Comparison of feature selection methods applied to Barkhausen noise data set

摘要

This paper discusses the feature selection problem and provides results from feature selection task in the field of non-destructive testing. The features are selected for quantitative prediction of residual stress from the Barkhausen noise signal. It is stated in the literature that the model behaviour depends on the used features and thus the selection must be carried out carefully. The selection methods studied in this paper are forward-selection, backward-elimination, simulated annealing and genetic algorithms. The used data set is divided into training and external validation data sets. The training data set is used in feature selection. The selection algorithms utilize leave-multiple-out cross-validation procedure in deciding which features are selected. The results show that backward-elimination performs poorly while the other three methods provide reasonable results. The results from the selection indicate that the stochastic methods outperform forward-selection but the external validation shows that in this case forward-selection provides results comparable to the studied more advanced methods. Even though the results indicate that forward-selection gives comparable results, it has been shown in the literature that stochastic methods are more likely to find the global optimum and thus should be used especially when the problem complexity increases.