Research On Smoke And Wildfire Location And Recognition Based On Deep Learning

Flammable materials such as mountains,forests,and trees often appear around the transmission line channel.If a forest fire occurs,it will pose a great threat to the safety of the line.Especially when the transmission line passes through complex terrain such as old-growth forest,steep mountain and large area of no man’s land,manual patrol operation is very difficult.Once mountain fire occurs,it is difficult to find and put out in time,resulting in rapid spread of fire and difficulties in safe operation of power equipment.So,combining real-time images captured by cameras on the tower with object detection algorithms to detect fires in a timely manner can greatly save labor costs and reduce safety hazards caused by forest fire.In this paper,the real smoke and mountain fire images collected from the transmission line patrol inspection are labeled,and the data set needed for the smoke and mountain fire detection algorithm is constructed.By analyzing the characteristics of smoke and mountain fire targets,two smoke and mountain fire location and recognition algorithms are proposed.On the one hand,from the perspective of improving the detection accuracy,a detection algorithm of transmission line mountain fire based on DETR is proposed;On the other hand,in order to complete real-time detection tasks on embedded devices,a lightweight network based on YOLO is proposed.Finally,the above algorithms are encapsulated and deployed in the smoke and fire detection system.The main work of this paper includes:(1)In order to improve the detection accuracy of smoke and mountain fire,this paper proposes an improved DETR smoke and mountain fire detection method.This method adds hole convolution and multi-scale information in the backbone network to improve the feature extraction ability,adds relative position coding in Transformer to make up for the lack of relative position information in the self-attention mechanism,and finally uses CIOU in the loss function to make the algorithm more easily convergent.The experimental results show that the average accuracy of the proposed model can reach 84.2%,6.25 percentage points higher than the original DETR.(2)In order to detect the mountain fire smoke target in real-time on the embedded device with limited computing power,this paper proposes a lightweight target detection model based on YOLO framework.Due to the limited computing resources of embedded devices,this method simplifies the network structure of YOLOv5-s from two aspects.First,it only simplifies the feature fusion network and the detection head part of the original network structure.The recall rate and accuracy are improved when the parameter quantity is nearly doubled compared with YOLOv5-s;Secondly,the modules of the backbone network were deleted,which greatly reduced the number of parameters of the model.The number of parameters and the size of the model were reduced by 14 times,and the recall rate and precision of the model were slightly reduced.Finally,it is deployed and tested on embedded devices.The results show that the lightweight smoke and mountain fire detection model proposed in this paper can achieve real-time monitoring effect in embedded devices on the premise of maintaining a high level of accuracy and recall.(3)In order to apply the improved algorithm mentioned above to practical work more conveniently and quickly,a visual smoke and mountain fire detection system based on Pyqt has been developed,and the above algorithm has been encapsulated in the system,so as to achieve automatic positioning and detection of mountain fire smoke targets in transmission line channels by embedded devices and back-end servers simultaneously.