This paper presents a deep learning framework for analyzing on-board vibration response signals in infrastructure health monitoring. The proposed WaveletInception–BiGRU network uses a Learnable Wavelet Packet Transform (LWPT) for early spectral feature extraction, followed by one-dimensional Inception-Residual Network (1D Inception-ResNet) modules for multi-scale, high-level feature learning. Bidirectional Gated Recurrent Unit (BiGRU) modules then integrate temporal dependencies and incorporate operational conditions, such as the measurement speed. This approach enables effective analysis of vibration signals recorded at varying speeds, eliminating the need for explicit signal preprocessing. The sequential estimation head further leverages bidirectional temporal information to produce an accurate, localized assessment of infrastructure health. Ultimately, the framework generates high-resolution health profiles spatially mapped to the physical layout of the infrastructure. Case studies involving track stiffness regression and transition zone classification using real-world measurements demonstrate that the proposed framework significantly outperforms state-of-the-art methods, underscoring its potential for accurate, localized, and automated on-board infrastructure health monitoring.

Rail grinding has been widely applied in railway networks worldwide to remove or prevent rolling contact fatigue (RCF) cracks. However, some concerns have arisen regarding grinding, that it may introduce initial damage to the rail and largely shorten the RCF life. This work aims to better understand the effect of grinding on the long-term degradation of in-service rails, particularly concerning White Etching Layer (WEL) and RCF cracks. Seven rail samples were selected and taken from the Belgian and Swedish railway networks, with different grinding histories, accumulated loads, and steel grades. The mechanical and microstructural properties of these samples were examined through the hardness test and optical microscopy. WEL and microcracks were observed in both ground and non-ground rails, suggesting that rail grinding does not create additional defects nor negatively impact the rail surface after long-term service. Macrocracks were observed only in rail samples that had undergone zero or a single grinding cycle, confirming the beneficial role of rail grinding in mitigating RCF cracks. Ratcheting is the dominant crack initiation mechanism under the examined conditions, while WEL may also contribute to crack formation, given that macrocracks predominantly occur at the transition between the WEL and the pearlite.

This paper investigates the growth and treatment of a major type of rail rolling contact fatigue (RCF) known as head checks (HCs). The analysis is based on extensive field data of 212 curved tracks made of R260 steel across the entire Belgian railway network. The HC crack depth was mainly measured by eddy current testing. The growth rates of HCs are analysed in relation to the curve radius, annual traffic load, and rail wear. The key findings are as follows: 1) Tracks with radii between 750 and 1000 m exhibit the highest HC growth rate of about 1.5 mm per 100 million gross tons (MGT) and the largest occurrence probability of about 25 %. 2) A counterintuitive result is that the HC growth per MGT is higher on lines with lower annual traffic loads, consistent with the trend observed in rail wear rates. 3) The artificial wear methods to control RCF, such as preventive grinding, should consider annual traffic load and service time, rather than solely accumulated tonnage, as is the current practice. Based on these findings, a new method is proposed to estimate the magic wear rate for the Belgian railways, which can serve as input for optimising grinding operations to mitigate HCs.

This paper proposes an onboard measurement technology that combines a Laser Doppler Vibrometer (LDV) and an Axle Box Accelerometer (ABA) to approximate the dynamic load-response relationship of railway tracks. Unlike existing track-side and onboard technologies, this paper eliminates the need for load measurement, estimation, or control, enabling continuous measurements under operational conditions. The LDV mounted on the moving vehicle captures the track vibration response contactlessly, while the wheel vibration measured by ABA is used to directly represent the dynamic vehicle load. The LDV and ABA signals are combined to approximate the load-response relationship in the frequency domain. Experimental validation on a vehicle-track test rig demonstrates the effectiveness of the developed system at different speeds. Further comparisons with the hammer test result confirm its ability to capture the local dynamic properties of consecutive track segments along a railway track and also its superiority in measurement efficiency. This paper offers a promising solution for monitoring railway tracks on a large scale and allowing prescriptive maintenance of rail infrastructures.

