Research On Fire Detection Method Of Electric Vehicle Charging Station Based On Deep Learing

Electric vehicles(EVs),which utilize clean energy,are essential to the development of low-carbon transportation as the availability of fossil fuels decreases and air pollution increases.By 2025,it is anticipated that the number of EVs in China will reach 17,82 million,the total number of charging stations will be approximately 9,391 million,and the ratio of vehicles to charging stations will be 1:0.53.To shorten the charging time of EVs,high-voltage,high-current,and other fast charging methods are commonly employed,resulting in severe battery heating and the risk of fire accidents,causing significant economic losses,and endangering the safety of individuals.Therefore,fire detection in real time for EVs charging stations is crucial.When a fire occurs at EVs charging station,the electrical equipment burns violently and emits a great deal of gray-black smoke.The flame merges with the smoke boundary and it is easily affected by wind and light,resulting in complex and variable characteristics.Traditional fire detection methods,such as smoke detectors,are unsuitable for this scenario due to their low detection accuracy under these conditions.This paper proposes a fire detection method for EVs charging stations based on YOLO(You Only Look Once),analyzes the challenges of flame and smoke target detection,and constructs a dataset,in terms of irregular target feature extraction and a lightweight algorithm model to optimize.Select the optimal model through ablation experiments and deploy hardware devices to optimize.This paper’s primary research content is as follows:First,beginning with the specific detection scenario of the EVs charging station,the fire’s cause,characteristics,and detection challenges are analyzed.Images of the flame and smoke dataset are subjected to screening,expansion,and labeling,resulting in a dataset suitable for fire detection in EVs charging stations.Second,based on the network structure and detection principle of YOLO,a Ghost Net fire detection algorithm is proposed for EVs charging stations.In the original YOLOv4 algorithm,the Ghost Net module,the K-means++ clustering algorithm,and the Coordinate Attention(CA)module are introduced.Monitoring equipment on-site to implement dynamic fire detection.Then,the model is placed on a central cluster monitoring platform to allow for centralized monitoring of multiple sites,and an improved YOLOv7-based fire detection algorithm is proposed for EVs charging stations to improve the accuracy of the model’s detection.Change the Rep Conv module in the head to a depth-wise over-parameterized convolutional layer(DO-Conv),improve the detection head’s ability to extract target features,train the optimized YOLOv7-DO-Conv model,and evaluate real-time detection performance.Last,the model is deployed on a low-power embedded platform to achieve realtime detection at the local end,and a fire detection algorithm for EVs charging stations based on improved YOLOv7-Tiny is proposed.Introduce the non-parametric attention mechanism module(Sim AM)and omni-dimensional dynamic convolution(OD-Conv)to enhance the network’s feature capacity.Through ablation experiments,the performance of each improved model is compared,and the optimal model is chosen for deployment on the NVIDIA Jetson Xavier NX.The experimental results demonstrate that the improved model satisfies the requirements for real-time detection on the lowpower computing platform in the local area.