Design And Research Of Mine Fire Edge Intelligent Detection System

Mine fire is always one of the key disasters in the field of mine safety,mine fire is mainly divided into external fire and internal fire.External fire is caused by the combustion of combustible substances caused by external high-temperature heat sources,and internal fire is mainly spontaneous combustion of mine coal seams.At present,the detection methods of external causes of mine fire mainly adopt the video-based detection method and sensor-based detection method.The sensor-based detection method has a high accuracy when the sensor is fully covered in the detection environment,but the cost is high,and when the laid sensor is far away from the fire source,it cannot detect the fire source in time.The video-based detection method is easily affected by factors such as mine underground environment,and the accuracy of early fire identification is low.Due to the limited underground network resources,when the deployment scheme of the fire detection system is based on the cloud,there will be transmission delay,response lag and other problems,and it is difficult to meet the real-time requirements of fire detection.In order to improve the accuracy and real-time detection of external fire in mines and improve the reliable identification of initial fires,this thesis proposes a method for detecting external fire in mines based on edge intelligence.First,the YOLOv5 s target detection algorithm model is introduced into the real-time detection of the fire caused outside the mine.In order to enhance the detection ability of small target flame,the YOLOv5 s algorithm is improved.The feature scale of the backbone of the network model is modified and a small target detection layer is added to fully learn the shallow features and improve the performance of small target detection.In order to reduce the loss of context information in the convolution process,an adaptive attention mechanism is added to the neck part of the model.The spatial attention mechanism is used to generate spatial weight graphs for each feature graph,and context features are fused with the weight graphs to generate new feature graphs containing multi-scale context information,so as to reduce the loss of context information and further improve the model detection performance.Secondly,a multi-source information fusion method is proposed to view the fact that the downhole video detection method is easily affected by environmental factors such as non-ignition light source,water mist,and dust.This method uses multi-sensor to assist decision-making,adopts the dynamic adjustment strategy of fire risk evaluation indicators,and fuses the information detected by video and the information detected by multiple sensors to make a fusion judgment,so as to improve the accuracy of fire detection.Thirdly,in order to meet the real-time requirements of fire detection,the designed fire detection algorithm model is transplanted to the intelligent edge processor.The information collected by the video and sensor is calculated and analyzed directly on the edge side,and the results are fed back at the field end,which solves the problems of data transmission delay and response lag existing in the cloud-based deployment scheme and improves the speed of fire detection.Before transplantation,the model is quantified as INT8 fixed-point model,and the size of the quantified model is only a quarter of the original model,which greatly reduces the number of parameters of the model,facilitates transplantation on edge devices,and completes the final fire detection task.Finally,a visual interface is designed based on LabVIEW graphical programming language to realize real-time monitoring and visualization of fire information and improve the level of management.Experimental results show that,compared with the YOLOv5 s model,the reasoning speed and detection accuracy of the proposed method are increased by 27%and 4.8%,meeting the requirements of fire detection.