Study On Discharge Characteristics And Feature Extraction For Suspended Movable Particles Defects In Transformer Oil

With the rapidly expanding of the smart grid construction, higher requirements onthe power system security and reliability have been proposed. Large power transformersis the most important and expensive equipment of the power transmission devices, itssafe and reliable operation plays a vital role for the entire power grid. Partial discharge(PD) is one of the main reasons for causing internal insulation deterioration of largetransformers, online monitoring of PD can promptly and accurately determine the statusof the transformer internal insulation to prevent power transformer accidents and ensurepower system security and stability, therefore, it has an important and practical value.In this paper, based on the actual suspended movable particles defects in powertransformers, designed a set of PD test platform that can simulate suspended movableparticles defect, used the ultra high frequency method for signal testing. For the randompulse disturbance can not be effectively suppressed, leading to the PD image added withwhite noise, developed Contourlet transform denoising algorithm, the simulationexperiments prove its effectiveness. Pulse coupled neural networks (PCNN) featureextraction algorithm is studied, the results proved the necessity and effectiveness of thePD image denoising, the extracted discharge characteristics are used in PD patternrecognition, enriched the theory of feature extraction methods in PD pattern recognition.The main achievements are as follows:According to the characteristics of suspended movable particles defect existing inoil-immersed power transformer, a set of PD test platform that can simulate suspendedmovable particles defect is designed, three typical particles defects are used in PDmeasurement to obtain a large number of artificial test data, a systematic analysis for thePD signals waveform characteristics in different experimental conditions are studied.First studied the PD characteristics of suspended movable particles defect and itsinfluencing factors, the increase of applied voltage and suspended particles content willlead to a more intense discharge development, while the the increase in the oil flowvelocity will help improve the insulation strength, there is a suitable temperature rangethat makes development of partial discharge severity to a minimum. Experimental andtheoretical analysis should be carried to identify the appropriate range of operatingtemperature and flow velocity range for each transformer, to improve the internalelectric field distribution of the oil duct, thus make the maximum extent to reduce the hazards of partial discharge.For the random pulse disturbance resulting the PD image dye-noise problem, firstproposed Contourlet transform noise suppression processing. Based on analyzing themain factor influenced the effect of Contourlet transform denoising, which are noiselevel, the decomposition layer, the decomposition of direction number and the differentPD types, it can be seen that noise level and the decomposition layer has great influence,revealing that Contourlet transform algorithm can correctly extract the distributioninformation from PD three-dimensional image, effectively suppress the white noise inPD image.First proposed a PD feature extraction method based on the PCNN output entropysequence, the PCNN output entropy sequence in a certain range has the character oftranslation and scaling invariance. The proposed criteria for β value makes theentropy sequence faster convergence and thus greatly reduce the length of entropysequence, improving the speed of the identification. Extracting the PCNN outputentropy sequence from the dye-noise PD images and denoised PD images, find that theoutput entropy sequence from the dye-noise PD image has huge mean square error withthe other types PD defects, while it is smaller in the similar discharge patterns, helpfulfor the accurate classification of the PD image and prove the effectiveness and necessityfor PD image denoising by Contourlet transform.