Vibration Measurement and Monitoring in Railway Vehicles

Abstract

The main purpose of this paper is to present a comprehensive and systematically organized overview of the current state of vibration monitoring and measurement techniques used in railway stock. It aims to raise awareness of significant technological developments in recent years and their practical applications. The scope of the analysis is strongly informed by established European norms, fundamental research efforts across the continent, and the practical needs of the railway sector. Last but not least, we hope this paper serves as a valuable reference point for engineers, researchers, and decision-makers working within the complex context of railway system design, maintenance, and evolving regulations. For effective monitoring of railway vehicle vibrations, a combination of specialized measurement methods and system architectures is recommended. These approaches are carefully developed to capture the dynamic responses of critical components of the railway vehicle, thereby providing invaluable data. This information is essential for thorough condition monitoring, improved ride comfort, and a deeper structural understanding of vehicle quality throughout its lifecycle.

1. Introduction

The main goal of this paper is to provide a systematic and integrated description of the status quo of vibration monitoring and measurement methods onboard railway vehicles. In this review, we intentionally emphasize recent key advances in technology as well as heterogenous applications stemming from these advances. The scope of the analysis is set in relation to the defined European framework, important projects developed on the continent and the needs of railway services that must be met. Finally, this paper is designed to be a key source for engineers, researchers and professionals working on the complex interplay between the design, operation and regulation of railway systems.

More specifically, the objectives of this article to be accomplished are as follows:

• Summarize the state-of-the-art of railway vehicle vibration-monitoring systems examining onboard integrated solutions in rolling stock and systems at ground level distributed along the track.
• Taxonomize the measurement literature and critically assess various measurement techniques (e.g., advanced accelerometer-based systems, novel wireless sensor networks for improved data acquisition, and advanced simulation-based techniques with predictive capabilities).
• Examine the various signal-processing methods necessarily employed in the interpretation of raw vibration data. This reporting will span classic known methods including RMS (root mean square) processing and FFT (fast Fourier transform) processing, to state-of-the-art AI-derived algorithms facilitating deeper understanding and intelligent correlation of investigative sound information.
• Pinpoint and articulate the major bottlenecks that continue to impede the efficient deployment and smooth performance of these advanced monitoring solutions. This encompasses challenges to do with the optimized position of sensors, data processing efficiency, signal interpretation, and the intricate process of system integration of the equipment into the railway infrastructure.
• Explore the complex body of regulations and standards that prescribe permissible vibration limits in railways, with special attention to the European context and the role these standards play in shaping the design and implementation of technical solutions.
• Explore a spectrum of practical applications, potential impacts, and future directions for vibration monitoring within the field of railway engineering, providing insights into emerging trends and prospective advancements.

2. Overview of Vibration Measurement and Monitoring Framework

2.1. Background and Motivation

The railway sector has long been a key building block of sustainable and efficient transport, and this has been true around the world [1]. The rise of rail in an age of rapidly advancing global urbanization and heightened environmental awareness; rail is being seen more and more as an essential way to reduce carbon and relieve unrelenting road chaos [2].

The vibrations generated by a railway vehicle running are not limited to the vehicle itself, but instead are transmitted throughout large areas of the vehicle’s structure and to nearby areas [3]. This “far-field propagation” can have far-reaching implications for many aspects such as passenger comfort, the structural health of the vehicles, the structural health of the infrastructure of the rolling stock as well as vehicle infrastructure and perhaps, even the health of the surrounding population near rail corridors. These dynamic vibrations usually result from a complex interaction of several influencing sources: intrinsic track irregularities, complex dynamic vehicle behaviors, the mechanism of interaction between wheel and rail, and soil/substructure properties.

The compelling drive for developing sophisticated and advanced vibration-monitoring systems in railway applications stems from several critical motivations:

• The comfort of passengers is the most important issue. High levels of vibration on bodies cause considerable discomfort, negatively influence the fatigue of passengers during longer travelling and even instigate health problems in extreme cases, especially during longer or high-speed journeys [4]. To combat this problem, regulations such as SR-EN 12299 [5] and ISO 2631-1 [4] have clearly defined acceptable levels of whole-body vibration exposure, with compliance to these standards now a key legal and operational obligation for railway infrastructure managers.
• Vibration monitoring plays a pivotal role in vehicle safety and maintenance. Amplitude instability is quite often an early symptom of an upcoming mechanical failure, including wheel flats, progressive suspension deterioration or wary irregularity of bogies [6]. Ongoing monitoring for these vibrational signals allows the introduction of a predictive maintenance approach that minimizes the expense of “dead” time in service and, most importantly, the cost of catastrophic failures, which may have significant safety implications.
• Train networks are also greatly served by infrastructure health. Vibrations generated from running train vehicles propagating to the track structures and/or the subballast/subgrade layer can significantly worsen the degradation process, resulting in unwanted issues such as track settlement, premature ballast breakage, fatigue damage in bridge areas, and so forth [1]. With proactive vibration monitoring, these emerging problems can be quickly identified and allow condition-based maintenance strategies to be deployed to key infrastructure assets.
• Urban and environmental impact is on the rise. Ground-borne vibration, as an inevitable result of the operation of railways, has the potential to cause negative impacts to surrounding buildings, sensitive equipment installed within buildings and to disturb human activity in urban habitations. In congested urban environments, it has become a major public concern and stricter measures are required, emphasizing the urgency of adopting effective vibration mitigation measures.
• Design optimization represents a continuous process. The comprehensive data amassed from advanced vibration-monitoring systems serve as invaluable input for the design and simulation phases of new rolling stock and infrastructure components [7]. This iterative feedback loop directly supports the development of future railway vehicles engineered with demonstrably improved dynamic behavior and a significantly reduced environmental footprint.
• In the highly regulated railway sector, the importance of regulatory compliance is absolute. Compliance with stringent standards and regulations concerning permissible vibration levels and passenger comfort is a must. State-of-the-art vibration-monitoring systems are essential tools that help railway operators ensure their vehicles conform to these rigorous requirements at all times [4].

In the past, vibration monitoring often depended on regular manual measurement with a portable sensor. Although these approaches provided useful insights for diagnosis, they were inherently void of the necessary fine temporal resolution and continuity required for modern applications. The last two decades have seen a drastic change in the vehicles from the disconnected multiple onboard sensor networks to the vehicle internal wireless and internet-connected systems for real-time data transfer [8]. This evolution has been further bolstered by utilizing advanced signal-processing methodologies (such as wavelet transforms, Hilbert–Huang analysis), exploiting the advanced machine learning and artificial intelligence (AI) paradigms for robust anomaly detection and predictive analytics, and leveraging the effective usage of digital twins that simulate the vehicle–track–environment interactions in a detailed manner [9]. None of these advances would be possible without the collective innovation that has disrupted vibration monitoring from being solely a reactive diagnostic tool, to a proactive, intelligent and deeply embedded element of a modern railway’s operational life-cycle.

And several important European research projects have had a decisive influence on this area. Initiatives such as SILVARSTAR, RIVAS and FINE2 have subscribed to the strategic priority to fulfil the ambitious objectives of the Shift2Rail program for which monitoring of vibrations had to be necessarily defined. Such dedicated joint research projects are focusing in particular on developing standardized tools for comprehensive vibration prediction and evaluation, with the aim of harmonizing measurement protocols between different member states and promoting research towards railway systems that are quieter, more comfortable and thus more sustainable. Notably, the SILVARSTAR Deliverable D1.1 clearly articulates the dire necessity for a modular, interoperable vibration-prediction system that can incorporate various field data, advanced numerical simulations, and accurate empirical models on the fly [10,11,12]. These are powerful tools for assisting infrastructure planning for both policy support and regulatory compliance, and increasing public acceptance for new and intensified rail construction.

2.2. Regulatory and Standardization Framework

The complex framework of vibration-monitoring-system evolution and implementation on railway vehicles is greatly influenced by a clearly defined, strict regulatory and standardization landscape. These protocols are established to ensure that unforeseeable levels are always kept within acceptable and safe limits, so as to safeguard passengers, infrastructure and the surrounding environment, including prospective future vibration-sensitive occupants’ buildings. In addition to safety, they contribute to common technical language and technical criteria and it is a precious instrument for engineers, operators and politicians that want to assess and compare system performance between regions and between different railway applications [4].

