The Partial Discharge UHF Signal Recognition System Of GIS Based On Random Forest Algorithm

GIS(Gas Insulated Switchgear) has played an important role in the power system because of its excellent performance,high security and stability.However,various external factors and prolonged operation may leave insulation defects in GIS and cause insulation failure.PD(Partial Discharge) UHF(Ultra High Frequency) signal classification and recognition has received widespread attention as an effective method to identify and analyze the types of insulation defects in GIS.In order to improve the efficiency and accuracy of GIS PD signal classification and identification,this paper presents a UHF PD signal recognition system based on random forest algorithm with combined feature parameters.First of all,this paper uses UHF signal acquisition method to obtain and pretreat the partial discharge signal of four typical insulation defects generated by the GIS PD simulation device as a data set,and then draws it as the scatter plot.According to the characteristics of the signal image,and 32 sets of feature parameters are selected by designing new feature parameters and choosing some traditional statistical features and image moment parameters to improve the data discrimination,and to establish a parameter space which can describe of the signals more accurately for the PD signal classification,to avoid the low accuracy caused by low parametric dimensions.Then,the paper divides the dataset into training datasets and test datasets by 10-fold cross validation to improve the usage of data samples and weaken the randomness of evaluating the classifier.Bagging(Bootstrap Aggregating) algorithm is used to obtain multiple sample subsets from the training set data,and the parameter subsets are obtained randomly from the characteristic parameter space with the same number of sample subsets but with indefinable dimensions to build random forest.By using a large number of decision trees in random forest to share the dimension of feature parameters,the inefficiency because of the excessively high characteristic parameters is solved.And the over fitting effect is not caused by the increase of trees,thereby overcoming excessive requirement about feature parameter dimension of other traditional pattern recognition methods.Based on this,the final classification result is got through the discrimination and voting the test data set between decision trees.Simulation and experimental results verify the effectiveness of this method in GIS PD signal classification.Next,based on the recognition program of random forest classifier,this paper improves its analysis function,designs and implements GIS UHF partial discharge signal recognition system.The system obtains the signal of UHF PD through UHF signal processing terminals,and uses ZigBee and WIFI to transmit data with the system host computer,and applies the computer which is installed the analysis identification program to analyze and classify the data.Finally,in order to verify the effectiveness of the recognition system,this paper carried out GIS simulation PD test and field operation test.The final test results show that the system has the ability to achieve practical applications.