Robust early detection of hunting on railway vehicles: single-sensor based methods and critical comparisons

The problem of robust and early detection of railway hunting is considered and two single-sensor ‘dynamics-based’ methods employing adaptive Recursive AutoRegressive (RAR) modelling of the car-body lateral random vibration signal are postulated. The first is based on a Degree-of-Stochasticity (DS) measure of the vibration signal and the second on the minimum underlying Damping Ratio (DR). Fully automated tuning for both methods is achieved via Bayesian optimization. The performance of the methods is systematically assessed via high numbers of Monte Carlo simulations under various Scenarios employing a high-fidelity SIMPACK-based vehicle model and three performance criteria: the True Positive Rate (TPR), the False Positive Rate (FPR), and the Detection Delay Time (DDT) with respect to conventional hunting initiation. The results reveal very good performance and robustness to suspension faults and worn track conditions for both methods, with the DR-based exhibiting an edge. Its performance and achievable robustness are also shown to be clearly superior to those of three alternative methods.