In this context, SR-EN 12299: Ride Comfort in Passengers is one of the reference standards in most European countries. This important standard describes objective numerical criteria for the comprehensive evaluation of ride comfort, primarily from whole-body vibration exposure. It provides detailed measurement instructions for placing the accelerometers on passenger seats and in the vehicle, and defines comfort indices such as the CRI (comfort ride index) and CAI (comfort acceleration index). These indexes are systematically analyzed based on the time-weighted root mean square (RMS) acceleration data, providing a measurable comfort metric [4]. In addition, the standard is closely harmonized to ISO 2631-1 [4], which gives general requirements for the evaluation of human exposure to whole-body motions. SR-EN 12299 [5] and ISO 2631-1 [4] combined provide the cornerstone for evaluating passenger comfort, first from the initial homologation of a new vehicle and then during the vehicle’s in-service monitoring.

While SR-EN 12299 is highly effective for evaluating broadband, low-frequency vibrations that cause general discomfort, it has known limitations in detecting the impact of transient or high-frequency events. The standard’s RMS-based weighting may not fully capture the sharpness of impacts from wheel flats, rail joints, or catenary arcing, which are often perceived as a “jolt” rather than a sustained vibration. Furthermore, the standard’s focus is on whole-body vibration, potentially under-representing localized discomfort or high-frequency events felt in specific body parts. To address this, some studies and national practices are exploring frequency-dependent discomfort criteria. For example, research suggests human discomfort is most keenly felt in the 4–8 Hz range for vertical vibrations and can be particularly sensitive in certain body parts at different frequencies.

Adding to this, the ISO 2631 Series: Human Exposure to Vibration is a fundamental set of international standards that is tailored for mechanical vibration and shock. Key integral parts in this series are ISO 2631-1:1997 [4] (General requirements for whole-body vibration), ISO 2631-2:2003 [13] (Vibration in buildings, 1 to 80 Hz) and ISO 2631-4 [14] (Guidelines for the evaluation and measurement of human exposure to whole-body vibration in public transport systems). These standards define a set of frequency weighting curves (e.g., Wk, Wd) depending on the body axis, specify exposure limits for comfort, health and performance considerations, and provide procedures to combine vibration signals over both the time and frequency domains [13]. What makes them important is the capability to contribute as interpretative frameworks for data acquired by onboard sensors and for guiding the development of performing vibration counter-measures.

In addition, the ISO 14837-1: Ground-Borne Noise and Vibration [15], whose full title is “Mechanical vibration—Ground-borne noise and vibration arising from rail systems—Part 1: General guidance”, provides a general reference for the assessment of the environmental effect of railway ground-borne noise and vibration. The standard contains essential directions for the appropriate modeling and prediction of ground-borne vibration as well as guidelines for use in measuring protocols and recording reports. This criterion is especially relevant for urban rail systems, because in such systems, trains come close to residential or business areas where vibrations have a significant effect on buildings and sensitive equipment and therefore require careful control and monitoring.

Numerous projects funded by the European Union have greatly enhanced the development, application and wider acceptance of these standards mentioned above. For example, the SILVARSTAR (Shift2Rail) initiative has not only produced a modular vibration prediction tool but it also placed strong emphasis on the tool’s interoperability with currently used standards, and proposed a series of useful extensions to ISO 2631 that would allow better integration of simulation-based assessments. In particular, the RIVAS (Railway Induced Vibration Abatement Solutions) project was dedicated to the study of suitable mitigation strategies for ground-borne vibration. Identified outputs have been used to feedback directly into ISO and EN standards with models substantiated by extensive field trial, proving real-world applicability [12]. Finally, the FINE2 (Future Integrated Noise and Vibration Evaluation Platform) project dealt with the necessity to have harmonized description and methods for the appreciation of the effects all over the European territory, ultimately presenting an integrated solution for the evaluation of noise and vibration across the whole continent [16]. These joint developments have not only raised the state-of-the-art of vibration monitoring, but have also had a significant impact on policy shaping and legislation, official certification requirements, and detailed procurement specifications of rolling stock and track; both the procurement of new trains and the procurement of railway infrastructure.

In addition to these international and European standards, many countries also have their own national standards and legislation. Examples include DIN 45672 [17] and TA Lärm [18] in Germany, BS 6472 (vibration), Vibration Standards in the United Kingdom and FTA/FRA (transit and high speed rail) in the USA. These national texts tend to add various details such as thresholds that are specific to the local context, specific protocols of measurement, and effective implementation strategies that are adapted to local conditions and legal frameworks.

But a wider view reveals that as robust as European standards have become, other large railway jurisdictions such as the US and Japan all use different (but complementary) frameworks. The noise and vibration portion, in line with the recommendations of the U.S. Federal Railroad Administration (FRA) and Federal Transit Administration (FTA), provides guidance on the assessment of impact leading to a set of mitigation measures directed toward communities and buildings. Similarly, the Japanese Shinkansen network is successful because of high track standards and vehicle design, developed through decades of operational experience at high speed. If global best practices can be integrated into the comparative table, it would emphasize different priorities—e.g., environmental impact in the U.S. vs. passenger comfort/interoperability in Europe—and give a more externally validated step-by-step approach to vibration monitoring.

2.3. Technological Advances in Monitoring Systems

Vibration monitoring in railways has evolved dramatically over the past two decades from static, low-resolution checks to dynamic, intelligent systems capable of real-time decision-making. This evolution has been fueled by advances in sensors, wireless communication, analytics, and system integration.

What was once primarily characterized by manual measurements and offline data analysis has matured into a highly sophisticated ecosystem comprising real-time, intelligent, and extensively interconnected systems.

The evolution from rudimentary manual measurements to smart monitoring represents a significant advance. Historically, comprehensive vibration assessments were typically conducted using portable accelerometers and standalone data loggers, primarily during scheduled maintenance inspections or dedicated test campaigns. While these methods offered some diagnostic utility, they were inherently limited by their low temporal resolution, the laborious manual data handling, and an inability to reliably detect transient or rare events that could signify impending issues.

The important change to integrated onboard monitoring systems was a fundamental advancement. These systems now embed sensors directly into the structural components of the vehicle—strategically placed on bogies, axles, the chassis, and even within passenger seating areas—enabling continuous data acquisition throughout the vehicle’s normal operational service. This paradigm transition has facilitated in-depth longitudinal studies of both vehicle and track behavior, enabled early fault detection through continuous trend analysis, and allowed seamless integration with broader fleet management systems, leading to more efficient maintenance planning.

Modern systems use wireless sensor networks (WSNs) and IoT integration [13]. These systems eliminate the need for complex cabling, which significantly reduces installation costs and greatly enhances scalability for railway applications. The key features of these modern solutions include low-power accelerometers equipped with onboard processing capabilities, robust mesh networking protocols ensuring reliable data transmission, and increasingly, edge computing capabilities that allow localized anomaly detection directly on the vehicle itself.

When integrated with Internet of Things (IoT) platforms, these individual sensors become components of a much broader ecosystem [13]. This comprehensive ecosystem typically encompasses cloud-based data storage and analytics platforms, remote diagnostic capabilities with automated alerts, and full interoperability with other critical subsystems within the railway environment (e.g., braking systems, suspension control, HVAC). Prominent European projects such as SILVARSTAR and FINE2 have consistently underscored the paramount importance of developing modular, interoperable architectures that facilitate the seamless integration of vibration-monitoring data into existing, often complex, railway IT infrastructures.

Vibration signals are complex—nonlinear, noisy, and multi-scale. Tools like wavelet transforms, Hilbert–Huang decomposition, and spectral kurtosis help to extract meaningful patterns. Meanwhile, machine learning and deep learning algorithms handle tasks like anomaly detection, failure classification, and trend prediction [11]. For example, unsupervised clustering can reveal unknown failure modes, while recurrent neural networks (RNNs) forecast future states based on historical data.

In parallel, the rapid advancements in machine learning (ML) and deep learning (DL) algorithms have led to their increasing application for automated pattern recognition within vast datasets, robust anomaly detection and classification, and sophisticated predictive maintenance capabilities based on historical operational trends [19]. For instance, unsupervised clustering algorithms can effectively identify latent or emerging failure modes that might otherwise go unnoticed, while recurrent neural networks (RNNs) are adept at forecasting future vibration levels under varying operational conditions, providing invaluable insights for proactive management.

The concept of digital twins and simulation integration has gained considerable importance in railway engineering. A digital twin is a dynamic simulation model that mirrors a real train [11]. These models integrate onboard data with finite element and multi-body simulations, enabling real-time scenario testing and adaptive control. SILVARSTAR strongly supports this hybrid approach, combining measurement, modeling, and better prediction.

Digital twins enable real-time synchronization between measured field data and sophisticated simulation models. This powerful integration facilitates rigorous scenario testing for design optimization and comprehensive risk assessment. It establishes dynamic feedback loops that enable adaptive control mechanisms and refined maintenance planning. These digital models are typically constructed using advanced multi-body dynamics (MBD) and finite element method (FEM) techniques, and are meticulously calibrated using real-world field data acquired from onboard sensors [3,8,9,20].

