Research On Pattern Recognition Of Partial Discharges For Oil Paper Insulation In Transformers Based On Parallel Feature Domain

Partial discharges(PD)are not only the deterioration precursor of oil paper insulation in transformers,but also one of the main causes for the transformer accidents.Pattern recognition of the PD defects in the transformer can provide the health conditions on the oil paper insulation system.Firstly,the mathematical model of the PD signals and the electrode models of the PD defects in oil paper insulation environment are constructed.The PD pattern recognition is studied from three aspects: noise suppression,single source pattern recognition and multi-source pattern recognition.In the aspect of noise suppression of the PD signals,many complex insulation environments bring a lot of noise to the PD signals.The time-frequency structure of the PD signals after energy attenuation is clearly demonstrated by using the synchro-squeezing wavelet transform.The proposed scheme on the multi-level threshold de-noising gradually suppresses the noise in the oil paper insulation in the form of steps.The high-resolution synchro-squeezing wavelet transform can resist the influence of the noise in the complex oil paper insulation environment,and the time-frequency coefficient matrix generated by it shows the characteristics of the noise and the pure signals.Through the noise suppression of the PD signal constructed by the mathematical model and the PD signals generated by the electrode model,it is found that the proposed method has stronger de-noising ability,and provides the sample of the relative pure signal for the research on the subsequent pattern recognition.In the aspect of single source pattern recognition for the PD signals in oil paper insulation environment,the traditional phase resolved pulse sequence can't provide more information for the PD defect faults because the information is affected by the physical and chemical characteristics of two kinds of insulation media.Thus the recognition rate of different PD defects is reduced.In this paper,the dynamic mode decomposition is used to show the fractal information of the PD time series,and the fractal dimension and the lacunarity are used to further quantify the fractal characteristics.Finally,X-means clustering is adopted to identify the different PD defects.It is found that the constructed feature is highly specific and shows the fractal behavior of the partial discharge under oil paper insulation completely,and the clustering algorithm has a higher recognition rate for three different PD defects.In the aspect of multi-source pattern recognition of the PD signals in oil paper insulation environment,the occurrence of the multi-source PD phenomenon is random and dispersive due to the oil paper insulation environment.Meanwhile the multi-source PD signals containing the information of the complex defect faults can't be described in the traditional feature extracted manually.In this paper,the stacked auto encoder algorithm is employed to train the PD signal sets in time-domain and in time-frequency domain simultaneously.And the transformed L1 norm with the guided stochastic subgradient proximally are utilized to solve the over fitting in training.Through the pattern recognition of four different multisource PD signals generated by the electrode model,it is shown that the parallel training method expands the fault information in the oil paper insulation environment contained in the characteristic quantity.The combination of the transformed L1 norm and the method of proximally guided stochastic subgradient can ease the over fitting problem in the deep learning model,and improve the recognition rate in the four multi-source PD defects.