Research On Abnormal Detection Method Of Idler Based On Deep Learning

The idler is one of the important components of the belt conveyor,which has the function of material transportation and support.If the idler fails during operation,it will affect the transportation system.Therefore,there are great importance of the research on the abnormal detection method of the idler,such as guaranteeing the transportation safety of mining production and preventing the malignant accidents.This paper is supported by the project"Research on Intelligent Patrol Conveyor for Mining Belt Conveyors" of Changzhou Institute of China Coal Industry and Engineering Group.It takes rollers as the main research object,adopts the combination of deep learning and machine vision.The specific work and research innovations are as follow:(1)The roller operation dataset is constructed.The current roller operation data is moslly monitoring video images.It is aimed at the problems of poor video quality,serious roller screen occlusion,time-consuming interception of relevant clips,and a small number of difficult to support deep learning training.This paper is simulated the running track of the inspection device in the simulation laboratory and three real mines,data collected by controlling the different influencing factors of the rolles.As the same time,the Labellmg open source tool is used for Labeling the video screen according to the frame sequence diagram.Finally,2160 videos are obtained for a total of 8.4h,and each video is about 10s～15s.and the dataset was namedT Gdata2019.(2)In the influence of the environment and the hardware factors of the inspcction device during the data collection process,some data screens cannot achieve the expectcd effect.In order to retain the motion information completation of the feature point,speeding up the optical flow extraction,a optimized optical extraction method is proposed.Firstly.is used to process the video datasd.The image is enhanced through fog:then an appropriatc asymmetric half-plane area(NSHP)image model is established,and the spatial model is converted into a system state equation applicable to Kalman filtering,and then passed "prediction+feedback"the frame picture is subjected to noise reduction;finally the optical edge is extracted by extracting the detailed edge feature points on the surface of the roller with the operation.Experimental results show that the optimized optical flow extraction method can improve the accuracy of feature point tracking,which is more obvious than the traditional optical flow extraction method,and the extraction speed is also significantly accelerated.(3)Based on the premise of the hardware requirements of the inspection device in the project and the classification scene of the roller operation,the dual-stream CNN+LSTM network model is constructed to address the deficiencies in the current mainstream video classification algorithm.In order to speed up the training of the model,the CNN model uses an improved lightweight VGG16 network structure.Firstly,a spatial pyramid pooling(SPP)layer is added to the last convolutional layer to adapt to the multi-scale change of the roller spatial texture feature map.At the same time,the global pooling is used to replace the deep features learned by the fully connected layer fusion network to reduce the redundant network parameters in the fully connected layer;Then,the deep descriptors learned by the convolutional layer are fused and their L2 normalized to accelerate Network training convergence.Experimental results show that this method greatly reduces the network parameters and speeds up the model training speed.This paper solves the problem of roller abnormally recognition that is difficult for traditional methods by constructing a suitable network model,improves the abnormally detection rate while optimizing network performance,and provides technology for coal companies to realize safety guarantee in the transportation chain supporting,at the same time,the construction of the roller operation dataset is also contributed to the subsequent research.