While a powerful tool, the implementation of digital twins faces significant challenges. These include the computational cost of running high-fidelity real-time simulations; the complexity of validating a virtual model against a dynamic physical asset; and the cost and effort required to integrate diverse data streams (e.g., vibration, speed, GPS, temperature) into a unified, synchronized platform.

Finally, modern vibration monitoring is not a standalone tool—it is part of larger fleet management and maintenance decision systems. These platforms link vibration data with scheduling, logistics, and investment planning, enabling a shift from reactive to prescriptive maintenance.

2.4. Challenges in Vibration Monitoring

Although significant progress has been made both technologically and through demonstrated benefits, there still remains a variety of underlying challenges for the implementation and operation of easily applicable vibration-monitoring systems on railway vehicles. They arise from basic restrictions at the sensor level to more complicated problems of system-wide integration that finally reflect in the accuracy, reliability, and overall usability of the huge amount of data taken. It is of utmost importance to gain a deep appreciation of these challenges, since they offer clear direction on how to conduct our research and development activities in these essential areas.

Sensor placement and calibration remain two of the most technical and enduring challenges [3,21,22,23]. It is not trivial to define where to put sensors on a railway vehicle or on its infrastructure. The dynamic behavior of railway vehicles is extremely complex and the nature of the rail–wheel interaction is greatly affected by many factors such as the type of the vehicle (e.g., high-speed passenger train vs. heavy freight wagon), the configuration of its suspension system and the track geometry and condition. Sub-optimal or incorrect sensor locations may result in partial or distorted measurements, which can be particularly the case for strongly localized or anisotropic vibrations. Moreover, the problem of calibration drift over long periods, which may be caused by a variety of factors like strong temperature variations, gradual mechanical degradation of sensors, or electric perturbation in the harsh railway environment, can affect the accuracy and robustness of the acquired track data. As such, the relevant literature in this domain stresses the importance of comparative studies on the optimization of sensor positions, many times using advanced statistical or modal analysis, to secure more robust and accurate deployments.

Secondly, and of no less importance, is that the raw data volume and signal complexity a modern monitoring system generates, is otherwise impossible to grapple with [1,11]. When many sensors are placed on a single vehicle or throughout a system, this system will invariably generate massive quantities of high-frequency data. Such an abundance of data generation results in several interrelated challenges: the limited storage capacity and network bandwidth, especially for wireless communication systems; the strict real-time processing requirements, in particular for resource-constrained edge devices that perform real-time data analysis on site; and the complexity of the signals themselves, which possess characteristics such as non-stationarity, multi-modality, and high levels of noise [19]. Advanced signal-processing tools like wavelet transform for multi-resolution analysis and Hilbert–Huang decomposition for non-linear signals, although important for meaningful feature extraction, are computationally expensive and often need careful expert tuning to deliver an optimal summary of pertinent characteristics.

Concrete examples of this challenge include a typical high-frequency accelerometer on a single bogie generating over 10 MB/h of data, leading to a total data volume of several terabytes per month for a large fleet. To manage this, effective strategies are needed, such as data compression (e.g., achieving a 10:1 ratio for a 100 MB file), localized edge computing to process data at the source, and a defined upload cadence (e.g., hourly updates with a larger nightly sync). Disconnection buffering is also critical, requiring sufficient onboard storage to retain data for up to 24 h of operation to ensure no data is lost during communication outages.

And yet interpretation and decision-making based on good data is still very much an unsolved problem. The inherent uncertainty of vibration signals and the fact that a given vibration pattern might come from different sources—a local track defect, a fault in a particular vehicle component or object-on-track—would imply that such information is only of limited value. This ambiguity makes fault-diagnosis, and hence decisions, extremely hard to make accurately. Although the machine learning-based methods provide high potential in automatic pattern recognition, a common issue is many black box models themselves are not intrinsically interpretable and are hard for human practitioners to fully trust them in safety-critical railway scenarios. The emerging field of explainable artificial intelligence (XAI) is aiming at addressing this issue and is designing approaches to offer clear explanations for AI predictive models, albeit with early stages of practical application in railway systems [1,11]. To mitigate this “black-box” concern, future systems must integrate XAI techniques such as feature importance analysis (e.g., identifying kurtosis or specific speed-synchronous harmonics as the primary drivers for a classification decision) or rule extraction to provide transparent, interpretable reasons for an anomaly detection alert, thereby building trust with human operators and reducing misclassification risk.

After that, environmental and operational variability add to the complexity. The level of vibrations in railway systems is dominantly affected by a variety of external factors including, but not limited to, the ambient weather conditions (e.g., heavy rain or snow, extreme temperatures), the natural variation in track conditions across a particular route (e.g., changes in ballast stiffness and rate), as well as the variation in several operational parameters (e.g., changes in train speed, braking events, different loading conditions). This high degree of inconstancy makes it particularly difficult to maintain reliable baseline vibration profiles or to confidently detect true anomalies. It also significantly complicates the validation of simulation models, which need to be carefully formulated in consideration of the wide range of possible operation conditions to better represent the reality [3,9,22,23,24,25,26]. In order to alleviate this type of problem, recent research explored countermeasures through domain adaptation and invariant feature learning that, respectively, aim to train models that are robust against speed, load and weather variations. It is vital to quantify the effect of such countermeasures, and while studies have demonstrated performance degradation with and without adaptation (i.e., 20% drop in accuracy or a 15% increase in accuracy with an adaptive model), these bring value depending on how effective they are.

Within the railway sector, system integration and interoperability are especially difficult tasks, as they continue to serve heterogenous fleets, consisting of different vehicle types and very often based on outdated systems. This complexity often increases the likelihood of failure to successfully integrate novel and advanced monitoring technologies. Problems stem from the diversity of file formats across different systems, the widespread absence of standardized application programming interfaces (APIs) for making data available, and frequently only limited support for piecemeal upgrading of systems. Acknowledging this dire need, the SILVARSTAR project focuses its attention on the development of interoperable, modular solutions adaptable to changing technology and regulatory trends. In the absence of such fundamental frameworks, the scalability and the long-term sustainability of advanced monitoring systems are greatly limited [11]. This challenge is exacerbated by the lack of standardization in data-sharing protocols, especially for legacy systems that often rely on proprietary formats. A common data schema and API, built on open standards, would be a major step toward enabling seamless data exchange between different systems and stakeholders, from vehicle manufacturers to infrastructure managers.

Last but not least, as vibration-monitoring systems start to integrate more and more, especially as the IoT and cloud become more and more popular, the fewer chances they have to be safe from the cluster of threats which are data security and privacy. In this environment, potential threats include data adulteration, which may cause dangerous false alarms; more dangerously, they may lead to critical missed anomalies; illicit penetration to sensitive operational data, which might jeopardize competitive advantages or safety procedures; and substantial privacy risk when dealing with personal data, in particular, with regard to passenger-data-monitoring scenarios [4]. As such, the need for strong and completely secure transmission and storage of data to ensure access only by authorized personnel is absolutely crucial, particularly when protecting key national sites. The integration of IoT and cloud services also introduces new cybersecurity risks, such as man-in-the-middle attacks on wireless communication channels or the compromise of cloud-based data repositories. This is particularly concerning for legacy systems with outdated security protocols, which can serve as a weak entry point into the wider network, underscoring the need for a comprehensive cybersecurity strategy.

3. Methods and Systems for Vibration Monitoring

The monitoring of vibration in railway vehicles in an effective way particularly benefits from the use of a combination of dedicated measurements, as well as modern system architectures. These various methods are carefully matched to achieve dynamic responses from several vital locations in the railway vehicle and to establish a treasure trove of data. Such information is crucial for the holistic evaluation of vehicle-condition monitoring, ride comfort, and structural health during the life of the vehicle, but also for the provision of appropriate warnings to minimize the negative effects of degradation on the components and to enhance the performance of the vehicle.

3.1. Onboard Measurement Systems

In-vehicle monitoring systems are a key component of the continuous vibration measurement process in railway applications that allow real-time or near real-time data to be transmitted from the moving vehicle [3,19,21,23]. These complex systems consist of several main components: firstly, sets of sensors positioned according to a decision-makers strategy, that translate physical changes (vibrations) to electrical signals; secondly, robust data-collecting devices, responsible for collecting and digitizing these electrical signals; and finally, communication modules for storage and further processing of the data.

