Study On The Fault Line Detection Using Machine Learning For Small Current Grounding System

Fault line detection for small current grounding system has always been adifficult point in the field of relay protection. Because single line selection method hasmany limits and single-phase earth fault is multiplicity and changeable, the accuracyof fault line selection is very low. In order to improve the accuracy of fault lineselection, the combination of many single line selection methods using informationfusion technology is studied. The key problem of information fusion technology ishow to handle sampled data. At present, sampled data of line selection which istreated by the existing algorithms has the huge sampled data, dimension disaster andgreat empirical risk.First, single line selection methods are verified. Among them, in order toimprove the accuracy of the transient line selection method, Hilbert-HuangTransform(HHT) is proposed. HHT is not affected by interference signal. Andcharacteristic component of transient signal is decomposed by HHT in turn. At thesame time, HHT can accurately reflect the characteristics of transient signal. Theresults of experiment show that HHT has the strong ability to extract features oftransient signal.Secondly, the method fault measure is established by the information gaindegree at the stage of the fault measure. By the method fault measure, the weigh ofsingle line selection methods is given and the internal data of one line is enhanced. Onthe basis, with the deep analysis of the characteristics of sampled data of line selection,the conclusion is showed that sampled data of line selection is high dimensions,imbalance and complex distribution.Thirdly, according to the characteristics of sampled data of line selection, thedimension of sampled data is reduced by personal computer assistant(PCA) and theimbalance is decreased by new sample methods-Grid Distribution Sample(GDS) andField SMOTE Sample(FSS). Through the analysis, when sampled data decreases totwo dimensions and the ratio of fault line and non-fault line is1:3or1:3.5, sampleddata is well distributed, less time consumption and the accuracy of fault line selectionis improved.Finally, the combination algorithm of ADABOOST and classificationperformance based on kernel function is proposed. The fused algorithm has the abilityof concatenation and parallel to treat sampled data. At the stage of experiment,algorithm performance is tested by adjusting the ratio of noise sample and total sample. It is showed that the modified ADABOOST algorithm improves the accuracyof the transient line selection method, and can be more suitable to hostile lineselection environments compared with other algorithms.