Transmittance function-based robust detection, isolation and precise estimation of degradation in railway vehicle suspension components under varying travelling conditions

Abstract A transmittance function-based diagnosis methodology for robust detection, isolation, and precise estimation of railway vehicle suspension components degradation under varying travelling conditions is presented. The methodology is based on transmittance function data-driven stochastic functionally pooled models, which are capable of representing the vehicle dynamics under varying operating conditions. This characteristic ensures robustness in fault/degradation diagnosis as the effects of varying travelling conditions on vehicle dynamics may fully or partly obscure the degradation effects. Degradation detection is accomplished in an unsupervised manner, while isolation and estimation are supervised. The travelling conditions are considered non-measurable in the inspection phase, where the methodology is assessed via different scenarios and the transmittance function is employed for the regularization of track excitation uncertainty. The methodology is validated through thousands of Monte Carlo simulations with a Simpack-based 54-degree-of-freedom vehicle model of the Attiko Metro company using acceleration signals from two points on the vehicle. A wide range of travelling speeds, passenger payloads, and different levels of degradation (through stiffness reduction) of the primary suspension conical rubber spring and the air spring of the secondary suspension are considered, combined with different realizations of track irregularities following the ERRI B176 standard. The results indicate excellent detection performance for even early stage degradation levels, while correct isolation is achieved for approximately 94% and 99% of the test cases considered in the vehicle’s primary and secondary suspension components, respectively. The corresponding degradation level estimation accuracy for these components is demonstrated via the mean absolute error, which is 2.45% ± 2.92% and 4.05% ± 3.66%. The methodology’s performance is additionally compared with that of a state-of-the-art data-driven method, showing an overall superior performance.