Most of the onboard systems are based on accelerometers as these are the key sensors for railway vibration monitoring [27,28]. These instruments detect acceleration independently of the static field, a basic dynamical quantity. From such acceleration plots, important kinematic quantities such as displacement and velocity can then be derived by mathematical integration, thus enhancing the picture of the vehicle’s dynamic status. A variety of types of accelerometers are used, made to perform more general or specialized measurements. Piezoelectric (PZT) accelerometers, such as those commonly used due to their high sensitivity and wide frequency range are ideal to capture high-frequency vibrations that are typical at wheel–rail interface or structure resonances. Micro-electro-mechanical systems (MEMS) accelerometers, on the other hand, can be a small and cheap alternative, if there is a space or budget limitation for the deployment. So, for low-frequency, high-precision gauging (e.g., body movements of large vehicles, very low-frequency oscillation), servo-accelerometers are commonly preferred because of their high accuracy and stability.

It is critically important to allocate these accelerometers in Cartesian space if significant dynamic responses are to be properly measured. Incorrect positioning of sensors may result in uncomplete and mistaken data, in particular, due to the local character and the anisotropy of vibrations in the intricate structure of a railway vehicle [4,11,12]. Common measurement points and strategic positions on a railway vehicle are as follows:

• Bogie frames: Sensors mounted on the bogie frames are used to look at crucial dynamic phenomena such as bogie hunting, performance and wear of the components linked to the vehicle’s suspension system and to estimate forces resulting from the complex interaction between the wheel and the rail.
• Axle boxes: The mounting on the axle box helps in the detection of major wheel defects like wheel flat; barring defects which have a typical vibration signature; and other defects arising from the direct wheel–rail contact area at an early stage.
• Automotive body/carbody floor: Accelerometers installed on the carbody floor or on the passenger coach floor are used to evaluate the ride comfort in terms of passenger comfort according to standards such as SR-EN 12299, ISO 2631-1, as well as to assess the general vehicle dynamic stability over the entire operations period.
• Passenger seats: The ascertainment of head and body vibrations experienced by passengers in their sitting positions is the most direct method of assessment for compliance with severe comfort requirements from the human perception point of view.
• Suspension components: Sensors can be attached to primary/secondary suspension parts (e.g., springs, dampers) to monitor their performance and detect signs of wear, providing health indicators of the complete suspension system.
• Pantographs: Studied for the interaction between the pan and the overhead line (or catenary). In electric multiple units, the dynamic behavior of the pantograph with regard to the catenary has to be monitored as the excess of vibrations there can cause arcing, abrasion and even separation, affecting the power pick-up and the operational reliability.

Apart from accelerometers, there are other types of sensors that together form a full onboard monitoring system [3,21]. Strain gauges are often used to measure the stress and strain that a critical structural member, such as a bogie frame or axle, is subjected to. This gives important information about the fatigue life of these parts as well as clarity on the dynamic load distribution in operation. Displacement sensors (e.g., LVDTs or high-end laser displacement sensors), measure the relative displacements of different vehicle components. That is to say, they can monitor the motion between the carbody of the vehicle and the bogie, or within a primary suspension system or between the primary and secondary suspension systems, and enable a more detailed assessment of the performance and integrity of the suspension system. Lastly, rotational speed sensors (encoders) are very important for accurately associating vibration data with both the speed of the vehicle and the rotational speed of the vehicle’s tires. A further such dependence occurs on the vehicle’s wheel—this is of practical importance, especially to perform a correct identification and diagnostics of wheel-related defects—since numerous fail signatures are speed-related or they appear at the harmonics of the rotation frequency (

Table 1. Comparative performance of accelerometer technologies.

Sensor Type	Sensitivity	Bandwidth	Cost	Durability/Robustness
PZT accelerometers	High (ideal for small vibrations)	Wide (up to several kHz)	High	High (robust, often hermetically sealed
MEMS accelerometers	Moderate (suitable for general purpose)	Narrow (typically < 1 kHz)	Low	Moderate (susceptible to temperature and shock)
Servo-accelerometers	Very high (excellent for low frequencies)	Low (down to 0 Hz)	Very High	High (built for precision and stability)

).

3.2. Data Acquisition and Transmission

The overall effectiveness and utility of any vibration-monitoring system are profoundly dependent on its robust data acquisition and efficient data transmission capabilities. These two elements ensure that the vast amounts of raw data generated by the sensors are reliably captured, processed, and delivered to where they can be meaningfully analyzed [21,28].

At the core of this process are data loggers and data acquisition systems (DAS). The reliability and practical value of any vibration-monitoring system is strongly based upon the strength of its data acquisition and data transfer systems. Here, these two components collect the huge flood of raw data from sensors and transmit it to a place where engineers can generate meaningful analysis.

At the heart of this are data loggers and DAS (data acquisition systems). These specialized blocks are intended for capturing analog electric signals from the sensors. Their primary role is to digitize these analog signals with high fidelity so that they can be further processed and stored as digital data, such as at, for example, a local or remote data center. Current DAS include advanced attributes like anti-aliasing filtering (AAF)—essential to prevent signal degradation during digitization—and they sample at high rates too, so as to preserve the fidelity and integrity in the data acquisition, adequately capturing even high-frequency events.

One of the most important paradigm changes in railway communications is represented by the introduction and diffusion of wireless communication (WC) technology in the railway domain, especially of wireless sensor networks (WSN) and their combination with the Internet of Things (IoT) principles [12]. These technologies have even changed the way data are transmitted from the vehicle to ground- or cloud-based systems. One of the major benefits wireless systems offer is the reduction in wiring infrastructure and complexity, resulting in lower installation costs and less disruption through the retro-fitment of monitors on to legacy rolling stock. Moreover, wireless functionality to stream data in real-time or near real-time is also available, which can offer operators immediate visibility into vehicle health and performance. Wireless connections may include Wi-Fi for local, high-bandwidth connections, cellular (e.g., 4G and increasingly 5G) for wide-area connections and high data throughput, and LoRaWAN for low-power, long-distance communication, namely with less data-intensive use cases. This wireless framework is a necessity for providing the capability for continuous, remote monitoring and for providing effective early warning of possible problems as they develop.

In order to circumvent the challenges of managing the huge amount of raw vibration data and to enable fast anomaly detection, in recent years, an increasing migration towards integrating the possibility of onboard data processing (edge computing) into the monitoring systems has been observed [12]. Rather than sending all raw data back to a central server or cloud provider, which would likely be expensive and bandwidth-intensive, edge computing involves the ability to process some of the data locally, with key feature extraction (e.g., RMS, simple FFT), and perhaps some very rudimentary machine learning inference (e.g., basic classification or anomaly detection) being performed on the vehicle. This “edge processing” facilitates communication in that it reduces the amount of data to be sent, thus saving bandwidth and lowering the communication costs. Crucially, it also enables faster turnaround of urgent alerts, saving middlemen and better identifying potential upsets before sensitive or mission-critical data are sent to the cloud to be analyzed and saved for a future activity. Specialized units are designed to collect analog electrical signals emanating from various sensors. Their primary function is to accurately digitize these analog signals, converting them into a format suitable for computational processing, and then to store this digital data, either locally or for subsequent transmission. Modern DAS often incorporate sophisticated features such as anti-aliasing filters, which are crucial for preventing signal distortion during digitization, and operate at high sampling rates to ensure that the fidelity and integrity of the acquired data are maintained, capturing even high-frequency events accurately. This approach presents a critical trade-off with traditional cloud-based analytics:

• Latency: Edge computing low-latency for critical fault detection, enabling real time alerts and actions that are essential requirements for safety during critical events. Cloud analytics, on the other hand, naturally introduces latency delays coming from the data transmission side which can be a no-go especially for time-sensitive applications.
• Power consumption: Edge devices are often battery-powered (e.g., mobile robots, portable gadgets) and should be designed for low-power operation, with minimal to no processing power. This burden is put on powerful data centers—a costly and environmentally unsound solution. That comes down to whether the energy cost of local processing is lower than that of constantly sending data back and forth.
• Fault detection accuracy: As a cloud-based analytics, you have a larger historical dataset as well as fleet-wide comprehensive big data information to train much more sophisticated, higher-accuracy models. Because models running on the edge necessarily have to be simpler and make do with less data, they may have lower accuracy or fail to spot some more subtle failure modes. A hybrid model can help by having edge devices watch for anomalies and send them to the cloud, where a more detailed, high-quality AI routine will analyze the behavior.
• Cybersecurity: Moving back to one of the points raised earlier, because they store it locally, businesses can limit how and where some sensitive data are accessed, which minimizes the size of their attack surface. However, locking down devices to a vehicle is challenging. Despite robust central defense and professional cybersecurity measures, vehicles remain high-value targets for sophisticated attackers, as they can serve as a single point of failure (SPOF). Edge computing is more decentralized in nature and thus may be even more resistant against a Serina vulnerability.

