Abstract
The efficiency and availability of modern railway infrastructure plays an increasingly strategic role in the sustainability, development and prosperity of communities and nations. Recent Artificial Intelligence (AI) algorithms, which enable the use of digital tools such as Data-Driven models that can automatically adapt system operation, make decisions and suggest strategies based on collected data, form the basis of modern Predictive Maintenance (PdM). PdM is considered a key opportunity for accurate Structural Health Monitoring (SHM), especially for railway infrastructure, where the transition from traditional preventive or periodic maintenance to PdM will reduce intervention times and costs. Furthermore, by directly correlating actual infrastructure conditions with measured information, SHM can utilise a limited number of sensors installed on critical components such as insulated rail joints. This review starts by clearly describing the different components that make up the railway infrastructure, the monitoring systems currently in use and the technical performance parameters that indicate their health status and goes on to examine the issues related to the SHM and related modern digital tools. All these topics are summarised to provide an effective theoretical and practical knowledge of SHM for railway infrastructure, to better understand the current transformation of the sector and to predict future developments.
Keywords
Introduction
The railway network and its infrastructure play a key role in the social, economic and sustainable development of nations, which is why many countries are steadily increasing investment in rail systems, which are capable of moving large volumes of goods and passengers with low energy consumption compared to other land transport modes.1–6 While this movement is essential to stimulate economic growth, the intensification of traffic is one of the main causes of the deterioration of infrastructure, which is increasingly exposed to stresses and forces that constrain railway systems, which must be effectively and regularly monitored and evaluated.7–9 In response to this critical need, this review provides a concise and effective summary of the main strategies used in railway infrastructure monitoring, aimed at all possible stakeholders, from researchers to public and private groups involved in the field of Structural Health Monitoring (SHM), with the aim of disseminating theoretical and practical knowledge to enhance their activities towards smarter, safer, more efficient and cost-effective solutions.
Climate change is responsible for an increase in the frequency and severity of extreme events, which are responsible for extensive damage to railway infrastructure,10,11 for example, deformations, displacements, settlements, damaged components, shear cracks during tunnelling and much more. 6 These potential sources of catastrophic failure can lead to dangerous accidents, justifying the search for an innovative maintenance paradigm, namely Predictive Maintenance (PdM), for railway infrastructure.6,12,13 In the past, infrastructure expansion helped to improve performance and meet increasing demand, but today it is too costly and time-consuming. In order to improve the availability, capacity and quality of service of existing infrastructures, it is therefore necessary to invest in modern SHM technologies and PdM tools that guarantee a fundamental optimisation of resources, leading to more efficient and sustainable infrastructures.6,14–18
Traditionally, railway infrastructure has been monitored through periodic inspections and preventive maintenance.19,20 Due to the vast extent of the rail network, classical inspection is generally carried out by sight, by walking along the tracks to look for visible defects and technical anomalies, 21 or by track-testing vehicles equipped with various types of sensors, such as accelerometers, ultrasonic systems and lasers, 22 which can provide information over a large section of track.23–25 Unfortunately, human inspection has several limitations: it is slow, costly and subject to potential human error in unfavourable environmental conditions. With the significant growth in rail traffic, traditional methods are time consuming and expensive for railway companies, while the high cost and low speed of diagnostic vehicles makes it difficult to implement a modern monitoring strategy that can gather information continuously.8,26 For these reasons, the current interest is in the development of new trackside instruments that can be installed at specific points to instrument only short sections of infrastructure, possibly activated by regular rail traffic, to provide information for innovative PdM methods.27,28 PdM based on Data-Driven models is one of the most promising approaches among emerging technologies, supported by the extraordinary growth of the Artificial Intelligence (AI) sector. 29 The AI performs complex statistical analysis using accurate Machine Learning (ML) and Deep Learning (DL) algorithms that automatically model system behaviour, detect trends and predict failures, improving system reliability. Data-driven PdM tools are a critical opportunity for the growth of digital SHM in the railway sector. 30 To achieve their objectives, these innovative methods will require new and sophisticated tools. As a result, interest in networking technologies, wireless sensors and integrated devices that can autonomously monitor infrastructure is growing rapidly.31–35 In particular, the use of Wireless Sensor Networks (WSNs) for railway monitoring is being investigated in several research papers,36–38 which analyse the possibility of using these new technologies together with modern AI monitoring tools for better and automatic data management. Therefore, the integration of state-of-the-art technologies and artificial intelligence models provides a key tool for collecting and processing real-time data to support decision-making, in this specific case by classifying, predicting or detecting anomalies to determine the current and future state of the monitored infrastructure.36,39–41 The new monitoring algorithms are also grouped according to their complexity and functionality, in line with the classification of damage identification levels commonly used in SHM. The simplest Level 1, Damage Detection, systems detect the presence of damage without locating it, while Level 2, Damage Localisation, also provides localisation information. Level 3, Damage Classification, instruments allow classification of the extent of damage and Level 4, Damage Severity, tools are able to estimate the consequences of damage and the remaining useful life. A Level 5 system, Damage Prognosis, is more complex, a set of bespoke hardware and software tools that provide diagnosis, prognosis and suggested solutions to the problem.36,40,42
To reliably identify track damage, the railway infrastructure can be analysed in three sub-categories: track irregularity, rail condition and rail support component condition assessment.5,43 Specifically, the track irregularities provide an assessment of the profile and track quality; the rail conditions are divided into surface defects, rail failures and the integrity of the joint elements; finally, the rail supporting components concern the trackbed stiffness, the ballast/sub-ballast/subgrade stiffness losses and the hanging sleepers. In addition, the infrastructure is subject to critical loading components such as rolling contact pressure, bending and shear forces due to the weight of the vehicle, thermal stresses due to the restrained elongation of continuously welded rails, and residual stresses due to manufacturing processes such as roller straightening and welding in the field.44–46 If the axle loads, deflections and thermal stresses on the rail are high enough, failures can occur, rapidly progressing from initial cracks to catastrophic damage such as derailments. The most common rail defects include rail corrugation, rolling contact fatigue then shelling, squat defects, oval taches, vertical or horizontal split heads and wheel or engine burns. 47 These defects can be categorised as follows: rail head defects as localised battering, flaking and long grooves, rail web defects as horizontal crack, vertical splitting, bolt hole fatigue within fishplate limits and diagonal cracking, foot defects as vertical and transversal crack. Despite significant technological advances in modern maintenance strategies aimed at train and infrastructure safety, rail failures are still too frequent. To reduce the risk of damage and breakage to railways and avoid potentially catastrophic consequences, there is an urgent need to invest in new technologies and strategies for SHM and timely PdM.
One of the main complexities of this challenging transformation is related to the thousands of kilometres of track that need to be continuously inspected and for which traditional solutions, that is, visual inspection, slow and time-consuming diagnostic trains, have not yet provided a sufficient answer. An effective response could come from factors such as the mass, stiffness and damping of the main components of the railway infrastructure and their interfaces, as well as the track irregularity profile, which have a direct impact on the dynamic responses to the passage of the train. 48 The purpose of modern monitoring systems is to assess and recognise the condition of infrastructure by correctly associating the measured responses with the corresponding damage or health states. This approach promises to provide effective and efficient monitoring of large sections of track using a limited number of sensors installed at critical locations such as insulated bolted joints.
In addition to adequate instrumentation, robust signal processing methods are required to accurately distinguish system responses caused by damage from those caused by external factors such as environmental disturbances.49,50 For this reason, a growing number of researchers and companies are dedicated to the monitoring and maintenance of railway infrastructures, with numerical and experimental studies aimed at developing innovative, increasingly accurate and robust methods.5,43,51,52 The importance and originality of this review is that it addresses in a synthetic and structured way the crucial topic of SHM for railway infrastructure, a topic that is often overlooked or treated in a fragmented way due to its articulation and complexity. For new readers, this is a starting point to distinguish the salient components of the infrastructures, tools and technologies used in railway SHM; for experts, this review delves into innovative PdM methodologies, particularly those based on Data-Driven models, Digital Twins and AI.
