Research On Wearing Detection Of Electric Power Safety Equipment Based On Convolutional Neural Network

In the safety management of power work sites,the correct wearing of power safety protective gear is crucial to the personal safety of power workers.At present,most of the existing power safety protective gear data sets are safety helmet datasets,and the actual safety protective equipment also has safety belts and tooling.Traditional manual extraction and deep learning methods rely heavily on complex feature extraction steps,do not focus on the safety shield area,and the obtained network model has low detection accuracy and redundant parameters,which is difficult to meet the end-to-end and real-time detection requirements.In response to the above problems,this paper conducts research and summarizes the following:(1)To address the problem of missing safety equipment detection data sets.Firstly,the images required for electric power safety equipment detection were collected and preprocessed to build an electric power safety equipment data set.Secondly,the images of the data set were labeled,and three data augmentation methods of gamma correction,mean filtering and Gaussian filtering were used to expand the data set,which helps increase the generalization ability of the network model and suppress overfitting.Finally,the effectiveness of the data set was verified using five methods.The experimental results show that the validation results on the data set have reached the initial accuracy requirements of electric power safety equipment wearing detection.(2)To address the issue of low accuracy in safety equipment detection models,an improved YOLOv3 method is proposed for detecting electric power safety equipment wearing.Firstly,a CA module is introduced into the main feature extraction network to emphasize the visual features of the relevant parts,improve feature discriminability,and a dense connection structure is adopted to enhance the model’s non-linear expression capability,thus improving the model’s prediction accuracy.Simultaneously,a bidirectional feature pyramid structure is introduced into the feature fusion module to achieve mutual transmission of high-level and low-level feature information,which can improve the performance of the model without increasing computational complexity.Secondly,the K-means clustering algorithm is used to optimize the setting of anchor boxes,and prior knowledge is utilized to improve the performance of the model.Finally,the weighted CIo U loss function is used to replace the mean square error regression loss function,optimizing the regression of predicted boxes to real boxes.The experimental results demonstrate that the detection accuracy of this method reaches 87.21%,which is higher than the accuracy of the YOLOv3 method,an increase of 6.13%.(3)To address the problem of redundant model parameters in safety equipment detection models,a pruning-based method for electric power safety equipment wearing detection is proposed.Firstly,the network is sparsely trained to optimize the redundant network layers and reduce the number of network parameters.Secondly,the channel proportion factors are sorted according to their importance,and channels with smaller proportion factors are filtered out.Finally,after the network model is trained and the weights are fine-tuned,the accuracy of the network model is improved,and it is compared with layer pruning and normal pruning methods.The experimental results show that the size of the pruned network model is optimized to 8.94 MB,the number of parameters is optimized to 2.64 M,and the forward inference time is optimized to 8.2ms.In the context of ensuring power safety,the method proposed in this article is a reliable means to ensure the personal safety of operators in scenarios such as substations,which is of great significance for promoting the development of energy internet in the State Grid.