In the world of today, railway networks are essential transportation means for both passengers and freight. The increasing success of high-speed rail and the demand for sustainable development has led to tremendous growth in railway traffic each year all over the world. Without a doubt, the increased vehicle speed, axle loads, load cycles, and traffic loads, amongst other factors, have a detrimental effect on the deterioration of the track quality and increase the need for maintenance practices. In ballasted tracks, which is the most common track type for railway networks, the largest chunk of maintenance cost is spent on the ballast layer. To ensure optimized maintenance practices and cost; and reduce significant damage occurrences in track structures, research on non-destructive means of monitoring ballast condition is therefore very necessary.

The present study involves the monitoring of the track systems’ health condition by using measured vibration behavior. Owing to the geomechanical properties of the ballast particles, it is necessary to consider the non-linear stress-strain behavior of the ballast layer for optimal settlement and stress prediction under varying loads. To adopt the model-based structural model updating approach for damage detection, the analysis of the system is, therefore, carried out in the time domain because of the material non-linearity. For the ballasted track, a 2D finite element (FE) model of the track across the traverse section is constructed by modeling the sleeper as a Timoshenko beam, the rails and rail pads as mass-spring elements, and the ballast layer as a series of elastoplastic tensionless springs. To consider the uncertainties involved in the time-domain analysis, the Bayesian theory is implemented using an enhanced Markov chain Monte Carlo (MCMC) scheme. Several damage scenarios of the ballast layer were simulated under laboratory conditions to verify all developed algorithms and schemes and measurements collected from impact hammer tests.

The first stage of this research involves a feasibility study on damage detection of the ballast by utilizing a time-domain model updating of the ballast layer with the introduction of a simple power-law (Ludwik law) to model the ballast’s stress-strain behavior in the vertical direction. With several numerical simulations and experimental verification, it was observed that ballast damage location and severity could be accurately identified. However, to remove some of the limitations observed, a more compact and modified form of the Ludwik law and a novel comprehensive stopping criteria scheme for the enhanced MCMC were developed to ensure better efficiency of the damage detection methodology.

In the next stage of work, the problem of optimal sensor placement for system identification and damage detection in structural systems is addressed by developing a robust method based on Bayesian theory. The enhanced sequential sensor placement (ESSP) algorithm also efficiently addresses the computational bottlenecks that may arise when many observable locations are considered for candidate configurations, while sensor redundancy problem in finely meshed models was addressed by considering the spatial correction of prediction errors at measurable locations and prescribing a minimum sensor interval. The ESSP algorithm shows improvements in the sensor configuration's optimality with the proposed method relative to the conventional methods when implemented on model updating scheme for ballasted tracks. The algorithm also proposes configurations to maximize the information gain from a given number of sensors.

Towards ensuring an effective maintenance regime of ballasted railway tracks, which involves the determination of differential settlement, track support stress, and stiffness, gaining a good insight into the mechanical behavior and the strain-hardening property of ballast, is non-trivial. Hence, the vertical stress-strain behavior of the ballast layer, which is primarily responsible for irrecoverable settlements in tracks, is studied by developing simple yet applicable elastoplastic models from well-known material laws. The dynamic ballast behavior is then investigated by using the enhanced Bayesian MCMC to handle the uncertainties and the Bayesian MCMC model class selection method to identify the most plausible models between the developed ones and the linear-elastic model, considering various ranges of impact forces under laboratory conditions. The results obtained demonstrate the importance of the non-linear ballast model and the applicability of the selected ballast model in field track monitoring.

For proper system identification, model-based structural health monitoring, and similar Bayesian inference studies, the importance of the system’s structural model and the prediction error model, is non-trivial. As against the common practice of assuming uncorrelated system responses, it was observed that the fine FE model mesh and high sampling frequency in the measurement test warrants the consideration of correlation in both spatial and temporal domains for the prediction error. Herein, the variance of the zero-mean Gaussian model of the prediction error is developed utilizing spatio-temporal correlation matrices, established from several distinct spatial and temporal correlation structures. Using the measured in situ sleeper vibration, the Bayesian MCMC model class selection method is then used to identify the most plausible Spatio-temporal correlation matrix model class from among 13 developed classes. Results show better health condition prediction performance when the prediction error correlation is considered.

Finally, for the adoption of the developed ballast monitoring methodology in the field and for real-time analysis, it is important to drastically the high memory and computation requirement involved, which stems from the non-linear FE model as well as the intensive enhanced MCMC scheme. To this end, Kriging surrogate metamodels are constructed to predict system responses and replace the FE models. In high-dimensional high-fidelity models, Kriging surrogates are usually avoided because of the complexity. Herein, six simplified Kriging surrogate schemes are developed with formulations for high dimensional structural responses. The schemes were compared for ballasted track damage detection based on prediction accuracies and prediction variance to identify the most suitable under a given load condition. With the optimal Kriging metamodel, about 90% of the computational time could be saved in Bayesian MCMC-based ballast damage detection analyses whilst maintaining a satisfactory accuracy level.