Beyond the technical complexity of monitoring systems, it is necessary to reflect on the methodological orientation and economic sustainability that they offer. In fact, these systems not only allow for proactive maintenance planning through the rational allocation of resources, but also improve safety through the empowerment of predictive tools that allow for the anticipation of potential failures. Adopting these methodologies thus ensures the longevity of the infrastructure, improves safety and enables railway operators to manage maintenance resources efficiently.
Section 2 identifies and examines in detail the different parts of the railway infrastructure, describing the parameters monitored for each component, the tools and techniques currently used with possible future developments, the critical failures and application examples from the literature. In particular, this work focuses specifically on the track and substructure and does not discuss in detail other components of the railway infrastructure, such as power supply systems, communication and signalling technologies, and civil engineering works, such as bridges and tunnels. Highlighting the growing attention to PdM and its benefits, Section 3 examines the digital transformation of SHM for railway infrastructure, strongly supported by modern AI tools. Finally, Section 4 presents the authors’ views and comments on the previous sections and suggests current implications and future developments.
The monitoring targets of a railway infrastructure on the track
To examine monitoring approaches in the railway domain, it is necessary to provide a concise description of the physical context in which rail transport takes place: the railway structure, which is designed to allow trains to move by absorbing the loads transferred to the track. For these reasons, the rail structure has a layered construction, with each layer having a different thickness and decreasing mechanical properties from top to bottom. 53 The characteristic elements of a classic railway track, also called ballasted track, are shown in Figure 1: are generally divided into two main categories, namely: (i) the superstructure, consisting of a layer of ballast on which the sleepers are positioned and fixed to the rails using fastening systems; it is the upper part of the railway structure and plays a crucial role in transmitting and distributing the wheel load to the lower layers, reducing it to an acceptable level 54 ; (ii) the substructure, whose function is to support the railway superstructure, is, therefore, designed to have limited settlement and a long service life. It consists of two foundation layers: the sub-ballast, in the upper part, generally made up of unbound granular materials or the more recent bituminous or cement-bound materials; the subgrade, in the lower part, composed of high bearing granular soil. 55 In this article, the authors focus on the infrastructure specifically associated with track development. However, it is important to note that the concept of railway infrastructure also includes track support (side slope, embankments, drainage channels, etc.) and track configuration (level crossings, tunnels, bridges, etc.), which are not covered by this review. In more detail, the individual elements are defined as follows 56 :
Rails: are steel profiles of specially shaped sections (typically of the Vignole type) with the function of supporting and guiding railway vehicles. The upper part, on which the motion of the wheel occurs, is called the rolling surface.
Sleepers: are prismatic elements with a variable cross section, symmetrical to the track axis. They link the rails together, maintain the correct spacing (known as the track gauge) and distribute the stresses caused by the rolling stock passing over the ballast. Today, they are usually made of pre-stressed reinforced concrete.
Fastening system: includes all the components which together constitute the structural connection between rail and sleeper. The fastenings must elastically absorb the rail forces and transfer them to the sleeper, they must dampen vibration and impacts caused by traffic, they must retain the track gauge and rail inclination, and they must provide electrical insulation.
Ballast: part of the superstructure made of crushed stone, used to ensure the correct distribution of vertical loads and to absorb the stresses induced by rail traffic and thermal variations giving elasticity to the track, as well as maintaining the geometric design conditions and guaranteeing the drainage of rainwater.
Sub-ballast: intermediate layer between the ballast and the subgrade with the function of distributing the dynamic loads on the platform surface and maintaining the moisture content of the subgrade facilitating the drainage of rainwater towards the drainage works.
Subgrade: first layer of the embankment, which must support a stable track structure, provide the desired line, and level for the track and provide a smooth, level surface on which the upper layers can be laid.
Characteristic elements of the traditional railway system.
Once the different parts of railway infrastructure have been identified, a specific subsection is created for each of them in which to investigate the monitoring techniques and instruments currently employed, the parameters of greatest interest, the main anomalies and faults that have been recorded, finally reporting successful application cases to confirm the importance of the modern development and diffusion of Structural Health Monitoring in railway field.
Superstructure
Monitoring of the railway superstructure can be carried out using sensors that consider the geometric characteristics of the track, analysing the superstructure, or by monitoring individual sections that need to be examined in detail because they are subject to critical failures of the component under investigation. After an overview of the geometric characteristics that are periodically monitored, we will first consider monitoring systems that assess the condition of the track as a whole and then, in specific subcategories, we will examine specific monitoring instruments and techniques for different railway components.
The parameters for assessing track condition have been clearly defined for the entire European railway system, by the Technical Specifications for Interoperability (TSI) annexed to Regulation (EU) No. 1299/2014 57 and the European Standard UNI EN 13848. 58 The European standard defines the basic parameters that characterise the track geometry, specifies the measurement requirements and data analysis methods. The track geometric parameters defined by the standards, which must be monitored to guarantee the necessary safety conditions for railway operations, are:
Gauge, the smallest distance, expressed in mm, between the lines perpendicular to the running surface that intersect each rail profile, it is measured at 14 mm below the rail running table. The gauge is also subject to a widening in circular planimetric sections, depending on the value of the radius of the curve, which is necessary to allow the railway vehicle to enter the curve, since the two wheels are rigidly keyed to the axle. Gauge errors occur when there is a displacement of one rail relative to another.
Alignment, the measurement of the transverse deviation of the successive positions of the point for each rail, expressed as a displacement from the mean horizontal position (reference line). The alignment or curvature defects occur when there is a displacement of one or both rails from their design position or their relative position. The individual defects are represented by the amplitude measured between the mean value and the upper or lower peak value.
Vertical alignment, the measurement of the vertical deviation of the distance of the rolling plane of each rail from the mean vertical position (reference line) and is calculated from successive measurements. The vertical alignment is closely related to the vertical acceleration at the axle boxes.
Cant, the measure of the difference in height between two adjacent rolling planes; it corresponds to the height of a right-angled triangle with the hypotenuse equal to 1500 mm and the angle at the vertex equal to the angle between the rolling surface and a horizontal reference plane. The cant errors occur when the two rails are not at the same height on straight track and when they have cant values different from the design values on curves.
Twist, the inclination expressed as a relative % of one row of rails compared to the other, calculated as the ratio between the transverse level difference between two track sections placed at a given distance, which is the basis of measurement of the twist, and the basis itself. The twist is representative of the torsional stresses acting on the track, causing different forces on the wheels and rails, leading to geometric deformations, oscillations and, consequently, derailment of trains.
Equivalent conicity, the conicity should have an axis with a perfectly conical profile for a given lateral displacement to describe a regular sinusoidal motion with the same wavelength (i.e. identical to that of the real axis). This parameter depends on the track gauge, railhead profile and rail inclination values.
About the rail profile, the TSI defines the characteristics of the rail head profile. Otherwise, rail defects can affect the profile, in this case we talk about wear defects, or the internal structure of the rail. Defects in the rail structure are cracks, which can be vertical, longitudinal, or transverse, cracks typically starting from the joint holes and defects in the surface of the rail head. Rail wear defects, such as corrugations, can affect the rail head with wavelengths ranging from a few centimetres to 10–20 cm, causing noise and vibration and further damage to the rails, and can be detected by wheel-rail vibration detection systems.59,60 Defects in the track are essentially the result of these parameters or a combination of them. In fact, there is a correlation between some of these parameters, for example it has been shown that skew defects are directly correlated with other geometric track parameters. 61 Therefore, to avoid such defects, it is necessary to control the parameters from which they originate. Data-Driven models and analysis, using Artificial Intelligence techniques, described in detail in Section 3, constitute one of the most effective and innovative methods for stabilising the relationship between skew and other geometric parameters of the track. The parameters that have the greatest influence on the skew defect are alignment and cant (transverse level), followed by vertical alignment. 61
Track quality monitoring systems generally consist of various types of sensors that can be installed either on board the train or at fixed points on the superstructure. The first type allows the monitoring of long stretches of track but is limited to the passage of appropriately equipped vehicles. Mobile diagnosis is therefore carried out using Track Recording Vehicles (TRVs) or Track Recording Coaches (TRCs), that is, rolling stock equipped with measuring systems that record parameters indicative of the condition of the track as it travels along the lines. An example of the diagnostic trains used on the Italian rail network is shown in Figure 2.