3.3. Wayside Measurement Systems

The onboard systems afford instantaneous and direct retrievals of the dynamic behavior of railway vehicles, but the wayside monitoring systems report important additional information [29]. These systems are strategically placed roadside and provide excellent information about wheel–rail interaction and track infrastructure condition. Because they are stationary, they can watch any train that goes by, regardless of the onboard gear installed in each car.

Key examples of wayside systems are WILD (wheel impact load detection). Such systems are intended to measure the dynamic forces that wheels of a train apply to the rail. They are usually based on the use of high-precision strain gauges or more sophisticated optical fiber sensors installed in the rail head area at some points. No description of the prior art concerning the detection of heavy defects on the wheel appear to have been made in the methodology described of the WILD truck, and, by itself, the WILD truck is an essential tool in recognizing serious problems of wheel defects such as wheel flats (local flat spots on the wheel tread), wheels out-of-round, or other forms of wheel damage which cause dangerously high “hit” loads on the track. By detecting these defects early, WILD systems are highly valuable to avoid potential track damage, minimize infrastructure wear, and—above all—to support the prevention of excessive dynamic forces that may lead to derailment.

A further key class is bearing acoustic monitors (BAM). These systems rely on dedicated acoustic sensors installed on/near the track to detect anomalous noises generated by wheel bearings during pass-by of the trains. The sound details caused by bearings in their early failure stages (such as grinding, squealing or rhythmic knocking sound) can be easily distinguished before significant vibration symptoms are developed. BAM technology has proven to be capable of accurately detecting these early indicators from noise and predicting potential bearing failures, so that maintenance can be performed in time before catastrophic accidents/weakening in service.

Not being directly a vibration-measuring-type system, track geometry measurement systems are close relatives, in that the track geometry is one of the originators of vehicle vibrations [22,23,30]. These systems—often mounted on dedicated track geometry cars or integrated with some onboard inspection vehicles (along with accelerometers and gyroscopes)—accurately measure key track parameters, including geometry alignment (straightness), twist (torsion), and super elevation (cross-level). Thus, data acquired on track geometry can be directly related to the vehicle vibration data to identify those routes which give excessive vehicle response thus directing preventive track maintenance activities.

Rail corrugation measurement systems ultimately serve to monitor and quantify periodic wear that occurs on the rail head in the form of rail corrugation. This effect is a major cause for high-frequency vibrations and high noise, thus influencing ride comfort and component wear [22]. Such systems are able to determine the depth and the wavelength of the corrugation, the information that is important for the planning of grinding measures for the reconstruction of the rail profile to eliminate these undesired vibrations (

Table 2. Comparative summary of onboard vs. wayside systems.

Feature	Onboard Monitoring Systems	Wayside Monitoring Systems
Cost	Installation: high for initial sensor deployment and integration per vehicle. Maintenance: low, as systems are part of vehicle maintenance cycle.	Installation: high for trackside infrastructure and civil works. Maintenance: moderate to high, due to environmental exposure and access challenges.
Accuracy	Vehicle-centric: high accuracy for vehicle-specific health (suspension, wheel wear). Provides real-time dynamic behavior data.	Track-centric: high accuracy for wheel–rail interaction and track geometry. Captures data on all passing trains.
Limitations	Infrastructure data: limited or no data on track condition, environmental noise, or wayside structures. Cannot monitor other trains.	Vehicle data: limited data on vehicle-specific components (e.g., bogie hunting, passenger comfort). Cannot detect some internal vehicle faults.
Best-use scenarios	Continuous monitoring for predictive maintenance of rolling stock, passenger comfort evaluation, and real-time vehicle performance analysis.	Wheel–rail interaction analysis, high-speed impact detection (WILD), and track condition monitoring for all fleet traffic.

).

3.4. Comparative Analysis of Measurement Methods

An integrated application of railway vibration measurement includes using a combination of various measurement techniques, each complements the others. Knowledge of these differences is important when designing a monitoring program that is capable of providing a balance between data richness, cost-effectiveness and operational reality. Below is a comparison of the main measurement techniques reviewed: onboard systems, wayside systems, and simulation-based methods (

Table 3. Overview of vibration measurement methods in railway vehicles.

No.	Measurement Method	Sensor Type	Processing Techniques	Primary Application	References
1	Onboard accelerometers	Accelerometers (1D/3D)	FFT, RMS, band-pass filtering	Dynamic monitoring, fault detection	[31,32,33]
2	Accelerometers on bogies and chassis	Accelerometers	RMS, frequency analysis	High-speed vibration analysis	[34,35,36]
3	Distributed sensors + strain gauges	Accelerometers, strain gauges	Modal analysis, SHM algorithms	Structural health monitoring	[37,38,39]
4	Accelerometers on passenger seats	Accelerometers	ISO 2631 filters, spectral integration	Ride comfort evaluation	[39,40]
5	Simulation-based measurement	Virtual sensors (MBD/FEM)	Multi-body dynamics, hybrid modeling	Model validation, design optimization	[2,41,42]
6	Wireless sensor networks (WSNs)	MEMS accelerometers	Edge computing, data aggregation	Real-time monitoring, scalable deployment	[43,44,45]
7	Sensor fusion systems	Accelerometers, gyros, GPS	AI, clustering, anomaly detection	Predictive maintenance, fault classification	[15,46,47]

).

3.5. Requirements for Monitoring Instruments

Various stringent demands have to be met in order to ensure that vibration monitoring instruments can supply reliable data under the extreme conditions typically encountered in the tough world of the railway environment [21,27]. These requirements are designed to guarantee their life, precision and safety in harsh working conditions.

• Robustness and durability: Railway environments are extremely hard on the instruments, which are subjected not only to fierce mechanical shock (impacts, coupling events, etc.), but also kept under continuous vibration, rapid temperature cycling (from sub-zero winters to scorching summers), high humidity, and a tremendous amount of electromagnetic interference from various electrical systems. The monitoring devices, therefore, need to be carefully engineered and built so that they can be placed in these conditions without loss of function. They need to last long term, eat up the miles, not the maintenance, for the lowest total cost, and highest availability.
• We all know that accuracy and precision are important for data that actually means something. Very accurate measuring equipment over the whole relevant frequency is required. For that, sensors should have low noise levels to not cover underlying true vibration signals, a large dynamic range to observe small and large events and as little cross-axis sensitivity as possible (i.e., a sensor should predominantly measure vibrations in the intended direction and dismiss the other ones). These properties are important for dependable fault detection and comfort evaluation.
• The sampling rate and bandwidth of the instrumentation should be carefully chosen to include all interesting vibrational features. The sampling frequency, the rate at which a signal is sampled, must be high enough to properly record the highest frequencies of interest in the railway domain. For example, the evaluation of ride comfort usually is a function of broadband frequencies up to ~ 1 kHz, but consideration of structural resonances or wheel–rail contact phenomena (impacts, squeal) may require yet greater bandwidths to minimize the effects of aliasing and to ensure complete data capture.
• Calibration and stability are required to ensure data reliability over time. Regular calibration is indispensable if the instruments are to deliver reliable measurements over the operating life. Additionally, sensors should, in general, also be stable over time, displaying little or no drift of their output values over time even when subjected to three different tests. This means recalibration is required less frequently and trend data quality is consistent.
• Power efficiency plays an important role, especially when dealing with onboard and wireless monitoring systems. For wireless sensor nodes powered by battery, power saving is critical to the life of the power source and refill cycle and thus the economy of operation and maintenance. It also allows sensors to be installed at remote or hard to reach places with no external power sources.
• It would need sufficient experience of data storage and communication for effective data processing. In addition, instruments should have ample onboard storage to buffer data, particularly during periods of connectivity loss and for offline analysis. Equally vital are the communication protocols able to end-to-end transmit acquired information to CPU or cloud platforms and to guarantee that data is always available for analysis and decision-making in terms of, e.g., Wi-Fi, OCC or LoRaWan.

And last but not least, it is a must to comply with regulations. Instruments must be compliant with applicable industrial standards in order to provide operational safety, electromagnetic compatibility and mechanical reliability. Prominent standards include EN 50121-1:2017 [48], Electromagnetic Compatibility (EMC) for railway applications, which covers interference; and EN 61373:2011 [49], Shock and Vibration tests, which verifies an instrument’s ability to physically survive the demanding dynamic environment of the railway vehicle. Compliance with these standards ensures the monitoring systems are trustworthy and do not compromise the security of railway operation.

