Transformer Fault Diagnosis Algorithm Based On Frequency Response Data

Transformers are considered as indispensable apparatuses in power systems, which is im-portant to assure reliable power supply. Once the geometrical structure or the solid insulationsof a transformer is damaged, the fault will probably develop into a catastrophic failure if it is fol-lowed by other accidents. Therefore the issues how to diagnose faults in transformers efficientlyand rapidly are widespread concerned.Among transformer fault diagnosis techniques, frequency response analysis(FRA) receivesabroad attention and research due to its high sensitiveness to structural changes in transformerwindings and core. However the current fault diagnosis methods based on FRA simply comparethe curves intuitively, without any standards to quantify the deviations between curves. Addi-tionally the methods of noise reduction and pattern recognition of the frequency data need to beimproved.In this article an adaptive rectangle structure element is utilized to denoise the frequencydata based on the binary mathematical morphology. The height of the structure element willchange dynamically according to the global amplitude range of the frequency response curvesand the frequency value of data being processed. Moreover, an algorithm to dynamically di-vide the whole frequency range into five subbands is presented, which is the foundation of thefollowing fault diagnosis processes.Subsequently, the equivalent gradient-area parameters are proposed to quantify the devia-tion between the currently measured curve and the reference curve, which avoids the difficultiesof calculating the shifts of resonance peaks and anti-resonance peaks directly. It is proved thatthese parameters can describe the deviation between curves quite accurately.Considering the condition that it is difficult to obtain many frequency response data be-longing to transformers that suffer from different faults, a transformer winding fault diagnosisalgorithm based on the equivalent gradient-area parameters and the support vector machinesis proposed. It shows that our trained SVM classifier can obtain high precision and accuracyunder small sampling condition, which illustrates that the proposed equivalent gradient-area pa-rameters are remarkable fault features for transformer winding faults. Besides, the classificationresults indicate that the proposed fault diagnosis algorithm has enormous potentials to be appliedin actual fault diagnosis process.