The diagnostic trains of Rete Ferroviaria Italiana: (a) diamante is the train used for the inspection of high-speed lines and (b) talete is the train used for the inspection of complementary electrified lines. 62
These monitoring strategies are widely used on high-speed lines, where wayside monitoring is more complex due to the dynamic effects induced by the passage of high-speed trains. However, the discontinuous nature of the diagnostic trains has favoured the development of instrumentation compatible with the passage of high-speed trains. The systems currently being developed must therefore address the specific problems of high speed and allow their use in this context. The measurement of track geometric parameters by mobile diagnostics according to EN 13848 is based on two types of systems: the string measurement system and the inertial measurement system. The values of these parameters can be acquired in predefined steps, and they are returned through graphical and numerical outputs to be manipulated analytically. 63 The first method involves measuring three points on a string using optical measuring units. The track measuring system has a local reference system, which can be the vehicle body if it is very rigid, or the bogie if it is not. Typically, non-contact sensors are used in this type of system. In contrast, inertial measurement systems consist of three gyroscopes, which define an inertial reference system, and three lateral and vertical accelerometers, positioned on the vehicle body, undercarriage, and axle boxes, which evaluate the vertical and lateral position of the unit with respect to the reference system. The axle box and body accelerations can be used to detect individual defects, for example, the irregularities on the rolling surface of the rails, known as rail corrugation.64,65 The track geometry monitoring systems installed on board diagnostic trains are often a combination of non-contact optical and inertial technologies (Figure 3). In general, dynamic monitoring, carried out by means of mobile detection vehicles, and static monitoring, carried out by means of fixed equipment positioned along the track, can be effectively integrated into the overall inspection and monitoring system. In fact, these two complementary approaches allow a more complete assessment of the condition of the infrastructure: dynamic monitoring provides data on the behaviour of the line under load and during operation, while static monitoring ensures a continuous and localised control of the structural condition, making it possible to detect anomalies that could not be detected by a single method.
Track geometry monitoring system for mobile diagnostics consisting of an inertial unit and two optical sensors. 66
Track level measurements can now be carried out using photogrammetric systems, using suitable wayside systems that do not interfere whit railway operations. Devices such as high-resolution cameras 67 can take measurements by recording objects of interest from different angles.68,69 By applying photogrammetric algorithms to the resulting 3D image, it is possible to measure parameters such as gauge, alignment or vertical level of the track. Wayside systems also include tools that use laser technology to monitor the parameters that characterise the continuous welded long rail during operation by means of specific targets positioned on the rail, 70 thus allowing the deviation of the continuous welded rail from the set temperature to be calculated automatically and in real time. Such a measurement could anticipate and thus prevent instability phenomena caused by static and dynamic loads acting on the rail. More recent applications include the use of unmanned aerial vehicles (UAVs) in railway monitoring. Inspections carried out by drones provide less accurate results than track-measuring vehicles, but they are a promising alternative as they are safe (they do not require track access), fast and cost-effective. 71
As mentioned above, civil engineering works such as tunnels are also part of the railway infrastructure. Tunnel outlines are also monitored using inertial-optical systems. In this case, retroactive optical systems are used to reconstruct cross-sections and inertial systems are used to assess the track elevation parameter. Monitoring of railway tunnels is useful to detect cracks and settlements in the structure and to minimise the risk of collision between rolling stock and the tunnel. In recent years, railway infrastructure managers have attempted to implement automatic track geometry measurement systems (UGMS) on commercial service vehicles, which, unlike diagnostic trains, don’t require additional tracks for normal train operation. 72 Rail defects, on the other hand, can be monitored using ultrasonic systems in combination with the above-mentioned optical systems, which make it possible both to assess the evolution of any cracks in the rails and to estimate the defect characteristics, 73 and to reconstruct the sequence of profiles to monitor rail wear. Although track quality monitoring via diagnostic trains is efficient overall, continuous measurement of the condition of the track, both under static and dynamic conditions (during train passage), at certain localised points on the railway network, for example, near a bridge or a switch, on long embankments or in tunnels, is necessary, in particular, to support the development and implementation of modern PdM strategies based on digital algorithms for SHM of railway infrastructures.
Several multi-sensor stations can be placed along the railway line to extrapolate laws and trend curves useful for evaluating the track parameters. Examples include systems that measure passing loads, compare them with the specifications of the line on which they are installed and compare these loads with the track capacity. These systems can also provide information on the type of vehicles in transit by identifying the number and weight of each axle of the trainset. By knowing the geometry of the passing train and therefore the distance between successive axles, these systems can also calculate the speed at which convoys pass. These instruments are generally strain gauges installed directly on the rails or sleepers.74,75 Fibre optic sensors (FOS) are currently widely used for fixed track monitoring systems. Thanks to their versatility, they can measure physical quantities (such as displacement, inclination and strain) along the entire length of the fibre, typically placed longitudinally on the rail. FOS technologies are based on changes in the physical properties of light waves propagating in optical fibres caused by external signals/stimuli. The most used fibre-optic sensors for measuring track geometry are Bragg grating (FBG) and optical time domain reflectometry (OTDR) based sensors. 6 Finally, it is worth mentioning infrastructure monitoring systems that use MEMS sensor technology to detect not only longitudinal and lateral track deformation, but also track inclination (Figure 4).
Non-conventional monitoring system designed for automatic monitoring of rail track geometry. 76
These systems are typically used to monitor areas affected by subsidence, which can cause track deformation. Micro- and nanoelectromechanical systems, MEMS and NEMS, are already successfully employed in various fields, from medicine to industry, from automotive to aerospace and promise to innovate the SHM field thanks to new miniaturised devices such as high precision sensors, nanoactuators and memory devices characterised by high accuracy and reliability with low power consumption. 77
Below are the detailed sections relating to the individual parts of the railway infrastructure.
Rail
The use of the Continuous Welded Rail (CWR) is now prevalent due to the advantages of improved ride comfort and reduced maintenance costs. However, the use of this solution meets some difficulties in areas where there are significant annual temperature variations, as instability problems can occur due to the excessive longitudinal forces caused by these variations, especially in curves with limited radius. 78 The uneven temperature change further increases the longitudinal force, which may lead to buckling or cracking of CWRs. 79 To avoid these problems, the rail can be fitted with fixed instrumentation at sensitive points to monitor these longitudinal forces, such as diagnostic systems with a magnetostrictive displacement sensor installed on small radius curves, 80 or other systems such as strain gauges 81 or the more innovative bi-directional fibre Bragg grating strain sensors, which can provide higher accuracy. 82
Rail joints
Rail joints are used to form a connection between separate lengths of rail. Geometric deviations must be small enough to limit dynamic effects. They have reduced mechanical properties compared to the rail, resulting in lower vertical flexural stiffness. This results in an increase in vertical deflection as rolling stock passes through, leading to increased deterioration. 83 This component is therefore subject to regular maintenance to ensure the safe running of trains. These operations consist of periodically retightening the bolts, checking the correct electrical insulation in the case of insulated joints, and finally replacing them when they have reached the end of their useful life. The problems of modern communication and control systems for rail traffic management, typical of high-speed networks, have been solved by the implementation of the European Rail Traffic Management System (ERTMS). However, in conventional rail networks using insulated rail joints (IRJs), the monitoring of these components is an essential strategy to ensure the safety, efficiency and reliability of the infrastructure. By installing sensors on the joints, it is possible to establish their range of normal movement and thus to identify appropriate alarm thresholds useful for detecting possible failures due to an increase in lability or the initiation of cracks from the holes. Non-destructive instrumental inspections, which typically involve the use of diagnostic trains or the measurement of axle box acceleration, 84 are often inadequate for detecting faults in time, due to their inherently discontinuous nature. Continuous monitoring would make it possible to anticipate these failures, allowing key parameters to be determined in time to prevent possible failures. Among the parameters that can be monitored in a railway joint are the wheel-rail impact 85 or the value of the joint gap, that is, the variation of the recorded opening caused by the stresses induced by the passage of rolling stock or thermal variations. 86 Continuous measurement of this parameter allows maintenance work to be planned since the actual condition of the joint. Fixed systems such as strain gauges, fibre optics or ultrasonic sensors can be used in this element, either mounted on the joint bar or directly by applying the sensors to adjacent rail heads as shown in Figure 5.
