A coupled Two-Dimension Convolutional Neural Network-Gated Recurrent Unit (2DCNN-GRU) model is proposed to evaluate and predict the hunting instability of high-speed railway vehicles in this paper. First, vibration accelerations of four measuring points on the surface of the bogie frame of a high-speed railway vehicle in good working condition and with hunting instability are obtained through a line test and model simulation. The vibration acceleration data under different conditions is cut into many pieces at equal intervals. Low-frequency band-pass filtering is applied to each piece to obtain filtered vibration data, which is then analyzed separately to get a sample set of spectrum images, including short-time Fourier spectrum, Hilbert time-frequency-amplitude spectrum, and marginal spectrum. Then, a 2DCNN model is proposed to extract features by deeply studying the spectrum images of each piece of the filtered vibration data. The root-mean-square (RMS) of the vibration responses of four measuring points on the surface of the bogie frame and the mean value of the filtered vibration response envelope are calculated and recorded for each piece. The Hunting Instability Index (HII) is proposed by considering the weighted mean of RMS and the envelope mean of the filtered vibration responses to quantitatively get the extent of hunting instability. Finally, the GRU method is applied to predicting the dynamic change of HII indicators, and the effectiveness and accuracy of the method are verified by typical examples. One contribution of this work is proposing a method to evaluate the hunting motion by image identification of the short-time Fourier spectrum, Hilbert time-frequency-amplitude spectrum, and marginal spectrum of vibration signals, and another is the definition of HII based on 2DCNN and statistics.
Read full abstract