We explore a data-driven approach for monitoring rail infrastructure from the dynamic response of a train in revenue-service. Presently, track inspection is performed either visually or with dedicated track geometry cars. In this study, we examine a more economical approach where track inspection is performed by analyzing vibration data collected from an operational passenger train. The high frequency with which passenger trains travel each section of track means that faults can be detected sooner than with dedicated inspection vehicles, and the large number of passes over each section of track makes a data-driven approach statistically feasible. We have deployed a test-system on a light-rail vehicle and have been collecting data for the past two years. The collected data underscores two of the main challenges that arise in train-based track monitoring: the speed of the train at a given location varies from pass to pass and the position of the train is not known precisely. In this study, we explore which feature representations of the data best characterize the state of the tracks despite these sources of uncertainty (i.e., in the spatial domain or frequency domain), and we examine how consistently change detection approaches can identify track changes from the data. We show the accuracy of these different representations, or features, and different change detection approaches on two types of track changes, track replacement and tamping (a maintenance procedure to improve track geometry), and two types of data, simulated data and operational data from our

We propose a novel recovery algorithm for signals with complex, irregular structure that is commonly represented by graphs. Our approach is a generalization of the signal inpainting technique from classical signal processing. We formulate corresponding minimization problems and demonstrate that in many cases they have closed-form solutions. We discuss a relation of the proposed approach to regression, provide an upper bound on the error for our algorithm and compare the proposed technique with other existing algorithms on realworld datasets.

No abstract

Collecting vibration data from revenue service trains could be a low-cost way to more frequently monitor railroad tracks, yet operational variability makes robust analysis a challenge. We propose a novel analysis technique for track monitoring that exploits the sparsity inherent in train-vibration data. This sparsity is based on the observation that large vertical train vibrations typically involve the excitation of the train's fundamental mode due to track joints, switchgear, or other discrete hardware. Rather than try to model the entire rail profile, in this study we examine a sparse approach to solving an inverse problem where (1) the roughness is constrained to a discrete and limited set of "bumps"; and (2) the train system is idealized as a simple damped oscillator that models the train's vibration in the fundamental mode. We use an expectation maximization (EM) approach to iteratively solve for the track profile and the train system properties, using orthogonal matching pursuit (OMP) to find the sparse approximation within each step. By enforcing sparsity, the inverse problem is well posed and the train's position can be found relative to the sparse bumps, thus reducing the uncertainty in the GPS data. We validate the sparse approach on two sections of track monitored from an operational train over a 16 month period of time, one where track changes did not occur during this period and another where changes did occur. We show that this approach can not only detect when track changes occur, but also offers insight into the type of such changes.

We present a data fusion approach for enabling data-driven rail-infrastructure monitoring from multiple in-service trains. A number of researchers have proposed using vibration data collected from in-service trains as a low-cost method to monitor track geometry. The majority of this work has focused on developing novel features to extract information about the tracks from data produced by individual sensors on individual trains. We extend this work by presenting a technique to combine extracted features from multiple passes over the tracks from multiple sensors aboard multiple vehicles. There are a number of challenges in combining multiple data sources, like different relative position coordinates depending on the location of the sensor within the train. Furthermore, as the number of sensors increases, the likelihood that some will malfunction also increases. We use a two-step approach that first minimizes position offset errors through data alignment, then fuses the data with a novel adaptive Kalman filter that weights data according to its estimated reliability. We show the efficacy of this approach both through simulations and on a data-set collected from two instrumented trains operating over a one-year period. Combining data from numerous in-service trains allows for more continuous and more reliable data-driven monitoring than analyzing data from any one train alone; as the number of instrumented trains increases, the proposed fusion approach could facilitate track monitoring of entire rail-networks.