Monitoring of rail joints by installing sensors at the head of adjacent rails.
Sleepers
Monitoring of the sleepers allows you to anticipate some problems, such as excessive bending deformation, which reduces the life of the track and makes maintenance interventions necessary to avoid vertical deviation of the rolling plane level or widening of the gauge. Therefore, geophones attached to the sleepers with brackets can be used to assess the dynamic response of the sleepers. 87 Using these systems, it is possible to assess the presence of track defects by measuring the movement of sleepers as rolling stock passes through, as geophones can accurately measure the peaks caused by passing wheels. In addition, geophone measurements of track movement can be supported by the installation of cameras which, by digitally correlating images using specially positioned targets, can measure sleeper movement by correlating it with the passage of trains.88,89 An alternative to the use of geophones is the more consolidated strain gauge, which can also be mounted on sleepers. Using this monitoring system, it is possible to obtain data on the dynamic bending of sleepers, which can be implemented in degradation functions that can represent, and thus predict, typical bending failure modes. 90 In the same way, the latest fibre-optic sensors can be used as an alternative to classic piezoelectric sensors, offering higher measurement accuracy, distributed coverage and greater resistance to electromagnetic interference. 91 In addition to evaluating deformation, these instruments can also measure vibration, which can be used to assess any rail running table defects or wheels of passing trains. The limitations of installing instruments on the surface of the sleeper can be overcome by using integrated monitoring systems (IMS). This innovative technology, already present on the market, pre-configures ‘smart’ sleepers with the aim of extending the service life of sleepers or other track components. 92 The innovation consists in the incorporation of sensors, either embedded or surface-mounted, capable of detecting the structural response of the sleeper as well as its performance in a railway track, without having to plan the fixing systems of the instruments to be used. The sensors placed inside the sleepers use Fibre Optic Technology (FOS), in particular Bragg Grating (FBG), and Radio Frequency Identification (RFID). FBG sensors can be embedded in two ways: longitudinally in the sleeper or horizontally in the centre and rail seats. The first method allows data to be collected on bending moments, the effect of train speed, load magnitude, ballast quality and track substructure integrity. It also enables the estimation of rail seat loading, the detection of cracking and the monitoring of the early ageing behaviour of concrete sleepers. The second method allows the observation of ballast freezing, rail pad stiffening in cold weather and sleeper mechanical behaviour, including maximum stress and strain. RFID sensors, on the other hand, have been used to detect the long-term change in impedance and radiation resulting from a fracture in a structure. As a result, the loss of material beneath or around the RFID can be monitored as a result of cracking. 92 However, this solution can also be achieved by using “smart” sleepers when renewing the track; alternatively, existing sleepers can be adapted using special systems available on the market.
Ballast
Failure of some parts of the ballast will result in high rates of track geometry degradation in the vertical and/or lateral directions, necessitating periodical track maintenance. Deterioration manifests itself in the changes in track geometry, such as vertical and lateral alignment errors. As mentioned above, the ballast consists of crushed stone, one of the functions of which is to give the track sufficient elasticity to distribute the stresses caused by the passage of rolling stock. Therefore, it is difficult to envisage a fixed monitoring system being installed in this part of the railway superstructure without extensive and potentially damaging excavation work. For example, the cone penetrometer, commonly used in soil testing, has been adapted for railway purposes. In the dynamic penetrometer test, a cone of predetermined shape is used to measure the force required to advance the tip continuously. This test makes allows to assess the depth of influence of the ballast on track deflection. However, this test does not provide parameters that can be directly correlated with ballast fouling, as the distribution of the data obtained is not standardised. 93 For the detection of problems that lead to reduced layer performance, such as ballast fouling, mobile detection methods are more suitable, taking advantage of the periodic passage of diagnostic trains or using vehicle-mounted instruments (the TRVs and TRCs mentioned above), such as LIDAR for ballast measurement profile or Ground Penetrating Radar (GPR). LiDAR which stands for Light Detection is an optical remote sensing technique that uses laser light and analyses the backscattered light. 94 The result of this instrumentation is a map that defines the missing ballast (including shoulder and crib ballast) per length of track. Furthermore, such instruments can be used to assess the drainage conditions of the ballast, and by supplementing this analysis with data from GPR, it is possible to determine track geometry degradation rates. 95 The GPR consists of a transmitter system, a receiver, a system for reading the waves and one for the graphic display of the survey. The impulse emitted by the source travels downwards in the ground by radiation until it is reflected by an obstacle or, more generally, by a discontinuous surface. The reflected wave is picked up by the receiver, which analyses it and displays it graphically. Therefore, GPR uses the transmission of high frequency electromagnetic radar pulses into the ground, measuring the time interval between the transmission and reception of the pulse. The pulse is transmitted into the ground via a transmitting antenna, which is positioned close to the ground surface and converts the electrical current into electromagnetic waves and vice versa. This antenna can be changed so that the instrument can operate at different frequencies. 96 In addition, this monitoring system can be supported by photogrammetric images taken by Unmanned Aerial Vehicles (UAVs), allowing quantitative risk assessment. 97
Substructure
Installing a monitoring system under the ballast is generally a more complicated operation than monitoring the superstructure; generally, these systems must be designed during the construction phase of the work. An example of this is the construction of the Bretagne - Pays de Loire High Speed Line (BPL HSL), where six test sections were instrumented using more than 100 sensors. 98 In this case, the design of the monitoring system, which included accelerometers and strain gauges placed at the top and bottom of the sub-ballast layer, allowed the dynamic response of the substructure to be assessed by comparing two types of sub-ballast used in the construction of the line. Conversely, examining the condition of the existing railway substructure using fixed systems requires inspection holes. These monitoring systems are different from those discussed so far in that they relate to slower phenomena not necessarily linked to the passage of trains. For example, if the aim is to assess the possible accumulation of deformations on the substructure, which could lead to differential movements that could affect the safety of rail traffic, a series of variable displacement transducers can be used to form multi-depth deflectometers. 99 However, the installation of such devices requires reaching great depths to be used as an immovable reference, making their large scale use difficult. Alternatively, the use of water pressure sensors is required to trace any problems to a change in the interstitial pressure of the subgrade. In railway applications, tube piezometers or vibratory piezometers are generally used, which can be installed at the interface between the ballast and the lower layers, but they are not exempt from the stresses induced by railway traffic, making long-term analysis difficult. 28 Furthermore, in interstitial pressures, it is possible to assess the vertical deformation employing settlement probes, which measure the pressure variation with respect to a fixed datum point. In this case, it is necessary to instal a probe in the ground connected to an external tank; such systems are small and low in cost. 100
Data-driven predictive maintenance for railway infrastructure
Having provided the reader with a concise and comprehensive analysis of the most interesting components for railway infrastructure monitoring and the most widely used sensors, it is necessary to understand how to make the best use of the measurements obtained, that is, that only through automatic information management methodologies will it be possible to overcome the current limitations and realise modern sustainable infrastructures with high efficiency, reliability, availability and safety. The multiple economic, environmental and developmental benefits for the communities involved are prompting an increasing number of public and private entities to invest in research into innovative digital technologies and techniques related to the vast world of Artificial Intelligence (AI), such as Machine Learning (ML) and Deep Learning (DL).
With the support of the material available in the literature, it is possible to describe the evolution, that is, the fundamental transformations that SHM has undergone in recent years, moving from corrective and preventive maintenance to the emergence of Data-Driven PdM, based on the use of large amounts of data managed and processed automatically thanks to modern AI tools, in fact, this is the era of Big Data.
To implement and employ a monitoring technique effectively you must clearly define and understand the system you are investigating. Currently, in fact, there is no universal approach, and it is typically not possible to directly use the same methodology for different systems. This explains, at least in part, why, although several innovative techniques are already successfully used for monitoring complex industrial systems, their use in the railway sector is only just beginning. In Duque et al. 101 a useful comparison is reported between an industrial system and a large-scale distributed system to which the railway infrastructure is comparable. In summary, the characteristics by which an industrial system differs from a large-scale distributed system specifically reported, 101 are a geographically compact system versus a geographically extensive system; a known uniform environment that is stable compared to different changing environments; limited number of complex components versus a large number of technically unsophisticated components.
