Small-amplitude bogie hunting identification method for high-speed trains based on machine learning

Hunting stability is a critical aspect of the dynamic performance of high-speed trains. However, the current evaluation criteria only address heavy hunting motion, and they require supplementation and improvement to meet the increasing requirements of dynamic performance. Therefore, a study was conducted to develop an identification method for small-amplitude bogie hunting (SABH) motion. A total of 39,816 samples were obtained from field-measured bogie lateral acceleration and manually labelled as normal and SABH. The distribution comparison, Pierce coefficient, and importance ranking showed that harmonic character-related features (the autocorrelation coefficient, approximate entropy and etc.) should be considered more for SABH identification than amplitude-related features (maximum and RMS). The most critical features for SABH identification were the autocorrelation coefficient and spectral frequency spread. The classifier performance comparison showed that the decision tree was the best-performing classifier, followed by the support vector machine, while the linear discriminant, K-nearest neighbour, and naive Bayes classifiers were less effective. Sensitivity analysis indicated that the proposed method worked well with a sampling frequency of 50-200 Hz and a window length of 5-10 s. This research provides theoretical support for hunting instability monitoring and real-time active control studies for high-speed trains.