A Research Of Data Augmentation Method For Transmission Line Insulator Images

The monitoring and maintenance of transmission lines are high-risk and high-cost tasks.In recent years,with the development of deep learning,more and more fields have applied deep learning,and there have been many advancements in grid monitoring,defect detection,and other areas.However,the application of deep learning requires a large number of labeled samples to supervise training,and data collection in the transmission environment is difficult due to technical protection and safety factors,making it difficult to achieve large-scale data collection.Therefore,how to perform data augmentation on transmission line environment images is a worthy research problem.Non-destructive insulator samples and defective insulator samples have their own domain requirements in deep learning.This thesis addresses the lack of image samples of non-destructive insulators and defective insulators in transmission lines,and proposes data generation models based on generative adversarial networks for each type.The main research contents are as follows:(1)A method for generating insulator data based on Cycle-GAN.In response to the problem of many types of non-destructive insulators,difficult data sample collection,and small data samples,this thesis proposes a model based on the Cycle-GAN architecture,which incorporates self-attention mechanisms and channel attention mechanisms.At the network structure level,a skip-layer structure is added to the generator to enable it to transfer feature information through skipping.The overall model uses an attention module to supervise the learning of the generator,achieving mutual transformation of different types of insulators while retaining background feature information of the samples before transformation,providing an effective solution for increasing non-destructive insulator samples under different backgrounds.(2)A method for generating defective insulator data based on transfer learning.In response to the problem of a lack of samples of defective insulators,this thesis proposes a method for generating defective insulators based on feature transfer.By using the attention module and generator trained in the non-destructive insulator mutual transformation network model,a compensation generator module and transfer discriminator are designed for secondary training.At the network level,an autoencoder structure is applied to the compensation generator part,enabling it to have directional generation capabilities.The discriminator adds a skip-layer structure,allowing the discriminator to rely not only on high-dimensional features but also on low-dimensional features for final discrimination.(3)Based on the proposed method for generating defective insulator samples,a targeted generation of defective samples was performed.Then,an effective comparison experiment was designed between a manually operated defective sample dataset and a dataset generated by the proposed method.Yolov5 algorithm was used to train models on the datasets before and after augmentation,and their performance was compared on the same test set.The experimental results show that the proposed model has better generation ability and produces samples of higher quality and better identifiability compared to traditional prior information models.It is suitable for deep learning datasets.