After investigating the system characteristics to develop and implement an effective SHM strategy for railway infrastructure, it is necessary to have a good knowledge of existing techniques in both general SHM and railway monitoring techniques as suggested in Brahimi et al. 102 where before moving on to the identification of technologies and techniques for the specific maintenance of the railway infrastructure, ample space is dedicated to the common aspects of the methodologies and tools used in Prognostics and Health Management (PHM). Convinced that this is one of the best ways to gradually present the topic, the authors have decided to adopt it, dividing Section 3 into two parts. The first part is aimed at exploring the topic of innovative PdM techniques based on Data-Driven and AI models, in the second part some significant applications of these techniques for monitoring the railway infrastructure are reported. The flowchart of Section 3 is shown in Figure 6.
The flowchart of the Section 3. The first part 3.1 investigates Predictive Maintenance techniques based on Data-Driven and AI models, the second part 3.2 reports some significant applications of these techniques for monitoring the railway infrastructure.
Architecture and development of Prognostics and Health Management System
Let’s start with a simple definition according to ISO 13881-1, Prognostics (P) refers to the ‘estimation of time to failure and risk for one or more existing and future failure modes’. The Health Management (HM) refers to the decision-making ability to effectively perform maintenance and other interventions based on diagnostic/prognostic information. 103 Furthermore, as specified in Brahimi et al., 102 the PHM is an innovative engineering discipline consisting of technologies and methods for evaluating the reliability of products and systems in the actual conditions of their life cycle to determine the onset of failures and reduce the system risk through the support of a Condition-Based Maintenance (CBM) strategy. 104 The CBM includes all those innovative maintenance methods in which interventions are planned based on the actual health status of the system, this is the core of modern PdM that significantly reduces maintenance costs by decreasing the number of unnecessary planned interventions and optimising availability at the same time, interrupting the service only in the event of proven criticality. 105
Although there is no canonical methodology for the implementation of a PHM strategy, several authors have investigated the field of prognostics, looking for common aspects.106,107 Some authors have even suggested a global architecture for PHM.102,108 This architecture is divided into seven layers which are, in turn, grouped into three main areas. The first is the Health Indicator Area which includes the first two layers, Data Acquisition (DA) and Data Manipulation (DM). The second Health Indicator Processing Area includes the Health Assessment (HA), Diagnostic (D) and Prognostic (P) layers. Finally, the third area, the Health Management Area, includes Decision and the Human-Machine interface. The first two layers, DA and DM, are used to collect monitoring data and preprocess it in order to extract useful features, provide parametric data or build health indicators (HI), which are used in subsequent layers, the HA and Diagnostic, to detect anomalies, estimate the current state and run diagnostics. 102 The Prognostic layer aims to predict future states using the outputs of previous layers. One of the most widespread prognostic outputs is the Remaining Useful Life (RUL) for a component or subsystem, then RUL passes to the Decision layer where recommendations and actions for maintaining the system are automatically generated. Although it is still difficult to identify the best PHM strategy for the system to be monitored, the reported architecture lays the foundation for innovative and effective PdM, which has all the potential to be successfully employed in SHM of railway infrastructures in the years to come.
The models used to evaluate the conditions of the railway infrastructure for diagnostic and prognostic purposes can be grouped into three categories: mechanical, Data-Driven and hybrid prognostics models.102,105 Mechanical models are developed starting from physical models that describe the behaviour of the system in steady state considering the possible occurrence of faults. Designing models of this type requires excellent knowledge of the system and the failure mechanisms, which is often not exhaustively available due to the numerous variables that influence the system. For this reason, simplified assumptions are typically used for the components of interest. On the other hand, Data-Driven models, which do not have such dependencies, have been increasingly applied in PdM of railway infrastructure. Data-Driven methods are based on the use of AI, for example, Neural Networks or Machine Learning models such as Support Vector Machines and Bayesian networks. The main advantages of this approach are the possibility of implementing the prognostic model more quickly than classic methods and with lower costs, furthermore these models provide an automatic, real-time, and reliable reading of data relating to the actual conditions of the infrastructure. On the other hand, the physical behaviour of the system and the failure mechanisms are unknown, making it difficult to automatically explain the causes of failures and thus generate adequate recommendations and correct maintenance responses. For this reason, it is interesting to consider a hybrid prognostic approach, that is a combination of the two previous models that can be organised in parallel by combining the outputs of each mechanical and Data-Driven approach, or in series, using Data-Driven models to adjust the parameters of the physical model therefore increase its accuracy and robustness to disturbances.
An overview of data-driven models
Since the analysis of railway measurement datasets with Data-Driven approaches has recently become an area of strong interest, both in academia and industry, this is the type of model the review focuses on. Data-Driven models discover feature sets and viable decision criteria from observed data. These methods include statistical models and ML models. 4 Both allow you to handle high-dimensional, multivariate data and find hidden relationships between infrastructure health conditions and monitoring data. The main difference between statistical and AI-based models concerns the objective of the analysis itself, that is, statistical models provide better inferences about the relationships between parameters, whereas AI models produce the most accurate predictions, both in regression and classification. 109 In detail, a statistical model estimates a limited number of parameters from a sufficiently large collection of samples, assuming that the data respect a precise hypothesis, such as a statistical distribution. 110 Instead, AI models estimate large numbers of parameters from huge sample datasets, without the need for a priori knowledge. 111 In Xie et al., 109 the trend of publishing research works on Data-Driven methods for PdM shows the growing interest of the scientific community in this sector, which has seen strong success since 2015 due to the capabilities of modern advanced models to exploit and cope with the modern characteristics of large-volume, multi-source, highly unbalanced and high-noise railway datasets. Furthermore, to better understand the current trend we report that 74% of the articles reviewed in Xie et al. 109 use AI models and only 26% classical statistical models. Therefore, Data-Driven methods, and especially AI models, help railway engineers to better understand the conditions of railway infrastructure and make corresponding maintenance decisions more efficiently and robustly, however, we must not forget that the performance of Data-Driven methods depends on the appropriate choice of data acquisition, preprocessing and analysis models. Therefore, the following paragraphs of the review are dedicated to understanding and in-depth analysis of these critical aspects.
Data acquisition for railway infrastructure
As systematically suggested in Xie et al., 109 to develop and implement an innovative monitoring strategy, it is necessary to understand which measurement methods and data sets to use, and how to select and deploy appropriate Data-Driven models for railway infrastructure PdM. The main characteristics of railway monitoring data are large-volume, multi-source, highly unbalanced compared to normal behaviour and high noise. 109 Specifically, we are talking about large volumes of data because it is necessary to collect enormous quantities of data both in the Time Domain (real-time or event approach) and in the Space Domain (thousands of kilometres for the railway network). 112 Multi-source refers to the various measurement methods from which data can be generated in multiple typologies. 113 Highly unbalanced since railway defects have a highly skewed distribution in the collected datasets, in fact most of the observations belong to normal states while only a small part is related to defects. 114 Finally, the high noise during data collection arises from two aspects, the intrinsic environmental uncertainty along the track such as soil type, climatic conditions, track profile and materials, on the other hand the performance of the sensors. 113
Data acquisition is the first level to implement a Data-Driven model. To better understand the next parts relating to Data-Driven models, we briefly summarise the most used measurement methodologies; a full detailed description is included in Zhao et al., 37 Alahakoon et al., 45 and Xie et al. 109 Human inspection is time-consuming and only provides data at a certain rate, making it unsuitable and unsafe for a modern automated strategy. Another limitation is that visual inspection cannot detect defects within the rails, such as rail breaks and internal cracks, which can be detected by ultrasonic testing. 115 Despite its speed, ultrasonic testing is unable to detect cracks at an early stage, so eddy current and magnetic flux leakage measurements are used. Another strategy is to use diagnostic trains, but these are rarely used on main networks due to their low speed and therefore produce discontinuous data sets. A step forward is to equip standard trains with daily service to obtain continuous measurements for Data-Driven analysis. 116 Finally, the wayside strategy involves the use of fixed sensors along the track that collect information on parts of interest, such as insulated bolted railway joints. Wayside sensors use a variety of technologies, measuring force, heat, sound, geometry and other values, 115 proving very robust and versatile, but at a high cost. One area that has found widespread use is that of fibre-optic Bragg grating sensors, which are favoured by their immunity to electromagnetic interference, multiplexing capability, long reach, lightweight and high signal fidelity. 117 Another fundamental aspect for the purposes of a correct Data-Driven model is that the wayside strategy can consider the environment in which the tracks are placed, including information such as temperature, humidity, and other atmospheric conditions.
