Research On GIS Equipment Fault Diagnosis Method Based On Deep Learning

As the application of GIS（gas-insulated substation）is expanding,how to ensure the safe and smooth operation of GIS and to effectively diagnose the faults of GIS equipment has become an important research content in the power industry.At present,the diagnosis methods for partial discharge fault of GIS equipment are based on the physical and chemical phenomena in the fault state,but these methods have different limitations.After analyzing and studying the existing partial discharge fault diagnosis methods for GIS equipment,it is proposed to combine SF₆gas decomposition components with deep learning to improve the convenience,practicality and universality of diagnosis.Therefore,a deep belief network GIS equipment fault diagnosis model is established based on the principle of SF₆gas decomposition component ratio,and the model is optimized using particle swarm optimization algorithm to achieve the purpose of being able to effectively diagnose the type of partial discharge fault of GIS equipment.The specific work is as follows:1.To understand the current use and fault situation of GIS equipment and the development of deep learning;to grasp the current status of research on the application of deep learning technology to GIS equipment fault diagnosis at home and abroad;to understand the fault diagnosis methods based on the physical and chemical phenomena occurring inside GIS equipment in the case of partial discharge,and to focus on the GIS equipment fault diagnosis methods based on the analysis of SF₆gas decomposition components.2.Propose the fault diagnosis method of GIS equipment based on deep belief network of SF₆gas decomposition component ratio,learn about deep learning and neural network,and select deep belief network as fault diagnosis model after analysis and comparison;master the basic principle of restricted Boltzmann machine and deep belief network,and sort out the fault diagnosis process;adopt max-min method To normalize the data,t-SNE method is used for data dimensionality reduction visualization to reduce data redundancy and complete the data set partitioning process;through a large number of experiments to determine the hyperparameters of the model,Dropout method is used to reduce the overfitting of deep belief network;finally,the trained network is used to diagnose the fault data of GIS equipment and establish BP diagnosis model for effect comparison.The experiments show that the DBN model is applicable to the research content of this thesis and can effectively make judgments on the four types of partial discharge faults of GIS equipment.3.According to the shortcomings of the established deep belief network model such as slightly weak feature extraction ability of fault data,comparing several current intelligent optimization algorithms,we propose to optimize the deep belief network fault diagnosis model by using particle swarm optimization algorithm,mainly optimizing 2hyperparameters such as the number of hidden layer neurons and learning rate.The optimization process is determined,the parameters of the optimization model are set by reviewing the data and experiments,the optimization model is established for simulation experiments,and the results are compared with the three models of DBN,BP and PSO-BP.The experiments show that the particle swarm optimization algorithm can effectively optimize the deep belief network model structure and improve its fault diagnosis accuracy,and the PSO-DBN model can be better applied to the diagnosis of partial discharge faults in GIS equipment,which also provides a new research idea and method for the application of deep learning to GIS equipment fault diagnosis.