Research On Key Technologies Of Belt Damage Detection System For Underground Coal Mine Belt Conveyor

China is short of fossil energy such as oil and natural gas,which account for a low proportion of the total energy consumption,and the energy consumption structure is dominated by coal.Underground coal mine belt conveyor is the main equipment for transporting coal during the process of coal mining.With the development of industrialization and intelligentization,high-speed heavy-duty belt conveyors have become the main development direction of the belt conveyor used in underground coal mine at present.High-strength belts are the main bearing unit of belt conveyor.Keeping belts in good condition is essential for safe and stable production in coal mine.However,the production environment of underground coal mine is harsh,the conveyor belt is wearing gradually by high-speed and heavy-load coal transportation.In addition,in the process of coal production and transportation,mixed foreign bodies such as metal rods and gangues may cause belt tearing damage,which is highly likely to lead to serious production accidents.Above all,it is essential to study the belt damage detection method for underground coal mine belt conveyor.The belt of underground coal mine conveyor is regarded as research object in this thesis.Different belt damage types in the coal transportation process is analyzed.Based on machine vision and deep learning theory,belt damage detection algorithm for underground coal mine belt conveyor is proposed and belt damage detection system is developed,which provides theoretical support and technical support for real-time detection of the belt operating state.Research contents of this thesis are as follows:(1)The overall framework of belt damage detection system for underground coal mine belt conveyor is established.The main structures and belt damage types of underground coal mine conveyor and belt are analyzed.Based on the research of pros and cons of mainstream belt damage detection methods,the overall framework of conveyor belt damage detection system is proposed.The hardware composition,main software architecture and corresponding deployment of conveyor belt damage detection system are designed.(2)MCC-Cycle GAN sample generation algorithm based on Generative Adversarial Networks is proposed.Various kinds of image sample generation algorithms are analyzed,and sample generation theory based on Generative Models are studied.A Cycle GAN sample generation model fused with multi-classification module is established.The training method for deep learning model based on mixed real-fake samples strategy and periodic decay mixing rate are proposed,and the training process of MCC-Cycle GAN model is optimized.The loss function based on gradient penalty Wasserstein distance and standardized Softmax with adaptive boundary penalty is constructed,which improves the training stability and generated sample quality of MCC-Cycle GAN model.(3)A De INN image denoising model based on Invertible Neural Networks is established.Image denoising and enhancing theory based on deep learning is analyzed,and high frequency noise separation strategy based on GLOW coupling and Haar wavelet transform layers is proposed.High frequency noise separation model based on INN is studied,and the structure of Trainable Guided Filter is designed.Loss function of De INN model based on multiscale structural similarity and edge gradient is established,which optimizes convergence process and stabilize training process of De INN model.(4)A belt damage detection model based on one-stage framework and a load balance algorithm based on LSTM are proposed.Channel-spatial attention strategy based on fusion information is designed to realize the efficient fusion of latent feature maps for backbone network.Deep learning model compress method based on Knowledge Distillation is studied.The pre-trained Teacher model is treated as soft labels for Student model so that the lightweight Student model achieves better damage detection performance.The statistical sample feature distribution theory is studied,and loss function of KDFA-Center Net model based on fusion loss of latent feature maps and response result is established.A load prediction and balance model based on LSTM is proposed to address the problems of network congestion,image backlog and sudden load surge of detection server.A load distribution strategy based on task complexity,residual computing power and fitness evaluation is established.By comparing the training results of various loss functions,the LSTM load balancing model based on Mean Square Error is established.(5)Verification experiments of conveyor belt damage detection system are carried out.Hardwares and softwares of conveyor belt damage detection system are designed.Experimental test bench for belt damage detection system is established and experiments are carried out.KDFA-Center Net damage detection model and De INN denoising model are deployed,and belt damage detection experiments with different belt speeds and industrial camera sensitivity parameters and industrial test are carried out.The experimental results in laboratory and industrial environment show that m APs of the proposed belt damage detection system have reached 95.96% and 98.1%,respectively.The load balancing experiment of conveyor belt damage detection is carried out.The experimental results show that average resource utilization rate decreased by 29.1% and average delay time decreased by 91.6%.It is verified that the proposed load balancing algorithm can effectively alleviate the image processing blocking when the load increases suddenly and improve the reliability and stability of conveyor belt damage system.The dissertation has 85 figures,31 tables and 144 references.