In light of these considerations, it is clear that there is a need to create an innovative monitoring system to collect real-time data from all the main components of the railway infrastructure, based on modern AI algorithms and Big Data management tools that have demonstrated increasingly exciting and powerful computing and processing capabilities in recent years. As described in Section 2, the use of various monitoring tools is required to correctly evaluate the different structural behaviours of the components. Given the limitations of pre-AI tools in managing combined data, the classic approach has been to use and interpret measures targeted at individual components. Today, however, we support the importance of developing an integrated approach for correct and effective PdM of the infrastructure, that is, the synergistic use of multiple indicators to achieve a comprehensive prediction of track failure by implementing an appropriate temporal management of resources.
The digitisation process of SHM for railway infrastructure is currently one of the major technological and scientific challenges facing the sector. To this end, Data-Driven models based on AI and Big Data are proving their usefulness and effectiveness. However, given the importance and confidentiality of monitoring data for the development of the national infrastructure, and given the large amount of data needed to train and test predictive tools, there is a need to bring together different stakeholders, from universities and private groups to the infrastructure manager, all heavily involved in the common field of railway PdM. This is confirmed by the successful Italian experience of the project MOST – Centro Nazionale per la Mobilità Sostenibile 118 – Spoke 4 Railway Transport – on efficient and effective rail transport, in which several academic and industrial partners are involved, driving the digitalisation of railways to improve safety, maintenance efficiency, railway asset management and environmental impact. MOST is an implementing project of the National Recovery and Resilience Plan (NRRP) as part of the Next Generation EU (NGEU) programme. 119
Data-driven models for predictive maintenance
Once the data has been collected, it needs to be processed with the appropriate Data-Driven models. Below is a brief but effective summary of Data-Driven models suitable for use in railway PdM, as more information on these models is available in the specialist literature.4,109,120,121 Traditional Data-Driven models are divided into statistical models and classic ML models. Among the statistical models we find regression modelling, probability distribution model, time series model, Bayesian methods, stochastic process. The main advantages concern simplicity, interpretability, stability; the disadvantages require prior knowledge to select the best fitting algorithm. Classic ML models are: Artificial Neural Network (ANN), Support Vector Machine (SVM), Tree-based model, K-nearest neighbours (KNN). For these methodologies it is more difficult to indicate common characteristics, we observe that Tree-based and KNN-based models are simpler and more interpretable but more sensitive to data distribution and noise, ANN and SVM are more robust and efficient but time-consuming and have poor interpretability. New advanced ML models use Deep Learning algorithms based on Neural Networks with large numbers of layers. The main advantage is the ability to learn directly from data, avoiding the long and sometimes complex pre-processing phase. These methods are proving to be suitable for working with different data such as images, video, audio, natural language reports. 122 The DL models used in the rail industry are Convolutional Neural Networks (CNNs), Recurrent Neural Networks (RNNs) and Long-Short-Term Memory (LSTM) models. 109 CNNs have revolutionised computer vision applications, this is extremely useful for tracking defect detection, ensuring human-like image recognition capabilities. In Faghih-Roohi et al. 123 a CNN approach was used to identify defects in the rail surface avoiding the classic and complex image feature extraction techniques. RNNs are proving useful in processing time series or sequential data from track measurements, such as track geometry data. 109 The advantage of an RNN is that by considering previous and current data simultaneously, this method provides better predictions for future trends. 122 On the other hand, the LSTM unit consists of input, output and forgetting gates that allow it to avoid the training error that RNNs suffer when the period becomes too long. Despite their proven value, advanced DL models still have some limitations, such as long training times and poor interpretability, which makes it difficult to explain and communicate the predictions obtained at an engineering level, in addition to the fact that when changes, even minimal ones, occur in the system to be monitored, it is not possible to make targeted changes to the solutions, but it is necessary to repeat the design from the beginning. Some of the latest influential AI technologies aim to overcome current limitations. In Ye et al., 124 an interpretable Track Detection Network (TrackNet) model based on self-attention and using transfer learning is proposed for the detection of ballastless track surface defects to ensure the safe operation of high-speed railways. This method enhances the fusion capability of global defect features and reduces its dependence on large datasets by directly transferring knowledge from publicly available datasets. Furthermore, in Min et al. 125 an algorithm for detecting rail surface defects based on semantic segmentation was presented. This approach uses the global information present in the image and avoids the problem of inductive preference. The efficiency of the model is further enhanced by window-based self-attention, which minimises the number of model parameters. A second type of advanced models involves the use of unsupervised learning for automatic pattern recognition from unlabelled data. 109 To date, the most used methods in railway engineering are clustering and dimensionality reduction techniques. These tools, described in detail in Xie et al., 109 Schalk et al., 126 and Li et al. 127 allow you to simplify the feature extraction, thus automatically identifying the indices that allow you to evaluate the conditions of the infrastructure. One approach that is attracting interest is the combination of two or more models to maximise individual advantages and compensate for disadvantages. They are called ensemble models and are mainly based on two techniques, namely aggregation and stacking, 109 which were employed and described in Cárdenas-Gallo et al. 128 to predict track geometry degradation and in Lasisi and Attoh-Okine 129 to study annual track fatigue defects. However, while ensemble models provide the best prediction from a scientific point of view, in practice these models require greater storage capacity and cost, with poor interpretability.
In practice, there is no method that is scientifically or technologically superior to others; experience in the field of AI for PdM must guide the choice of the appropriate method. This review aims to be a practical reference point for readers wishing to start or consolidate their knowledge in the field of innovative monitoring of railway infrastructure, a fascinating but complex sector that is constantly and rapidly evolving.
The choice of tool depends on the type of input data, the most common being time series, images, video, discrete values such as temperature and, in railway monitoring, text records. Cars and in-service vehicles that record track geometry provide time series data from which, using appropriate signal processing techniques, it is possible to extract features with obvious physical meaning, which are used to develop accurate classifiers. 109 For example, track degradation has been investigated using the Fourier transform to extract frequency domain features from vibration signals collected from in service trains. 130 Other anomaly detection applications have involved the use of wavelet transform and simplified indices such as the Track Quality Index (TQI), the latter of which has proven to be a useful decision support tool in the railway monitoring sector. Using classifiers such as linear regression, SVM, KNN it was possible to identify condition indicators useful for correctly determining maintenance activities. A CNN approach is presented in Sun et al. 131 for railway joint detection using acceleration data. Compared to human inspections, artificial vision represents a notable step forward, in fact Data-Driven models, and particularly those based on DNNs, can process a large quantity of images in a short time. 132 Some examples reported in literature show a real-time automatic vision tool to detect defects or missing components, such as connection plates, ties, and anchors 133 and a system capable of combining track inspection images and growth data of cracks from ultrasonic inspection. 134 Discrete-value data is log data collected with warning or fault messages during a specific event. They contain various information, temperature, track age, tonnage, infrastructure condition and can be analysed to identify the causes of a failure. Since this type of discrete fault information is only available in small datasets not suitable for modern DL techniques, Bayesian models have generally been used until now. 135 In addition to the proliferation of increasingly accurate and widespread sensor networks, Digital Twins, virtual environments in which faithful and accurate failure models can be developed and simulated, are used to extract critical information from the data to train predictive models. This approach, already successfully applied in the automotive and aerospace industries, promises to contribute in terms of efficiency, safety and sustainability to the future transformation of the railway sector. Finally, the diffusion of innovative RNN capable of automatically interpreting the content of human-generated textual data has made it possible to identify the causes of an accident directly from the railway reports that accompany periodic checks or anomalous events.
