Development of a railway subgrade condition assessment method using a sparse autoencoder approach

Track subgrade condition is one of the major aspects that influences overall track support conditions and railway dynamic response. However, current vehicle-based railway monitoring solutions do not provide reliable data on this, mostly due to an overall focus on the more accessible superstructure elements (e.g., rails, sleepers or the ballast layer). A new methodology to assess railway track support condition is currently under development and assessment, which is based on modal analysis of the characteristic frequencies of the multi-element system composed by an instrumented rail vehicle and the railway infrastructure under assessment. During this process, an unsupervised data-driven learning tool was developed to apply the proposed track monitoring methodology, enabling the automatic extraction of relevant features from the collected data. This tool was implemented to improve the reliability of the proposed methodology when working with information either from experimental data or numerical simulations. This paper provides an overall description of this track monitoring methodology and the current unsupervised data-driven tool. It also presents two cases-studies that were used to assess the developed solution, in terms of the theoretical concepts behind the methodology and the data processing tool itself. The first case-study is based on simulated data obtained from a numerical model created in Simpack ® , and the second is based on experimental data. The obtained results demonstrate the overall potential of this methodology to provide a valuable method to enable adequate subgrade condition monitoring.