4. Data Processing and Analysis Techniques

The original data collected from the raster of vibration-monitoring systems is of large volume, complex and sometimes noisy. Unprocessed and unanalyzed, this amount of data is not intelligible. Effective processing and advanced analysis are not just nice to have but are crucial for extracting valuable information, correctly identifying anomalies, diagnosing faults, and to enable sound decision-making processes in the context of railway operations and maintenance [3,27].

4.1. Time Domain Analysis

Time domain analysis serves as the essential tool for vibration signal investigation. It serves to look at the amplitude of a signal as a function of time directly to give visual and intuitive understanding of its most basic properties. This procedure is frequently the essential first phase in any vibration-data treatment process.

The simplest technique when directly analyzing in the time domain is raw time-series visualization. With just the raw acceleration, velocity or displacement signal intrinsic to a critical event, the scenario can be seen visually when signals are plotted vs. time. Such a visual inspection can disclose impulsive, high-amplitude transient events (such as impacts from wheel flats) as well as sharp, short duration shocks resulting from track irregularities, or large amplitude excursions associated with system instabilities. This kind of immediate feedback provides a very rich resource for a first, qualitative assessment of the data, as well as the determination of particular interesting points for closer scrutiny.

Attached to visualization are the quantitative measures. RMS values are one of the most commonly used metrics to measure the integrated energy content (or overall intensity) of a vibration signal throughout a given time-window period. The RMS value is a measure of the average power of the vibration. For important applications such as ride comfort evaluation, especially for the standards SR-EN 12299 and ISO 2631-1, specific frequency-weighted RMS values of acceleration are used because they are important parameters that correlate well with the human vibration perception. By embedding a time-history level of analysis, the time-dependent evolution of RMS values can give definite evidence to progressive deterioration on vehicle components and infrastructures, or changes in an operational condition altering the vibration levels.

Other key points or metrics are peak values which are simply the maximum and minimum amplitudes of the signal. These values represent the maximum positive or negative amplitude of a vibration signal. Maximum values can be critical for the detection of impulsive or extreme loads that are short in duration and might cause structural failure or material fatigue when they exceed design thresholds.

The crest factor (CF), which is the ratio between the peak and the RMS value, is a reliable indicator of the impulsiveness or “spikiness” of the vibration. A high crest factor indicates separate and distinct impacts, sharp shocks, or very concentrated failures (wheel flats hitting the rail). This measure may be capable of revealing issues not easily identified when using RMS values that smudge the signal in time.

Finally, a selection of statistical parameters, covering broader ranges, enables us to accrue further understanding of the distribution and the nature of signal vibration. Variance and standard deviation are parameters that measure the data distribution about the mean. Skewness quantifies the asymmetry in the distribution of the amplitude of the signal, that is, whether the signal has a biased distribution of positive or negative amplitudes. Kurtosis, as a fourth-order statistical moment is of great interest because it is very sensitive to the existence of impulsive components or impacts in the vibration signal. A high order kurtosis index is an eight indicator for the incipiency of faults, such as bearings and gears, where local defects cause repetitive impacts. Analysis of these statistical parameters can enable engineers to understand further the fundamental sources of vibrations.

4.2. Frequency Domain Analysis

Even though time domain analysis can give a rapid indication of the nature of the signal, the frequency-domain analysis transforms the time domain signal into its frequency component parts that show the fundamental and harmonically related resonant frequencies. This transformation is extremely necessary to detect the cyclic faults generating vibrations with certain frequencies and to obtain comprehensive insight into the dynamic characteristics and natural frequencies of railway system components.

The fast Fourier transform (FFT) is the most common method to perform this transformation. The FFT algorithm is a computationally fast manner of transforming a time-domain signal into the frequency domain where the amplitude (or power) of each frequency component versus frequency is plotted. By using FFT, engineers can pinpoint the primary frequencies that appear in the vibration signal and measure their associated amplitudes. In railway systems, FFT has been widely employed for the identification of frequencies associated with more predictable phenomena such as wheel revolution, bogie hunting onset and characteristics, rail corrugation presence and wavelength, and detection of structural resonance in the vehicle and the railway track. Monitoring the variations of these frequency components over time may indicate faults in process.

One more important quantity in frequency domain studies is the power spectral density (PSD). FFT gives you amplitude at some specific frequencies but PSD gives you how the power (or energy) of the signal is distributed throughout the range of frequencies. PSD is very useful for analyzing random or wide-band vibrations; for example, caused by the rail surface roughness itself. Plots of the PSD can be used to visually emphasize ranges of frequency which contain the most important vibrational energy. For example, it might be an indication that the track is in a worse condition, or that some device is resonating at or near the critical speed.

Harmonic analysis broadens the applicability of FFT to details based on integer multiples of the fundamental frequency itself. The majority of critical railway disturbances appear as a series of individual harmonics of wheel rotational speed or other “known”, characteristic system frequency. For instance, a growing wheel flat will produce heavy blows at a frequency equal to the speed of the wheel rotation and at the harmonics thereof. Also, bearing defects may generate vibration harmonics at particular frequencies which are a function of bearing geometry and shaft rotation speed. Tracking these harmonics over time, or by comparing them with vehicle speed (as typically presented in waterfall or cascade plots) can be a very effective way of detecting certain types of faults, such as wheel flats, or bearing degradation at an early stage. This frequency domain focused analysis is important for accurate fault diagnosis.

4.3. Time–Frequency Analysis

The traditional time-domain and frequency-domain methods are strong but have some limitations in the sense that they focus on stationary signals, in which the power spectral density of the signal does not change much over time. The non-stationarity behavior frequently found on railway vibrations is caused by dynamic parameters such as train velocity variations, track condition changes or transient events such as impacts and leap braking. For such situations, time–frequency methods are indispensable, since they supply information about how (the) signal (energy) is distributed over time and frequency at the same time.

One of the basic time–frequency methods is the short-time Fourier transform (STFT). This essentially amounts to performing a complete FFT on a series of small, overlapping segments (or “windows”) of the signal as it slides through time. By repeating this process as the window moves along the time axis, a sequence of spectra in time is produced, resulting in a time-varying frequency representation. Although conceptually straightforward and universally employed, the fundamental limitation of STFT is determined by the Heisenberg uncertainty principle, which dictates a trade-off between time resolution and frequency resolution. While a short window gives good time localization and poor frequency resolution, a long window gives good frequency resolution but poor time localization. This compromise implies that STFT could perform sub-optimally for signals comprised of both sharp impulses and slow time-varying frequency components.

In order to address the drawbacks related to a fixed-resolution, more elaborate and flexible tools for frequency analysis are used, i.e., wavelet transform (WT). Wavelet transforms use “wavelet” oscillations that oscillate in both time and frequency space that are scaled and shifted across the signal for analysis. This multiresolution analysis property is a crucial advantage: wavelets can adequately supply a good timing resolution for high-frequency transient events (by means of narrow stretched wavelets), and at the same time supply a good frequency resolution for low-frequency sustained phenomena (by using broad compressed wavelets). Wavelets: both CWT and DWT are widely used in railway vibration signals for purposes such as reliable signal denoising, accurate feature extraction for machine learning and fault diagnosis. For instance, certain wavelet coefficients may be particularly effective at emphasizing the impulsive nature of wheel defects, rendering them detectable in a noisy background.

The Hilbert–Huang transform (HHT) is another high-level technique that is especially appropriate in non-linear and non-stationary signals such as those found in railway dynamics. In contrast to the above two types of transforms that use fixed basis functions such as Fourier or wavelet basis functions, HHT is an adaptive method. The principal operating mechanism involves two steps: firstly, signal decomposition into a finite, typically small, number of intrinsic mode functions (IMF) via an iterative procedure known as empirical mode decomposition (EMD). Each IMF is regarded as a simple oscillatory mode of the signal. Secondly, we perform independent Hilbert transforms on IMFs to obtain instantaneous frequency and instantaneous amplitude. HHT has shown great value in the identification of subtle variations and interaction between the different modes in rail vibration signals which could be overlooked by conventional FFT analysis, presenting a more sophisticated perception of actual mechanical processes.

4.4. Advanced Analysis and Machine Learning

AI- and ML-based approaches are steadily becoming core to support decisions and automate practical decisions in a time when vibration data have been increasing rapidly in volume and in-built complexity, especially with continuous monitoring of railway systems [11].

