Research On Deep Hybrid Networks And Its Application In Transformer Fault Diagnosis

Dissolved gas in oil analysis(DGA)technology analyzes dissolved gas in transformer to monitor its operating status and becomes the main means of t he fault diagnosis of oil-immersed transformers.Based on exploring the structure of deep learning model and combining with DGA data,this paper attempts to apply the deep learning methods to the fault diagnosis of transformers to solve the problems of low diagnostic accuracy and unavailability of unlabeled samples,help maintenance workers judge the operation state of the transformer and decide whether to carry out the overhaul.The paper proposed a deep learning hybrid network(DLHN)algorithm and applies it to the transformer fault diagnosis method.Based on the comparative analysis of the advantages and disadvantages of Auto-encoder and Restricted Boltzmann Machine,a deep learning hybrid network model is built and DLHN is implemented with the data of dissolved gas analysis technology.The method is based on the stacking of AE and RBM,which can pre-train effectively with a large number of unlabeled data,obtain relative optimal initialization parameters,and use a small amount of tagged data to fine-tune the parameters to further improve the diagnostic accuracy.The paper also proposed an improved deep hybrid network(IDHN)algorithm and applies it to the transformer fault diagnosis method.This paper introduces Denoising Auto-encoder and Gauss Bernoulli Restricted Boltzmann algorithm to improve DLHN to make it has a stronger robustness and anti-noise ability,which further improves the diagnostic performance.It is applied to the transformer fault diagnosis and improves the transformer fault diagnosis performance.The two fault diagnosis methods mentioned in this paper are used to test the actual data of power plant transformers.The test results are compared with the results of the fault diagnosis methods based on BP and SVM.The results show that the two methods proposed in this paper can use a large number of unlabeled samples available on the project site effectively,and have better diagnostic performance,which is suitable for practical engineering needs.