Axle box acceleration (ABA) measurements can be used for continuously monitoring rail infrastructure and detecting rail surface defects such as squats. However, accurately detecting squats is challenging due to their short-duration responses and low occurrence in ABA signals, particularly for light squats that exhibit subtle ABA responses. To address this challenge, we propose using a spiking neural network (SNN) with time-varying weights to enhance the detection performance of rail squats based on ABA measurements. Our approach employs a simple SNN architecture without hidden layers, trained using a method that combines genetic algorithms, k-fold cross-validation, and multi-start gradient-based approach to optimise hyperparameters and weights. The proposed methodology demonstrates competitive accuracy compared to other state-of-the-art SNN-based methods on UCI benchmarks for both binary and multi-class nonlinear problems. Part of its advantages include higher efficiency with a simpler architecture and training approach that reduces computational times while achieving effective spatiotemporal pattern detection. As shown by real-field measurements from Dutch and Swedish railways in anomaly detection, it effectively captures subtle changes in light squat defect responses in ABA signals and achieves a detection performance of 100% for severe squat defects and over 93% for light squat defects. Furthermore, we show that the spike responses, postsynaptic potentials, and membrane potentials can be used as a new way to explain and analyse the ABA signals. The proposed method using time-varying weights highlights a correspondence with the physical problem and offers an ability to capture sudden and subtle changes in the responses, which is crucial, particularly for detecting defects in their early stages.

Transition zones in railway tracks often degrade faster than other locations, yet traditional health assessments rely on infrequent track geometry measurements, limiting early detection of dynamic changes. This research presents an approach for more frequent evaluation of transition zone health by integrating data sources from multiple monitoring technologies: track geometry, interferometric synthetic aperture radar (InSAR), and axle box acceleration (ABA). Missing InSAR data are addressed through spatio-temporal interpolation, and track longitudinal levels are predicted using a hybrid neural model that includes a hybrid convolutional neural network (CNN) with gated recurrent units (GRU) network and a hybrid CNN with a long short-term memory (LSTM) network. The models fuse historical and interpolated data from InSAR and ABA, enabling higher-frequency insights. A novel key performance index (KPI) based on predicted longitudinal levels is proposed to quantify track condition. The framework is validated on a transition zone at a railway bridge between Dordrecht and Lage Zwaluwe in the Netherlands. Results show that the hybrid model outperforms standalone methods and offers a good balance between accuracy and computational efficiency. The proposed approach enables earlier detection of irregularities, supporting prescriptive maintenance decisions.

Nowadays, rolling stock can be equipped with high-frequency vibration sensors to continuously monitor rail infrastructures and detect defects. These moving sensors measure at high speeds and sampling frequencies, generating a massive amount of data that covers each track position with very short signal durations. These data contain a variety of dynamic and transient responses that vary significantly along the track and are affected by noise. This leads to a large amount of unlabeled and noisy data, complicating the extraction of dynamic responses for effective anomaly detection. To address these challenges, this paper proposes an unsupervised representation learning methodology to automatically capture and extract characteristic features of dynamic responses that reflect the conditions of rail infrastructures. The unsupervised nature allows exploratory analysis of high-frequency vibration signals when prior knowledge or reference information about infrastructure conditions is unavailable or very limited. A collaborative optimization process that synchronizes empirical mode decomposition (EMD) with a convolutional autoencoder (CAE) is presented. The EMD level is tuned to remove noise while preserving effective vibration responses. The CAE is trained using demodulated signals that are considered normal to generate representations that ensure reconstruction quality and differentiate between normal and abnormal conditions. Furthermore, a Gaussian mixture model is used to showcase the effectiveness of the learned representations for rail infrastructures. Applied to validated axle box acceleration data for rail defect detection and train-borne laser Doppler vibrometer data for rail fastener monitoring, our method outperforms other variants of autoencoder-based models and the wavelet-based CAE in accurately identifying the conditions. It achieves an average improvement of 16% with the axle box acceleration data and 21% with the laser Doppler vibrometer data.

