Fault Diagnosis for Transformers Based on FRVM and DBN

Dissolved gas analysis (DGA) of insulation oil is widely used in potential fault analysis for transformers. In order to improve the accuracy of fault diagnosis, a hybrid model which combines the FRVM with the depth belief network (DBN) is proposed to establish the mapping relationship between gas and fault types. Considering that DBN needs to extract a huge amount of feature information, this paper uses FRVM to separate the discharge and overheating faults, and then uses DBN to realize further fault diagnosis. The diagnosis accuracy is studied when IEC ratio, Rogers ratio, Dornenburg ratio and non-cod ratios are used as input parameters, and the results show that the correct rate of diagnosis is highest when the non-cod ratios are used as characteristic parameter. In addition, the method has better performance compared with single DBN, support vector machine and artificial neural network, and it has the ability to diagnose multiple faults.