The necessary first steps in applying any machine learning algorithms are feature extraction and selection, respectively. The raw vibration signals are abundant in content but unstructured. Accordingly, the meaningful numerical features should be carefully extracted from these raw time-series or converted signals. They can cover a broad range of parameters: these could be statistical parameters (like root mean square (RMS) values, kurtosis, skewness), specific features in the frequency domain (such as dominant frequencies, bandwidth of energy concentration), time frequency features (for example, coefficients from a wavelet transform), or even model-based parameters from approaches like modal analysis. The following step of feature selection is to detect the most discriminating and representative features, which can bring high ability to accurately diagnose the fault, and remove the redundant and noise information so as to enhance the model’s ability efficiently.

After features you care about have been extracted, then you start using classification algorithms. Supervised machine learning algorithms are generally used to classify various vibration patterns into predefined classes or labels. These classes often reflect different modes of operation or failure types (e.g., to classify a pattern of vibration as “healthy” vs. “wheel flat” vs. “bearing fault” vs. “track irregularity”). Common classification algorithms such as support vector machines (SVMs), renowned for high dimensional stability, k-nearest neighbors (k-NN), which relies on surrounding features for classification, decision trees and random forests, renowned for being humanly interpretable, and artificial neural networks (ANNs), capable of handling complex non-linear classification, have been some of the more successful algorithms for this setting.

In addition to classification, anomaly detection methods are equally important, especially in cases of unexpected or new types of failures. These generally make use of unsupervised learning techniques in order to determine the anomalies or variations in the vibration that are different enough from typical safe operation of the normal operational conditions of the machine and most importantly without having to have any validated labelled examples, which goes hand in hand with all the fault types. The most common algorithms used for anomaly detection are one-class SVMs, isolation forests, which create a model that isolates the anomalies, or autoencoders, a specific network that is trained to reconstruct normal data and recognize an anomaly as data that it was not able to reconstruct well. By training on “healthy” vibration data, an autoencoder learns to reconstruct the normal signal. When a new signal containing an anomaly or a significant deviation from the norm is fed into the model, the reconstruction error will be high, acting as a clear indicator of an anomaly. Similarly, other unsupervised techniques like clustering algorithms can group similar vibration patterns, making it easier to identify outliers that represent potential faults, even if the fault is previously unknown. Anomaly detection is especially important for identifying new or otherwise non-correlated failure modes, serving as an early warning to categories of known faults.

As part of such proactive maintenance, predictive maintenance (prognostics) is gaining importance. Machine-learning techniques can train effectively on historical data that includes vibration trends, multiple other types of measured parameters, and corresponding health states, and can accurately forecast the remaining useful life (RUL) of a piece of equipment, or predict future degradation trends. This approach usually uses regression algorithms to estimate continuous states (e.g., time to failure), or more sophisticated sequence models such as RNN, long short-term memory (LSTM) networks, which are proficient at learning complex structures from time-series vibration data and associating them with other operating parameters in order to obtain an acceptable future prediction.

With the advent of deep learning, vibration analysis has been further impacted. Deep learning models, of which one type is LSTM, a special kind of neural network with multiple layers of processing units or nodes, duo trace analysis and end-to-end fault diagnosis and prediction have been widely used. For example, convolutional neural networks (CNN), a model commonly used in image processing, can be used to form image-like representations of vibration data (such as spectrograms or scalograms), naturally learning very discriminative features from the raw data. The RNNs, being naturally suitable for sequential engineering data, fit the raw time-series vibration signals. Deep learning has a strong advantage in learning complicated and hierarchical features from raw data automatically, which makes it possible to significantly alleviate or even remove the labor-intensive feature engineering process, thereby obtaining more robust and generalized models.

Deep learning has a strong advantage in learning complicated and hierarchical features from raw data automatically, which makes it possible to significantly alleviate or even remove the labor-intensive feature-engineering process, thereby obtaining more robust and generalized models. For example, a known case study involves using a 1D CNN to directly process raw accelerometer data to classify different levels of wheel wear with high accuracy, without the need for manual feature extraction. Similarly, LSTM networks were successfully applied to predict the RUL of bogie bearings by learning degradation patterns from long-term vibration time series.

One idea that is gaining traction at the moment is digital twins. They are digital, high-fidelity copies of physical assets, enabling the coupling of vibration data during operation with complex simulation models. For railway vibration monitoring, the full digital twin provides a permanent and real-time synchronization between the measured data of physical sensors and the different behavior that is simulated for its virtual proxy. This makes it possible to develop capabilities such as predictive analysis over current and historical data, and test “what if” scenarios “in silico” (e.g., how a new type of track defect would impact vehicle dynamics) for robust feedback loops for adaptive control and optimized maintenance planning. For the railway DT, this would include the virtual representation of an entire rolling stock or a segment of a track of infrastructure, degrading its componentry, predicting its faults, optimizing scheduled maintenance, assessing its performance and improving future designs through virtual testing.

There needs to be a thorough model calibration and validation process to verify that these FEM/MBD models are accurate compared to real-world data. This often takes the form of simulating outputs (such as an acceleration profile over a particular track irregularity) then comparing them to measured data from an instrumented test vehicle. Even though no single agreed-upon benchmark exists, it is standard to use various documented test runs (like a known-segment constant speed run) as a validation dataset on which the results of different control algorithms are compared. They then use the differences between reality and the model to adjust the parameters of the model, e.g., spring rates or damping coefficients, so that their simulation can reproduce what happens in the real world. In order for these models to be useful in predictive maintenance and design optimization, the accuracy needs to be as perfect as possible.

4.5. Software and Platforms

The complex methodologies of acquiring, handling and analyzing large databases of railway vibration data are supported by a variety of heavy software tools and powerful analytical environments. These vary from general purpose mathematical modelling tools to industry-specific highly customized analytics tools that come with their unique features for engineers and scientists [3,28].

One of the most popular general purpose tool is MATLAB/Simulink (2024). With a variety of toolboxes available, including signal processing, machine learning, and control, MATLAB is a strong platform for research and development. It is often used for the development and testing of complex algorithms, as well as detailed signal processing and modeling, and can perform similarly complex MBD simulations, e.g., for a railway vehicle. Simulink is a block-diagram environment operating in the multi-domain and model-based design of the MATLAB, and is perfect for simulating and testing the dynamics for the vehicle–track interaction.

Python (2025) has quickly become a dominant open-source programming language for data analysis, machine learning, and scientific computing. Its powerful libraries and thriving ecosystem make it is very flexible. NumPy provides a solid array computing foundation and then targets for optimization. SciPy extends this functionality with mathematical, scientific, and engineering functions, and Pandas allows efficient and flexible data manipulation as well as data analysis. With machine learning, Scikit-learn provides a large array of methods; and with deep learning, TensorFlow (2025) and PyTorch (2025) are two of the major open-source frameworks. Its flexibility, large community, and many, many libraries have caused it to become the language of preference for custom analytics, algorithm prototyping and integration with more complex data pipelines.

LabVIEW (2025) is widely used among researchers who are developing real-time data acquisition and control systems. Its graphical programming interface also makes it highly suitable for direct interaction with hardware, and it is also used extensively in experimental or laboratory test rigs for prototype or custom high-speed data and control logging systems for railway testing. So, its power to design data visualization or system control intuitive user interfaces is strong in this case.

Apart from these generic solutions, an increasing variety of specialized railway software and the commercially available platforms provide customized solutions exclusively for railway asset management, condition-based monitoring, and predictive maintenance. Such end-to-end software and systems also receive and process data from various unrelated sources (including but not limited to vibration data, track geometry data, wheel profile scanner data, operational parameters), consolidating the data in one platform (AKA world), for an accurate cross-sectional view. They commonly supply domain-specific data analysis, premade models for prevalent railway defects and adapted reporting, reducing the complexity of the knowledge extracted from the raw data to aid the rail operators and infrastructure managers in maintenance and operation-level decision-making. These solutions focus on giving a more comprehensive perspective on asset health beyond what dedicated monitoring systems can provide.

5. Practical Applications and Case Studies

In practice, the testing and analysis of rail vehicles with vibration measurement and monitoring technology offer a range of advantages.

These advances significantly increase system availability, passenger comfort and overall system energy efficiency for the railway industry as a whole [1,4], fundamentally overhauling maintenance from a reactive to predictive discipline.

5.1. Passenger Comfort Evaluation

Likely, the single most widespread and fundamental mandate for practical vibration monitoring is the passenger comfort specification.

The presence of vibration in rail vehicles is one of the most critical aspects as, in short trips, high vibration magnitude is generally associated with passenger discomfort and problems with fatigue and can result in long-term health symptoms in people because of the effects of strong or continuous exposure.

To relieve these hardships, powerful onboard vibration-monitoring systems are frequently employed to analyze the acceleration in passenger cabins, more specifically at floor level and even in the passenger seats.

Readings are internationally adjusted according to the norms SR-EN 12299 and ISO 2631-1, always considering all necessary precautions.

