296 / 2018-09-25 21:21:01

GIS Partial Discharge Type Identification Based on Optimized Support Vector Machine

Support Vector Machines; Overfitting; Kernel function parameter; Regularization; Principal Component Analysis; K-fold Cross Validation

In order to accurately determine the type of GIS partial discharge, the Support Vector Machine (SVM) was employed to identify the discharge type. When SVM in the classifier training, the isolated data points in the data samples will cause over-fitting because of the existence of interference signals. The training error was continuously reduced and the complexity of the model was continuously improved, which eventually leads to an increase in generalization error and a decrease in recognition rate. Selecting appropriate kernel parameter and regularization was proposed in this paper to optimize the SVM to improve the accuracy of GIS partial discharge type recognition. Principal Component Analysis (PCA) was applied to reduce the dimension to improve the recognition speed.
Firstly, the K-fold Cross Validation (K-CV) was adopted for cross-validation to determine the optimal kernel function parameters, the kernel function used in the SVM is a radial basis kernel function. Gaussian radial basis nucleus has good anti-interference ability, its parameters determine the scope of the function because of its strong locality. The K value was taken as 5, and the training samples were divided into five groups, four of which were selected as training groups and the other selected as test groups. Changing the training group and the test group until all 5 groups were selected as test groups, the average accuracy of each group and the corresponding kernel parameters were recorded after each iteration function by performing 50 cross-validation, and the optimal kernel function parameters obtained was σ=1.6.
Then the regularization was adopted to make certain restrictions on the parameters of the model, so that the model prefers simpler parameters. In this paper, the training of SVM was considered as a constrained optimization problem with the objective of maximizing the boundary. During the GIS partial discharge type identification based on optimized SVM, the σ=1.6 was set and the L2 regularization constraint was applied. By SVM confusion matrix, the optimized SVM has a classification accuracy rate of 96.5%, a training time of 5.326 seconds and a prediction rate of 850 obs/sec.
Finally, the PCA was applied to reduce the dimensionality of the data. By solving the eigenvalue vector of the covariance matrix, discarding the small variance and retaining the term with large variance, the dimension of the 7-dimensional data in this paper can be reduced to 5 dimensions. After dimension reduction, the optimized SVM has a classification accuracy rate of 96.2%, a training time of 4.829 seconds and a prediction rate of 970 obs/sec. Again, the PCA was adopted to study the variance contribution rate of the five groups of features to reduce the dimension to four dimensions. After dimension reduction, the optimized SVM has a classification accuracy rate of 95.8%, a training time of 3.933 seconds and a prediction rate of 1100 obs/sec.
The results show that the optimized SVM achieves higher classification accuracy. After the dimension reduction by PCA, the recognition accuracy has a 0.4% reduction, however, the sample training speed was significantly improved and the over-fitting phenomenon was suppressed. The classification accuracy and rapidity of GIS partial discharge type identification were improved.