In addition to the type of input data, the approach to be used depends on the railway defects you wish to monitor. The most used Data-Driven models for monitoring railway infrastructure concern geometric irregularities and structural defects. 109 Rail geometry irregularities can be responsible for catastrophic consequences such as derailments, therefore it is necessary to promptly identify factors such as wide gauge, excessive warp or twist and horizontal and vertical rail deformations and keep them below acceptable safety thresholds. 129 To guarantee the safety of railway operations it is necessary to monitor the structural defects that appear due to wear in curves, fatigue such as surface or sub-surface initial cracks and plastic flows such as rail corrugations. 136 Among these defects, geometry irregularity, rail head defects and missing rail components are the most investigated with modern Data-Driven models that have only recently been applied to railway breaks and still present difficulties in predicting the substructure failures.
A further aspect to consider improving the reliability of maintenance approaches concerns the methods on which the predictive models are based, such as statistical methods, or models based on the Remaining Useful Life (RUL) of the components or directly on maintenance AI tools. 109 Typically, statistical models are used to predict the behaviour of the railway infrastructure in the long term by calculating appropriate degradation indices that allow the timing of maintenance interventions to be optimised, thus guaranteeing the correct state of the railway components. In PdM, RUL represents the time from present to the end of the useful life of railway components. The time series data and discrete data we described above are used to estimate an accurate medium-term RUL prediction and then plan appropriate maintenance responses. 137 In addition, it is now possible to use advanced AI tools to directly identify the best maintenance strategy. Since, among the various tasks, those that occur at high frequency or require long maintenance windows have a major impact on track availability and network capacity, it is always advisable to integrate automatic scheduling functions into maintenance strategies. 138 An example is reported in Allah Bukhsh et al. 139 where a tree-based model is used to predict and organise the maintenance interventions necessary for the safety of railway switches. Further evidence of how predictive tools can process data to guide infrastructure maintenance is reported in Zhang et al., 140 where the detection of absent sleeper support in ballasted track was investigated using integrated model-based and Data-Driven methods. It is found that the absence of sleeper support has a significant impact on the dynamic response of the coupled vehicle-track system, making its monitoring critical for railway safety. The vertical displacement of the sleepers and the rail-sleeper force were analysed when a waggon passed through the site with 1–6 hanging sleepers, showing a significant increase in the absence of support. A Three-Layer Convolutional Neural Network (TLCNN) was then trained to classify new experimental data into the six sleeper failure conditions. The model has demonstrated a high degree of accuracy in proposing developments that follow the evolution of the conditions of the infrastructure being monitored. In particular, by comparing the measured values with the limit values set by the standard, the Data-Driven will be able to plan the necessary maintenance work.
In conclusion, PdM, through Data-Driven models, provides innovative and automatic support for decisions that ensure the continuity of rail services in the long term, based on the detection and prediction of rail infrastructure failures. This explains the growing scientific and economic interest in moving from corrective and planned maintenance to the modern predictive approach. However, PdM cannot be based on data and AI models alone, as there is critical information such as cost considerations, machinery and personnel involved, scheduling aspects of activities characterised by different times, places and operators; therefore, to understand and manage the system as a whole, this information can only be best interpreted and processed through close collaboration between digital models and human experts in innovative railway maintenance and AI techniques.
Innovative applications of AI models in railway infrastructure maintenance
In the second part of Section 3, the importance and usefulness of innovative methods based on AI for railway infrastructure monitoring are confirmed through some significant examples selected from the literature for their valuable contribution in terms of the completeness of the research carried out, the quality of the results obtained and the fundamental ability to report them clearly.
Investigation of railway fastener using image processing and augmented deep learning
Rail anchors, tie plates, chair and rail fasteners constitute the rail fastening system, an indispensable part of the tracks that must be periodically inspected to ensure safety, efficiency and sustainability of the rail network. 8 Despite the difficulties of positioning and the practical limits imposed by environmental conditions, automated visual inspection of railway infrastructure has found a new and promising interest in recent years thanks to the combination of image processing and Deep Learning algorithms. 141 Research conducted in Chandran et al. 8 uses this approach to detect missing clamps within a railway fastening system using images acquired during field inspections along the Borlänge-Avesta line in Sweden. Missing clamps in the rail fastening system are responsible for a dangerous reduction in the force holding the rail to the sleeper, which can lead to slippage, excessive gauge widening and low lateral resistance, which can further lead to derailment.
Railway images were collected using a grayscale CMOS line camera mounted on an inspection car. For good detection accuracy it is necessary to reduce the positioning error of the fasteners, for this purpose in Chandran et al. 8 the raw images were merged to form a long-concatenated image of the railway track therefore, after using appropriate image processing techniques and anti-noise filters, single frames containing a sleeper with two fasteners are obtained. The images are then categorised into three different classes based on the structural health of the infrastructure, that is, healthy fastening system with both clamps intact, fastening system with one clamp missing and fastening system with both clamps missing.
The fastener detection task was studied using two DL algorithms, namely Convolutional Neural Network (CNN) and Residual Network (ResNet-50). 142 CNN algorithms convolve input images with filters or kernels to extract features, which can reduce trainable parameters compared to traditional Artificial Neural Networks, taking advantage of their notable features such as subsampling, weight sharing and local field. ResNet-50 is a very deep network that stacks building blocks of the same connecting shape called residual units, then it increases the performance of the network using a shortcut or skip connection. 143
The original dataset in Chandran et al. 8 contained over 6000 cases of healthy clamps, 116 cases of fasteners with one clamp missing, and only 47 cases of fasteners with both clamps missing, which meets the expectations of an operational rail section. However, an unbalanced dataset, no matter how large, will not allow you to accurately solve a multiclass classification problem. Image augmentation techniques were used to update the dataset making it fit for purpose. The final dataset contains 3000 real or augmented images, 1000 images for each class which are used as 2040 samples for training, 510 samples for validation, and 450 images for testing.
To evaluate the CNN and ResNet-50 models, performance indicators such as accuracy and cross entropy (loss) were investigated during the training and validation phases and indicators such as precision and recall during the testing phase. Training and validation accuracy for both the CNN and ResNet-50 models was greater than 98% with minimal loss. While ResNet-50 has the highest accuracy, the CNN has shorter training and testing times, due to the larger network structure of ResNet-50. Ultimately, both algorithms were able to achieve over 94% accuracy in detecting fasteners in a variety of environments during the testing phase. It is interesting to note that a standard laptop with Matlab (R2019b), Python 3.6, with packages such as Numpy, Pandas and Keras and Jupyter Notebook was sufficient to manage the software part of this research. 8
Thermal imaging analysis for railway infrastructure monitoring
To ensure efficient rail transport in difficult weather conditions, such as low temperatures and risk of ice, electric heating devices (EOR) have been introduced, which allow automatic switching on based on temperature variations, in particular EORs are crucial elements of railway turnouts. In Stypułkowski et al. 52 an innovative method based on thermal imaging of infrastructure elements and ML model is used to monitor the operation of trackside EOR devices with the aim of increasing the safety level of the infrastructure by making the process faster diagnostics and capable of preventing emergencies.
A critical factor for those involved in monitoring and managing the railway infrastructure is the availability of reliable, robust, and easily installable tools. For this purpose, data analysis based on ML methods and thermography, as a non-destructive tool, represents an interesting choice for monitoring railway turnouts through the conditions of EOR devices. The most critical area is the gap between the resistor itself and the turnout switch rail, which requires heating to remove snow and ice. This is done using resistance heaters connected to the resistor foot.52,144 EOR devices placed in the track line of Warszawa Zachodnia station under normal atmospheric conditions were studied in Stypułkowski et al. 52
EOR devices have guidelines for standard operation and maintenance and require continuous monitoring, as summarised in detail in Stypułkowski et al., 52 where a thermovision method has been adopted to track the processes associated with the differentiation of thermal images of individual EOR objects. 145 For the tests, they used a camera with a non-cooled SC 660 detector mounted on a tripod positioned in the axis of the track. The thermograms, visible and infrared images, were recorded over a temperature range of −10°C to 120°C. Setting the colour palette, temperature range and isotherm is the basis of classic manual inspection, but when combined with thermographic imaging it becomes fundamental to an accurate automated approach that enables real-time image analysis during EOR operations to identify the occurrence of any anomalies. In Stypułkowski et al., 52 a Python software was developed to analyse and predict anomalies and failures of railway EOR equipment using thermal imaging and ML for image analysis. Once trained, the AI automatically interprets the results of thermal imaging analysis to effectively assess the condition of components and then suggests the correct maintenance strategy, from quick inspection to repair or replacement, increasing the safety and sustainability of infrastructure.
