Research And Application Of Smoke And Fire Recognition Technology In Electric Power Field

Smoke and fire are ruthless,which are more incisively reflected in electric power filed related to the national economy and people’s livelihood.Once the fire spreads,it could cause a short circuit and trip,at worst,burn all kinds of equipment,cause a major power outage,and threaten people’s lives and property.If the hidden dangers can be discovered and dealt with in time at the early stage,serious consequences can be effectively avoided.At present,the electric power field still uses traditional means such as manual inspection,smoke detection,and manual video surveillance to prevent smoke and fire,which have problems such as low accuracy and slow timeliness.Therefore,in order to liberate manpower and improve work efficiency,it is imminent to study efficient and convenient smoke and fire recognition technology for the electric power field.In response to this problem,this thesis focuses on a vision-based intelligent smoke and fire recognition method in electric power field,which introduces emerging deep learning and computer vision technologies,and improves the technical solution from the three perspectives of data,model,and post-processing.We propose a smoke and fire recognition method based on improved YOLOv3 to solve the difficult problems encountered during the application landing process such as false positives caused by complex backgrounds,false negatives caused by too small target scale,and slow and unstable reasoning speed.The main research work can be introduced as follows:Aiming at the problem of false alarms caused by complex environmental interference,in terms of data,three data argumentation methods such as brightness transformation,Gaussian noise increase,and image mixing are introduced,which reduce the influence of various external conditions such as brightness changes and noise on the model detection effect.Improving the adaptability of the model to the complex and changeable environment.In terms of model,considering the changeable and irregular characteristics of smoke and fire,this thesis implements the optimization of the convolution method by embedding a deformable convolution layer that can effectively adapt to the target deformation in the residual module,which can better capture the characteristic information of the target and reduce the interference of the complex background in the environment.The effect of the model has been further improved.Aiming at the problem of missing report caused by too small target scale,a new detection scale is added on the basis of the three existing detection scales.Small targets are represented by high-resolution features on the shallow large feature map,and targets are predicted from four scales,so that the model can adapt to the scale change of smoke and fire.Besides,this thesis introduces a target box clustering analysis method based on K-means++algorithm.In the specific implementation,we clustered to obtain a priori box that meets the target characteristics of the smoke and fire recognition task according to the scale characteristics of the target in the recognition task.Hence the training effect of the recognition model are improved.Aiming at the problem of slow and unstable reasoning,this thesis proposes inference and deployment optimization methods based on TensorRT and TensorFlow Serving,TensorRT is used to accelerate the reasoning of the model,and the effect of speeding up and reducing memory usage is achieved with a small loss of accuracy.TensorFlow Serving tool is used to build a model deployment framework that meets industrial applications,and meets the deployment requirements for reliable and stable operation.