Research On Conveyor Belt Damage Detection Technology Based On Advanced Lightweight Convolutional Neural Network

The stable operation of the conveyor belt under the mine is an important prerequisite to ensure safety in production,but the underground environment of the coal mine is complex.Because the coal is mixed with sharp objects such as gangue and thin rods,long-term operation can easily cause the surface of the conveyor belt to be damaged,and even lead to tearing in severe cases.If the traditional machine vision method is used to detect the running status of the conveyor belt,the pre-processing to enhance the damaged target of the image often increases the calculation time in the early stage of detection;At the same time,because the damaged features extracted by the algorithm lack location information,it is impossible to locate the damaged area.Therefore,based on the regional convolutional neural network,this paper improves the feature extraction module and feature fusion mode,and puts forward an improved detection method of conveyor belt damage using lightweight regional convolutional neural network,which not only realizes quick identification of damaged targets by using lightweight convolutional neural network in real-time,but also improves the learning ability of minor damage by multi-scale feature fusion and attention mechanism in detection accuracy.The innovative work of this paper is as follows:(1)As there are few data sets of conveyor belt damage published online in open source,this paper sorts out and classifies the existing data sets of conveyor belt,and photographs the image samples of conveyor belt in complicated scenes such as dust and dim light.At the same time,in order to effectively increase the number of samples in the data set,the whole experimental data set is expanded by data expansion method,and the effective samples are marked to form a complete data set of conveyor belt damage,which lays a foundation for the selection of benchmark model and the research of follow-up algorithm.(2)Aiming at the problems of low detection accuracy and efficiency of the traditional detection algorithm for damaged targets,an improved lightweight regional convolution neural network is proposed.In feature extraction of the original network model,its deep convolution level leads to high computational complexity.Therefore,this paper introduces a lightweight convolutional neural network(Mobile Net V2),which uses deep separable convolution to retain the feature representation ability,reduce the model parameters and improve the detection efficiency of the network.Secondly,in order to avoid the problem of insufficient receptive field of the model caused by the small scale of network convolution,a large scale convolution kernel is used in the feature extraction stage to expand the receptive field of the network and enhance the feature extraction ability of the model.(3)Aiming at the problem that the convolution feature map of the original lightweight regional convolution neural network does not effectively combine the feature information,which leads to the low detection accuracy of micro-damage,this paper introduces the feature pyramid structure to learn the deep and shallow feature information through multi-scale feature fusion,which is helpful to enhance the detection ability of micro-damage.At the same time,in the stage of feature fusion,because the feature scale transformation is easy to cause the loss of feature information,attention mechanism is adopted to adjust the feature weight,so that the network can effectively select important channel information and highlight the target area of features.