Research On Miners’ Unsafe Behavior Recognition Algorithm Based On Lightweight Neural Network

The development of the coal industry plays an important role in the supply of national energy.In the coal industry with miners as the main body,coal mine accidents caused by non-standard occupational operations seriously affect personnel safety.Therefore,the research on unsafe behavior of miners is of great significance in reducing the accident rate of coal mines.To accurately identify unsafe behavior of miners in the complex coal mine background environments such as low light has become the key to solving this problem.With the development of deep learning technology,it provides new opportunities for the research of unsafe behavior identification algorithms of miners,but it also brings a lot of resource equipment consumption.Due to the complexity of the deep convolutional network and the limited resources and equipment in the mine,the research on the unsafe behavior of miners based on deep learning usually transports the collected data to the cloud for analysis and judgment,which leads to a delay in the judgment process and cannot guarantee real-time of unsafe behavior.Therefore,lightweight network with fewer parameters is studied for the real-time response problem of unsafe behavior.The primary research of this thesis is as follows:(1)Aiming at the static image discrimination task of miners’ unsafe behavior,dynamic convolution kernel based on efficient attention mechanism is proposed and applied in lightweight network to discriminate unsafe behavior.The feature expression ability of the lightweight network is increased through the dynamic convolution kernel.Four channel compression methods are used in the attention mechanism of the dynamic convolution kernel to solve the redundancy problem and decrease the parameters.The adaptive down-sampling method is adopted to reserve useful information and enhance the performance to discriminate images.(2)Since the static image network model does not consider the timing information,3D lightweight convolutional neural network is proposed for the action recognition task of the video set.The two-dimensional lightweight network is improved into a three-dimensional network,which not only learns the characteristics of the time series,but also ensures the lightweight of the model.Three-dimensional dynamic convolution kernel based on the attention mechanism is designed,which can extract the important information in the network through the attention mechanism and improve the identification performance of network.(3)In order to further verify the effectiveness of the proposed network model in judging unsafe behaviors of miners,a static image set and a video set of unsafe behaviors of miners are respectively constructed.By verifying the dynamic lightweight network with efficient attention mechanism on the static image set,it is found that its accuracy is improved by 2.4% compared with the static network.The number of parameters is reduced by 20.17% compared with the standard dynamic network.The improved 3D dynamic lightweight convolutional network is used for video recognition,the results show that the recognition accuracy rate reaches 88.58%,and the average F1 value reaches 87.27%.The article has 30 figures,24 tables and 89 references.