Research On The Identification Of Coal And Rock Infrared Thermal Image Damage Area Based On Deep Learning

With the rapid development of China’s economy,the demand for energy is increasing day by day.China’s energy dominated by coal resources will not change in the short term.Coal consumption has gradually depleted shallow coal resources,and the mining of deep coal has become an irreversible trend.However,coal and rock dynamic disasters caused by the uncertainty of the deep coal mining environment seriously threaten people’s lives.Through the research,it is found that the coal and rock damage and destruction activities are the precursors before the coal and rock dynamic disasters.By detecting the damage area of coal and rock and then providing early warning of coal and rock dynamic disasters,the occurrence of coal mine disasters can be effectively avoided.Existing detection methods for coal and rock damage area are not only inaccurate but also require manual participation.The rapid development of infrared imaging technology and deep learning technology provides the possibility for intelligent detection of coal and rock damage area.To this end,this thesis studies the detection method of coal and rock damage area based on deep learning.The main work and contributions of this thesis are as follows:(1)In order to solve the problem of a lack of coal and rock infrared thermal image datasets,this thesis conducts coal and rock uniaxial compression experiments according to the requirements of coal and rock mechanics,and collects infrared thermal images in the process of coal and rock damage and destruction.Firstly,grayscale is performed to reduce the occupied space.After data enhancement with operations such as rotation,translation,zooming,and clipping,the damaged area is marked with Label Me and the format is converted after labeling to establish coal and rock infrared thermal image datasets.(2)In order to remove the noise in the infrared thermal image,this thesis proposes a dense residual image denoising algorithm based on autocorrelation network.Firstly,an asymmetric multiscale convolution module is introduced to extract features for the first time,and the asymmetric structure is used to reduce the parameter amount of the module to realize multiscale feature extraction;secondly,the lightweight dense residual cascaded autocorrelation block is introduced,and the one-dimensional fast convolution is used instead of the dense connection in the pseudo 3D auto-correlation blocks,and the two-dimensional structure is used to simulate the three-dimensional convolution to achieve the integration of the horizontal,vertical and channel directions,and the dense residual connections are introduced to fuse feature information and enhance feature propagation;finally,skip connection and global residual connection are introduced to facilitate cross-layer flow and retain more prior information,which can improve the training effect.The reconstruction module outputs the denoised image after network reconstruction and compares it with BM3 D,Dn CNN,FFDNet and IRCNN denoising algorithms.The experiment proves that the denoising algorithm proposed in this thesis can effectively denoise coal and rock infrared thermal images.(3)In order to improve the accuracy of identifying coal and rock damage area,this thesis proposes a coal and rock infrared thermal image damage area segmentation algorithm based on U-Net network.Firstly,the convolution layer in the middle of the encoder and decoder is replaced with the serial atrous spatial pyramid pooling module,which can effectively extract the edge information of the damage area by using the atrous convolution to increase the perceptual field size,and the series structure is used to solve the problem of partial information loss caused by atrous convolution;then,the skip connection is replaced with the attentional feature fusion module to achieve the fusion of deep and shallow semantic information.The feature information is filtered to reduce interference while paying more attention to the smaller damage area;finally,dense conditional random fields are introduced after the output,and the global information of the image is used to refine the edge segmentation and solve the problem that the edge boundary of the damaged area in the infrared thermal image is blurred,which makes it difficult to accurately identify the damaged area.The algorithm in this thesis uses transfer learning to train on the coal and rock infrared thermal image datasets,and compares it with U-Net,U-Net++,Atten-UNet,Seg Net,FCN-16 s and Deeplab V3+segmentation algorithms.The experiments show that the accuracy of the algorithm in this thesis reached 94.36%,and the MIo U value reached 86.93%.Compared with other algorithms,it can better complete the segmentation of coal and rock damage area.(4)Based on the above research,this thesis designs a coal and rock damage area detection system,realizes the automatic detection of the damage area,and proves the reliability and usability of the system through system testing.This thesis includes 37 figures,17 tables,and 91 references.