Recognition And Application Of Environmental Hazards In Distribution Channels Based On Convolutional Neural Networks

The identification of hidden dangers in the distribution line channel environment is an important research content of the intelligent distribution network line channel safety inspection.Forests,farmland,rivers,construction sites and other power distri-bution lines pass through the terrain as the weak link of the power distribution line channel,which is a frequent occurrence area of potential hazards such as tree barri-ers,bird damage,falling poles,and external damage.However,the traditional manual inspection method to find potential hidden trouble areas is inefficient and dangerous,which brings great troubles to power operation and maintenance personnel.Therefore,there is an urgent need for automatic hidden danger location and identification technology for power distribution lines to facilitate operation and maintenance personnel quickly eliminate hidden dangers.Based on relevant research at home and abroad,this paper designs a distribution line channel environmental hazard identification algorithm based on convolutional neural network and remote sensing detection technology,which can provide powerful power grid operation and maintenance technical support.The main tasks of this paper are:(1)Based on the wide coverage of satellite imagery,high definition and can re-flect the authenticity of distribution line channels features such as geomorphology and geomorphology,this paper proposes two GIS-based distribution line channel environ-ment capture algorithms,which are used to capture the actual distribution line channel environment and follow-up distribution line channel terrain and landform recognition data source.In order to solve the current bottleneck caused by the lack of environmental data sets of distribution channels in the power field,this paper uses the Google Map-based distribution line channel capture algorithm to establish a terrain and landform detection data set dedicated to the field of distribution networks,and uses Labelimg labeling to establish further identify the data set and divide it into four types of hid-den terrains: forests,farmland,rivers,and construction sites.At the same time,a map API-based multi-level route channel capture algorithm suitable for most GIS platforms is proposed,and the high-resolution surrounding environment of the tower is obtained using the Gaode map as an experimental case.(2)Aiming at the distribution line channel hidden danger terrain and landform de-tection data set,this paper proposes the VGG16 feature extraction network + Faster RCNN framework + the proposed lightweight convolutional neural network to real-ize the location and detection of the distribution line hidden danger terrain and land-form.First,the VGG16 deep convolutional neural network is used to extract its high-level feature expression,and then the distribution line channel terrain detection model is trained based on the Faster RCNN intelligent algorithm.A total of 20 models are saved in the training phase,and the Loss curve is analyzed and the converged model is se-lected.In the test,the model with the highest AP value is selected as the final detection model of this category of topography.For the hidden terrain and geomorphic hazards detected by Faster RCNN with low confidence,this paper proposes a simple structure and efficient convolutional neural network to reclassify them,to comprehensively im-prove the recognition rate of hidden hazards in distribution lines.The network uses L2 regularization combined with the Dropout layer to avoid network overfitting,and uses Adam optimization algorithm to update parameters.Experiments show that the network has a high recognition rate for hidden hazards of different terrains.The correct rate of construction site classification is 94.9%,the correct rate of farmland classification is76.7%,the correct rate of mountain forest classification is 94.4%,and the correct rate of river classification is 70.6%.