recall rate,the Focal Loss function and Cut Mix data augmentation method are selected to improve the YOLOv3 algorithm model.The results show that the algorithm model shows better detection accuracy under the two strategies,which determined as a non-coal foreign body detection model.The underground laying conditions of computers are harsh,and the use of edge computing units can solve the problem of difficulty in the mine well.In this paper,the non-coal foreign object detection model is deployed on the Jetson NANO development board,with industrial cameras and human-computer interaction components for video stream acquisition and prediction results display.Due to the limitation of the computing power of the edge computing unit,the Tensor RT engine is used to accelerate the quantitative optimization of the algorithm model and accelerate the reasoning,and reduce the number of model parameters.At the same time,use the industrial camera SDK for secondary software development,and write industrial camera call,video acquisition,and video frame processing programs.In addition,the use of multi-threading technology enables video stream frame processing and algorithm model inference and prediction to be synchronized and run quickly to meet real-time performance requirements.In order to test the overall performance of the non-coal foreign object detection video system,offline experiments and online experiments of the system are designed.The test results show that the system detection speed is significantly accelerated under the acceleration of the Tensor RT engine,and the average working frame rate is close to 5 FPS,high detection accuracy can be maintained,and the system basically meets real-time application requirements.The research of this non-coal foreign object detection video system provides a feasible solution for the early warning of foreign matter in the underground belt conveyor belt.The system judges the existence of foreign matter by obtaining the video information of the conveyor belt,and predicts the type and location of the foreign matter at the same time.For the prevention of belt conveyor belt breakage and tear,it is of great value to ensure the long-term stable operation of the belt conveyor.