Modern smartphones, widely available and equipped with multiple sensors, offer the potential for railway infrastructure condition monitoring through mobile crowdsensing without disrupting operational railway services. This paper investigates key factors influencing the use of smartphone accelerometers for railway track quality assessment and highlights the associated challenges. Differences in accelerometer characteristics across three budget smartphones, including variations in sampling frequencies and dynamic sensitivity, are analyzed. A case study conducted on in-service passenger trains examines the effects of operational conditions, specifically vehicle speed and smartphone placement within the car body, on the frequency content and magnitude of recorded signals. Additionally, this study compares smartphone measurement results with a conventional track quality index derived from track geometry measured by specialized vehicles. Promising correlations are observed between the standard deviation of the smartphone vertical acceleration signals and the longitudinal level-based track quality index, demonstrating the potential of smartphones to assess track quality and detect local anomalies. However, variations caused by vehicle speed and smartphone placement pose challenges for standardizing track quality assessments. These findings highlight the potential of mobile crowdsensing for railway infrastructure monitoring while emphasizing the need for strategies to address variability in operational conditions and device characteristics.

The growing volume of available infrastructural monitoring data enables the development of powerful data-driven approaches to estimate infrastructure health conditions using direct measurements. This paper proposes a deep learning methodology to estimate infrastructure physical parameters, such as railway track stiffness, using drive-by vibration response signals. The proposed method employs a long short-term memory (LSTM) feature extractor accounting for temporal dependencies in the feature extraction phase, and bidirectional long short-term memory (BiLSTM) networks to leverage bidirectional temporal dependencies in both the forward and backward paths of the drive-by vibration response in the condition estimation phase. In addition, a framing approach is employed to enhance the resolution of the monitoring task to the beam level by segmenting the vibration signal into frames equal to the distance between individual beams, centering the frames over the beam nodes. The proposed LSTM-BiLSTM model offers a versatile tool for various bridge and railway infrastructure conditions monitoring using direct drive-by vibration response measurements. The results demonstrate the potential of incorporating temporal analysis in the feature extraction phase and emphasize the pivotal role of bidirectional temporal information in infrastructure health condition estimation. The proposed methodology can accurately and automatically estimate railway track stiffness and identify local stiffness reductions in the presence of noise using drive-by measurements. An illustrative case study of vehicle–track interaction simulation is used to demonstrate the performance of the proposed model, achieving maximum mean absolute percentage errors of 1.7% and 0.7% in estimating railpad and ballast stiffness, respectively.

This paper presents a new approach to simulate forward and inverse problems of moving loads using physics-informed machine learning (PIML). Physics-informed neural networks (PINNs) utilize the underlying physics of moving load problems and aim to predict the deflection of beams and the magnitude of the loads. The mathematical representation of the moving load considered involves a Dirac delta function, to capture the effect of the load moving across the structure. Approximating the Dirac delta function with PINNs is challenging because of its instantaneous change of output at a single point, causing difficulty in the convergence of the loss function. We propose to approximate the Dirac delta function with a Gaussian function. The incorporated Gaussian function physical equations are used in the physics-informed neural architecture to simulate beam deflections and to predict the magnitude of the load. Numerical results show that PIML is an effective method for simulating the forward and inverse problems for the considered model of a moving load.

A primary challenge of physics-informed machine learning (PIML) is its generalization beyond the training domain, especially when dealing with complex physical problems represented by partial differential equations (PDEs). This paper aims to enhance the generalization capabilities of PIML, facilitating practical, real-world applications where accurate predictions in unexplored regions are crucial. We leverage the inherent causality and temporal sequential characteristics of PDE solutions to fuse PIML models with recurrent neural architectures based on systems of ordinary differential equations, referred to as neural oscillators. Through effectively capturing long-time dependencies and mitigating the exploding and vanishing gradient problem, neural oscillators foster improved generalization in PIML tasks. Extensive experimentation involving time-dependent nonlinear PDEs and biharmonic beam equations demonstrates the efficacy of the proposed approach. Incorporating neural oscillators outperforms existing state-of-the-art methods on benchmark problems across various metrics. Consequently, the proposed method improves the generalization capabilities of PIML, providing accurate solutions for extrapolation and prediction beyond the training data.

