Abstract Summary
Using axle box accelerations (ABA) measured from rail vehicles is a cost-efficient way of monitoring railway track conditions. Existing techniques are focused on identifying a specific type of degradation, usually at a single track layer, such as at the rail top, fastenings, insulated joints, ballast, etc. The basic principle is to identify unique ABA features that correspond to a particular degradation. For instance, increased magnitudes of wavelet power spectrum between certain frequencies were used for the detection of singular rail surface defects, such as squats. A common approach for identifying such ABA features is through sensitivity analysis, which employs a vehicle-track interaction (VTI) model and simulates track degradations via parameter changes in the VTI model. However, different types of degradation may cause similar changes in specific ABA features. In this case, a wider range of ABA features should be investigated to determine if they can distinguish between different degradations. In addition, conventional sensitivity analysis only varies one parameter at a time while fixing the other parameters, making it difficult to consider the combined effects of multiple degradations. As a result, it is still challenging to distinguish between various types of track degradations, particularly when they occur at different track layers at the same time. Instead of focusing on a specific type of track degradation, this paper aims to identify interpretable and generic ABA features measured on a baseline track without degradations. Furthermore, the most contributing track parameters or operational variables to different ABA features are determined through global sensitivity analysis, where multiple parameters of a VTI model are varied simultaneously to simulate the effects of combined degradations under various operational conditions. The outcome of this paper can be used to establish a quantitative relationship between ABA features and track parameters, which can be applied to track condition monitoring. First, ABA features are identified through field measurements at a well-maintained plain track. We apply the synchrosqueezed wavelet transform to the ABA signals and construct a feature space with peak frequencies and magnitudes of the power spectrum. Furthermore, based on hammer tests and pass-by measurements, different peak frequencies are associated with different mode shapes of the VTI system and thus can be used to monitor the condition of their corresponding components. Identified ABA features serve as a baseline for calibrating and validating the VTI model. Subsequently, a high-dimensional parameter space is defined to represent track conditions (such as the track geometry and railpad/ballast stiffness) and operational variables (such as the axle load and vehicle speed). The parameter space covers both the nominal and degraded values, as well as the uncertainties associated with them. We sample from the parameter space to obtain multiple sets of space-filling input parameters for the VTI model. In such a way, different track parameters are varied simultaneously across multiple samples. Through simulations with sampled parameters, a variance-based global sensitivity index is calculated to determine the most sensitive ABA features associated with each track parameter.
