Research On Image Classification Of Bird Species Related To Transmission Line Faults Based On Deep Learning

The classification of bird species related to bird fault in transmission line is of great significance and practical value to the prevention and control of bird damage in power grid.Statistics show that the number of line trips caused by bird damage is second only to lightning stroke and external force damage.The power grid operation unit has carried out a risk assessment of bird failures and installed bird protection devices on the transmission lines so as to prevent line failures caused by bird activities effectively.However,the lack of knowledge of the types of birds involved in birdrelated failures by operation and maintenance personnel makes precise prevention and control remain impossible.Due to the continuous breakthrough of deep learning in the field of fine-grained image recognition,it is a trend to intelligently identify the bird species related to bird-related damages,and then achieve differentiated bird prevention.According to the image classification of bird species related to transmission line faults,thesis has carried out the following three research tasks:(1)Based on the statistical data and the investigation results of bird species,the basic information table of bird species that causing damage to transmission lines was extracted.The fine-grained images of related birds are collected by manual shooting and network crawling,and the image samples are manually labeled one by one.A bird image datasets including 5400 images of 90 bird species was established.(2)A transfer learning method based on Image Net datasets and pre-trained deep convolutional network model is proposed.According to the mainstream deep model,a network model of 18 architectures is built.And through comparative analysis,the interlayer function that matches the network model is determined.The weight parameters of the model run to convergence on the Image Net datasets are imported into the built model for fine-tuning and retraining.Experimental results show that the Res Net101 model based on global fine-tuning has the best classification effect on bird species image datasets,with an accuracy of 87.96%.This proves the applicability of transfer learning on small sample datasets and the advanced nature of the deep residual network architecture.(3)The backbone network and multi-scale feature map were optimized on the SSD framework,the influence of background noise on the feature extraction from the model was effectively reduced.Res Net101 model is used to replace VGG16 model,which overcomes the defect of poor characterization ability of the original SSD model,and the feature Mosaic module is designed.The ability of the model to segment background and target is improved by combining the shallow features and the top features.The experimental data shows that the optimized model achieves 91.6% average accuracy and 92.41% classification accuracy on the bird species datasets,which proves the feasibility of the above optimization scheme.The method proposed in this paper can identify the types of harmful birds on transmission lines in real-time and accurately,and provide a reference of bird knowledge for the operation and maintenance personnel to carry out the prevention and control of bird hazards.