This paper proposes a novel framework for simulating the dynamics of beams on elastic foundations. Specifically, partial differential equations modeling Euler–Bernoulli and Timoshenko beams on the Winkler foundation are simulated using a causal physics-informed neural network (PINN) coupled with transfer learning. Conventional PINNs encounter challenges in handling large space–time domains, even for problems with closed-form analytical solutions. A causality-respecting PINN loss function is employed to overcome this limitation, effectively capturing the underlying physics. However, it is observed that the causality-respecting PINN lacks generalizability. We propose using solutions to similar problems instead of training from scratch by employing transfer learning while adhering to causality to accelerate convergence and ensure accurate results across diverse scenarios. The primary contribution of this paper lies in introducing a causality-respecting PINN loss function in the context of structural engineering and coupling it with transfer learning to enhance the generalizability of PINNs in simulating the dynamics of beams on elastic foundations. Numerical experiments on the Euler–Bernoulli beam highlight the efficacy of the proposed approach for various initial conditions, including those with noise in the initial data. Furthermore, the potential of the proposed method is demonstrated for the Timoshenko beam in an extended spatial and temporal domain. Several comparisons suggest that the proposed method accurately captures the inherent dynamics, outperforming the state-of-the-art physics-informed methods under standard L²-norm metric and accelerating convergence.

A train-borne laser Doppler vibrometer (LDV) directly measures the dynamic response of railway track components from a moving train, which has the potential to complement existing train-borne technologies for railway track monitoring. This paper proposes a holistic methodology to characterize train-borne LDV measurements by combining computer-aided approaches and real-life measurements. The focus is on the speed-dependent characteristics because the train speed affects the intensity of railway sleeper vibrations and the intensity of speckle noise, which further affects the quality and usability of the measured signals. First, numerical models are established and validated to simulate sleeper vibrations and speckle noise separately. Then, a vibration–noise separation method is proposed to effectively extract speckle noise and structural vibrations from LDV signals measured at different speeds. The parameters of the separation method are tuned using simulation signals. The method is then validated using laboratory measurements in a vehicle-track test rig and applied to field measurements on a railway track in Rotterdam, the Netherlands. Further, the speed-dependent characteristics of train-borne LDV measurement are determined by analyzing the competition between sleeper vibrations and speckle noise at different speeds. Simulation and measurement results show that an optimal speed range yields the highest signal-to-noise ratio, which varies for different track structures, measurement configurations, and operational conditions. The findings demonstrate the potential of train-borne LDV for large-scale rail infrastructure monitoring.

A transfer function (TF) is an effective representation of the load-response relationship of railway track structures. To fill the gap in measuring track structure TFs over a wide frequency range from a moving vehicle, we develop a TF measurement system and the associated TF estimation methodology. Accelerometers are utilized to estimate the dynamic vehicle load to track structures, and a laser Doppler vibrometer (LDV) is used to scan track structures and measure their vibration response. First, operational modal analysis is applied to vehicle impact response over joints to identify its modal parameters, which support the estimation of dynamic wheel-rail forces from vehicle vibrations. This combination eliminates the need to pre-define the vehicle stiffness, vehicle damping, and vehicle body mass and enables the vehicle parameters to be updated under operational conditions. Meanwhile, a signal processing method is applied to LDV signals to reduce speckle noise and compensate for the effect of vehicle vibration. Then, a continuous track structure is segmented into distributed sections, and a TF is estimated for each track section using the estimated wheel-rail force as input and the extracted track vibration as output. We validate the methodology in a vehicle-track test rig on different track sections (with or without joints) and at different speeds (from 8 km/h to 16 km/h). The results are further compared with trackside measurements and hammer tests. We demonstrate that the track vibrations extracted from the LDV signals are consistent with those measured by trackside accelerometers. The shapes and resonance frequencies of the estimated TFs are in good agreement with those measured from the hammer tests in the frequency range of 200–800 Hz. The developed system captures differences in the TFs between different track sections, suggesting its potential to be used for structural health monitoring of railway tracks.