Research On Arc Fault Detection Algorithm Based On Convolutional Neural Network

Reliable and effective arc fault detection technology is the key to preventing electrical fires.However,in AC electrical systems,the randomness,diversity and complexity of arcs bring great challenges to arc fault detection.There are shortcomings in the existing arc fault detection technologies,which have limited the promotion of arc fault detection technology.Based on the existing technology,this research included the following aspects:1.The main arc fault detection technology is to manually extract arc current features and classify them by model,but the artificial extracted current features are subjective and limited.To some degree,there are problems such as poor portability between datasets and poor generalization.In response to these problems,this thesis designed a convolutional neural network model for arc fault detection that directly took current signals as input,and used dilated convolution and other techniques to optimize it,and compared with methods of manually extracting features,which indicated that the convolutional neural network was effective and better in arc fault detection.2.Introduce attention mechanism into fault arc detection technology.By adding convolutional block attention modules to the convolutional neural network,a convolutional neural network combining channel attention and spatial attention was proposed,which was used for arc fault detection.By experimental comparison,it was proved that the attention mechanism was helpful for arc fault detection.In addition,by the visual analysis of the channel attention and spatial attention learned by the model,it showed the interpretability of the attention mechanism in the arc fault detection.3.By studying the correlation between the arc current characteristics and the type of load which generated the arc fault,a multi-task learning model for fault load detection was proposed.This model can not only determine whether a fault arc occured in the circuits,but also identify the specific type of load which generated the arc fault.By experiments we proved that the performance of the model on both tasks was higher than the single task model.In addition,by redefining the load type and modifying the loss function,a arc fault detection model with universal applicability based on multi-task learning was proposed,which achieved the best arc fault detection performance in the study of this thesis,it proved that introduction of the fault load identification task was helpful for improving the arc fault detection performance of the model.