SVM Hyperparameter Optimization Using a Genetic Algorithm for Rub-Impact Fault Diagnosis

Abstract

Differentiating data samples within various classes is a typical example of machine learning applications. The support vector machine (SVM) is a robust and useful tool capable of resolving the problems of data classification with high accuracy. One of the most reliable kernels that are widely used with SVM to map features into high-dimensional feature space and perform classification is a Gaussian radial basis function (RBF kernel). However, the classification performance of SVM with RBF kernel strongly depends on two hyperparameter values of the trained model. This paper proposes the use of genetic algorithm (GA) as an optimization technique to select proper hyperparameter values for the SVM classifier. GA is known as an intelligent optimization approach based on heuristics that allows finding a relatively good solution to the optimization problem, especially when the exhaustive search of optimal values can be computationally expensive or may not deliver a proper result. In the experimental part of this manuscript, the classification accuracies of the trained one-against-all multiclass SVM (OAA-MCSVM) classifier with hyperparameter values adjusted by the conventional exhaustive grid search optimization algorithm and the proposed one were evaluated and compared using five datasets containing mechanical rub-impact faults of various intensity levels.