5.2. Predictive Maintenance for Rolling Stock

There are mainly two drivers for the development and utilization of advanced vibration monitoring (aside from protection): proactive maintenance and predictive maintenance of rolling stock.

Vibration monitoring is used to detect and diagnose a wide range of errors in critical components, e.g., wheelset health, bogie-suspension system or even the pantograph (early warning). By carefully monitoring the subtle shifts in vibration signatures—for example, certain harmonic frequencies might begin popping up (perhaps at multiples of the wheel rotation speed)—potential failures can be flagged up long before they actually occur.

Such methods substantially lower maintenance costs, reduce the number of operational downtimes and enhance railway operations safety.

5.3. Infrastructure Condition Assessment

There are also new post-processing analysis methods that can be used for the detailed assessment of track infrastructure conditions, based on GPS information combined with onboard vibration data.

The vibration response of a rail vehicle travelling on track can be used to diagnose the quality and condition of the rail profile, track geometry, and integrity of the rail joints and sleepers. For instance, a sudden change in the amplitude of vibrations may represent a failure (e.g., broken rail joint) which can be further inspected and repaired earlier than later by performing proactive maintenance before it becomes a bigger issue.

This way, vibration monitoring might add to the traditional track inspection methods that infrastructure managers use to plan better maintenance activities [50,51].

5.4. Applications in Freight Trains

Even in industries where the comfort of passengers is not a top priority, such as freight trains, critical both for safety during operations and maintenance as well as preservation of cargo integrity is the measuring of vibrations. When it comes to goods (freight), the detection of wheel defects or nascent bogie problems, by means of vibration analysis, as early as possible is a must. Early warnings of such derails are essential for safeguarding against catastrophic derailments and damage to sensitive/hazardous materials en route, thus ensuring the safe movement of essential materials. These various use cases emphasize the widespread and necessity of vibration monitoring in the railway industry, and how it improves every aspect of railway operation nowadays.

6. Conclusions and Future Work

6.1. Conclusions

The field of vibration measurement and monitoring in railway vehicles has experienced a substantial change from intermittent diagnostic practice to an integrated and core subject in the whole lifecycle of railway systems, as well as a full stage of development. This phenomenal transformation has been powered by immense technological developments in several aspects, such as advanced sensor technologies, ubiquitous development of wireless networks, emergence of efficient edge computing, and, more strikingly, by the path-breaking advancements in artificial intelligence (AI) and machine learning (ML). Together, these new breakthrough technologies have significantly changed the perception, assessment and control of vibrations in the railway sector.

The following are conclusions derived from the extensive study:

• Multidisciplinary importance: It is quite evident that vibration monitoring is not only useful, but absolutely necessary in several aspects of the railway system. Its importance goes well beyond making passengers feel comfortable and includes issues as vital as increasing operative safety, making predictive maintenance highly efficient for rolling stock and infrastructure and playing a major role in alleviating the environmental impact resulting from all railway activities.
• Technology convergence: The modern railway vibration-monitoring landscape is characterized by the strength of convergence and compatibility of new technology. This covers new types of sensors (like MEMS and piezoelectric devices with high sensitivity) and robust IoT platforms, high-speed communication networks (5G) and power supply technologies data analysis platforms that include increasingly important digital twin concepts. This combination is what has allowed the emergence of truly intelligent monitoring systems that analyze in real-time and effectively predict probable faults before they grow too big.
• Role of standards and research: The existing regulatory framework, particularly through standards like SR-EN 12299, ISO 2631 and ISO 14837, and the important European research projects: Shift2Rail, SILVARSTAR and RIVAS have had an absolutely decisive role. Such frameworks and initiatives have facilitated standardization of measurement practices, contributed to the emergence of new diagnostic tools and played an important role in the promotion of the pervasive application of advanced diagnostic technologies throughout the European railway scene.
• Advanced vibration data analysis: Being able to work the complex data characteristics of most vibration systems is essential. This requires the employment of a broad spectrum of signal-processing approaches, which range from classical techniques that are well-known in the field (e.g., root mean square (RMS) computation, fast Fourier transform (FFT)) to very advanced, AI-based algorithms. These advanced methods which are called time-frequency analysis techniques, such as wavelet transforms and Hilbert–Huang transforms (HHT), along with knowledge-driven machine learning algorithms for classification, anomaly detection and prognostics, are essential for capturing actionable information from large amounts of acquired data.
• Persistent challenges Although there has been tremendous progress and technological advances made, there still exists a gamut of challenging problems associated with the deployment and efficient functioning of these advanced monitoring systems. These are challenges which involve optimized sensor placement, data management on extremely large-to-grand scales, the ambiguities of signal interpretation, system adaptation to differences in operation and environment, and robust collaborative system integration within typically heterogeneous railway systems. These regions are active and fertile for further research and development.

6.2. Future Work

Railway vibration monitoring is dynamic in nature and sustainable innovations are still possible, as well as transformational developments. It is expected that the future direction of research and development in this area will fall into several directions with the goal of improving the intelligence, autonomy, and integration of monitoring solutions.

A prominent example will be the development of smart and autonomous sensors. Next steps would include the development and deployment of sensors which consume even less power, thus increasing their duration of operation and relaxation of maintenance requirements. These next-gen sensors will also more than likely have significantly improved onboard processing compared to their once “cloud-only” predecessors, with edge AI that can process data instantly on-the-device. Moreover, self-calibration algorithms will certainly also need to be included to compensate for calibration drift and to increase long-term measurement precision, in order to diminish the reliance on manual calibration processes.

The expansion of digital twin adoption is another key trend. Efforts will be made to apply the concept of digital twins to rail vehicles and overall infrastructure on a large scale. This will make it possible to achieve truly continuous monitoring, to forecast asset behavior with very high precision, even under extreme or unexpected operational conditions, and will decrease the operational process and lifecycle of railway asset optimization. This will require tighter integration of advanced physical models and streaming data from in-the-field sensors in nearly real time.

Explainable artificial intelligence (XAI) will be a topic of research attention with AI models. In safety critical railway applications, the need to build explainable AI models that can offer transparent and interpretable justifications for their decisions, especially to raise trust and acceptability in decisions made by AI algorithms, is more critical. This will enable human operators to understand the “why” a certain fault was detected and correct mistakes when needed in a fail-safe scenario.

Moreover, future work will be focused on multi-sensory data integration. Combining vibration with other sensor sources—like imaging data from cameras, audio signatures from microphones, temperature information, accurate GPS location, and even real-time meteorological conditions—will represent an incomparably holistic view of vehicle health and track conditions. Such multi-modal data fusion can effectively reduce the ambiguity in fault diagnosis and give more robust information.

Efforts for standardization and interoperability at a European level should be continued. There will also be continued challenges to harmonize measurement processes, data formats and system interfaces among various railway operators and system suppliers. This should create convergence that will assist in the transparent, smooth exchange of data management and in the harmonization of incompatible observation systems, while accelerating the widespread implementation of innovative technologies throughout the whole European railway network.

The construction of robust predictive models in variable environments will also be an important core target. Future machine learning models will have to be specifically adapted to become more resilient and accurate to the typical variations in operating conditions (e.g., train speed, load, passenger density) and environmental conditions (e.g., temperature extremes, precipitation). This increased stability will result in a decrease in false positives and will provide much improved accuracy and reliability of fault prediction.

As monitoring systems become more connected, the most critical issue will be cybersecurity and data privacy. Future studies will put a great deal of effort into enhancing the cybersecurity of these connected monitoring systems in order to safeguard sensitive information. Such proactive security measures are crucial to the resiliency of critical infrastructure, and guard against cyber threats that could threaten safety and operational uptime.

While the focus of this paper is firmly on railway applications, it is important to acknowledge that many of the challenges and methodologies discussed are not unique to this sector. The core principles of vibration monitoring, signal processing, and predictive analytics are shared across various engineering disciplines. For instance, the aviation industry utilizes vibration analysis extensively for the mechanical health monitoring of jet engines and airframe structures, a practice that shares common methodologies with railway condition monitoring. Similarly, the broader application of vibration and frequency analysis for mechanical health monitoring in other specialized fields, could be offering a more generalized context for the techniques presented here. This cross-sector relevance underscores the universal importance of vibration-based diagnostics and suggests potential for future knowledge transfer and methodological innovation from these other domains.

To conclude, vibration-monitoring applications in railway vehicles will follow a trend towards a world of deep blending, greater intelligence and omnipresent proactivity. The overriding ideal is the never-ending pursuit of developing a form of railway transportation that is visibly safer, vastly more comfortable, conspicuously more efficient, and in the end, one that is sustainable for future generations.