Machine Learning Method for Small-Amplitude Bogie Hunting Identification of High-Speed Rail Vehicles

Hunting stability is the foundation of the dynamic performance of high-speed trains. Due to factors such as track irregularities, the evolution of wheel-rail relationships, and degradation of vehicle suspension parameters, small-amplitude bogie hunting (SABH)was discovered in the line test of high-speed trains. To meet the higher requirements of existing high-speed trains and standard EMUs for vehicle dynamic performance, it is necessary to study the accurate and rapid identification method for SABH. The distribution differences of vibration characteristics between normal and SABH samples were compared using the Wasserstein distance. Power spectral entropy (PSE) and wavelet energy ratio (WER) have the farthest Wasserstein distance. The importance of features was ranked using MRMR and Rrelief, and the classification importance score of the autocorrelation coefficient (ACC) was the highest. Comparing the application effects of different machine learning methods in identifying SABH, the accuracy of semi-supervised and supervised learning methods is comparable for existing samples, but the unsupervised classification methods tend to misclassify the normal samples as SABH. Among semi-supervised learning methods, the random forest classifier performs the best; In supervised learning methods, decision tree classifiers perform best. Finally, the impact of the number of features on the accuracy of the classifier was analyzed, with 2–3 features being optimal. As an example, the newly obtained experimental data was labelled using semi-supervised learning methods.