Fault detection methods for rail vehicle suspension systems include data-driven methods and model-based methods. Existing model-based methods often face a trade-off between computational efficiency and accuracy: standard linear models fail to adapt to time-varying track irregularities and inherent structural nonlinearities, while nonlinear models are computationally prohibitive for on-board real-time monitoring. To address this challenge, a novel fault detection framework combining the variational Bayesian adaptive Kalman filter (VBAKF) and the cumulative sum (CUSUM) algorithm is proposed, employing the VBAKF algorithm to estimate the time-varying measurement noise covariance online. By modeling the noise variances with Inverse-Gamma (IG) prior distributions, the algorithm adaptively compensates for linearization errors and environmental uncertainties via variational inference. Subsequently, the CUSUM algorithm is applied to the VBAKF residuals to statistically monitor shifts in mean and variance, enabling sensitive detection of suspension faults. The proposed method is validated on a full-scale roller test rig. Meanwhile, the principal component analysis (PCA) is introduced, and the VBAKF-PCA model is established for comparison. Experimental results demonstrate that the VBAKF-CUSUM method has superior performance compared to the VBAKF-PCA method, exhibits robustness under dynamic conditions, and demonstrates remarkable effectiveness in detecting multiple fault types. It also proves that a reasonable model-based method can reliably detect suspension system faults.
Gao et al. (Thu,) studied this question.