Railway track faults detection and localization using acoustic analysis
The great breakthrough in digital technologies, and particularly those related to AI, has allowed146,147 to develop an autonomous approach based on the Internet of Things (IoT) for the identification, the localisation and classification of railway track faults based on acoustic analysis. In fact, they identified six different track faults linked to a specific frequency: wheel burnt, loose nuts and bolts, crash sleeper, creep, low joint, and point and crossing. There are numerous research studies based on electromagnetic approaches, for example a differential eddy current sensor combined with a ML tool was used to detect missing clamps in the rail fastening system, 148 a magnetic flux leakage allows minor imperfections on the surface of the rail to be identified. 149 Guided wave systems, such as non-destructive ultrasonic approaches, are employed to measure the angular velocity, elastic deformation, and rail angular displacement 150 and to identify the onset of a fracture on the railhead surface. 151 Systems based on computer vision constitute an innovative method for monitoring railway infrastructures, where the use of drones has recently been tested. 152 As reported in Siddiqui et al., 146 given the extraordinary image processing capabilities of DL models, especially those based on Neural Networks, there are more and more vision-based approaches for the identification, localisation and real-time classification of defects, such as missing or damaged components and surface degradation.
Modern PdM techniques are based on accurate data collection and management, for this purpose the development of high-performance real-time communication systems, such as IoT-based systems, becomes essential. An example is reported in Nayan et al., 153 where an IoT-based real-time railway fishplate system monitors the position of each bolt on each fishplate and the central system if any bolt is loose. In Rifat et al., 154 a prototype solar-powered vehicle uses ultrasonic and infrared sensors to detect cracks and obstacles in railway tracks; if an anomaly is detected, the vehicle stops automatically and sends a warning message to the central control via the Global System for Mobile communications (GSM) module which at the same time allows the defect to be localised. In Hashmi et al., 147 a monitoring system based on acoustic methods combined with DL models, that is, CNN and RNN, has been developed to improve railway performance and reduce railway accidents. Considering these approaches, an IoT system for railway fault detection based on acoustic analysis was presented in Siddiqui et al. 146 Two microphones record the acoustic signal caused by the friction of the wheels and the track while a GPS sensor records the position of the cart which moves at an average speed of 35 km/h therefore, the collected data is sent every 5 s to the central computer via a WiFi network. For each type of fault, in Siddiqui et al., 146 are collected between 200 and 300 audio files from which to extract the acoustic frequency features used to train and test the DL algorithms, with a 7:3 ratio, to correctly classify the six track faults. Wheel burns are caused by the drive wheel of a locomotive slipping on the rail fastenings, when the tractive effort of the locomotive is insufficient to support the weight of the train, wheel slip occurs, causing the temperature of the rails to rise and the rail surface to melt. 155 Rail creep is defined as a longitudinal movement of the rail relative to a sleeper. The sleeper provides the permanent route with longitudinal and lateral support by ensuring the rails are properly sized and aligned, 146 therefore crashed sleepers are dangerous to rail service. Likewise, you need to check for missing or loosening of the nuts that held the tracks to the rail seat. A rail joint in good condition reduces the effects of wheel passage through steel rail connection areas, while improving the stability and continuity of passing trains, 156 for this reason it is necessary to identify low joint faults. Points and crossings are critical components used to move rail vehicles from one track to another, representing a weak point that must be monitored and maintained in good operating condition.
The time domain and spectrogram images of the acoustic signals of these faults are visually different 146 and this is a key feature to implement an accurate classification tool. In Siddiqui et al. 146 several approaches to the task of classifications have been performed, from classical ML models such as adabost (ADB), Random Forest (RF) and Logistic Regression, to advanced DL models such as Artificial Neural Networks (ANN) and multilayer perceptron (MPL). Such research demonstrated that DL approaches, especially the MLP model, outperformed classical ML models, having higher classifier performance indices, namely accuracy, precision, and recall. Furthermore, given the general nature of the developed models, they are suitable to be integrated into future IoT systems to monitor other parts of the railway infrastructure.
Railway infrastructure maintenance using deep reinforcement learning integrated with Digital Twin
The best monitoring performance to date has been achieved by PdM using advanced DL classifiers based on supervised learning. Recently, a new approach using the integration of Reinforcement Learning (RL) and Digital Twins has been proposed in Sresakoolchai and Kaewunruen 157 to improve the maintenance efficiency of railway infrastructure based on track geometry parameters and track component defects. In RL models, agents are trained and learn from environments that have rules that the agents must follow, such as constraints and available actions. The agents will interact with the environments by performing actions, then the agents will receive rewards or penalties, this process will be repeated until the end of the training with the aim of maximising the rewards or minimising the penalties and in the railway field maintenance costs and defects can be considered as penalties that the agent must minimise. 158 The technique used to develop the RL model is the Advantage Actor Critic (A2C) which has better performance and shorter process times compared to classical techniques. 159 Furthermore, in Sresakoolchai and Kaewunruen, 157 the advantages of an approach able to integrate ML techniques and digital twins are investigated, such as the possibility of improving the efficiency of data management, since all data can be stored in a single model and used to make decisions during the different phases of the project.
In Sresakoolchai and Kaewunruen, 157 data is collected from a 30 km railway section over the period 2016–2019 using track geometry cars. Seven track geometry parameters, such as superelevation, longitudinal level, alignment, gauge, and twist, are used as input so the agent can select subsequent actions. Each of the track geometry parameters is associated with a safety threshold which can be interpreted as a priority level from 1 to 4. Priority 1 means that the track geometry parameters are very bad, and the track sections need to be maintained as soon as possible, while priority 4 means that the track sections need to be included in the regular maintenance plan. Inspection reports represent a second source of data containing information on 71 different types of track component defects grouped for simplicity into five categories based on track components: ballast, fasteners, rail, sleepers, switches, and crossings. Therefore, considering the track geometry and component defects, we get a total of 12 categories used as states in the RL model to train the agent to perform maintenance tasks.
The study in Sresakoolchai and Kaewunruen 157 finds a strong correlation between track geometry defects, in fact a section of track that has some track geometry defects, other than gauge, tends to have other defects, on the other hand the correlation between track component defects is lower as these defects are more independent. Furthermore, crossing and fastener defects have a high correlation with track geometry defects. Maintenance records represent the third source of information which contain seven maintenance activities consisting of tamping, rail grinding, ballast cleaning, sleeper replacement, rail replacement, fastening component replacement and ballast unloading, these activities describe the possible action space for the RL model, which correspond, in practice, into 128 maintenance actions. To update the states of the RL model, changes in each state are considered using field data from track geometry measurements, defect inspection reports, and maintenance logs. In summary, there are 12 states in the environment corresponding to 7 track geometry parameters and 5 track component defects occurrences, the agent takes a maintenance action, after taking the action, the environment will respond by generating a series of new states considering the proper values of each state based on the field data. This process is repeated until the end of the training. The rewards are defined in 2 categories, rewards in terms of maintenance costs and penalties in case of defects.
At this point, having acquired and prepared the data for RL, a Digital Twin model was developed using Autodesk Civil3D. 157 This model allows you to store, process and share data and information with the RL algorithm to create a sophisticated multi-variable model that can simultaneously consider different aspects such as cost, time, emissions and maintenance scheduling. 160 The RL model developed in Sresakoolchai and Kaewunruen, 157 can increase maintenance efficiency, decreasing the number of maintenance activities by 20% and reducing the number of defects by 68%. On the other hand, supervised and unsupervised learning cannot guarantee this performance because their prediction is made only once, whereas RL uses historical data and actions to continuously update itself as it learns. The combination of RL and Digital Twins is a promising technology for improving the efficiency of railway maintenance, reducing the number of defects, maintenance costs and downtime for railway maintenance, improving safety, passenger comfort and railway sustainability.
Discussion
This section aims to confirm the importance of the review in helping the reader to acquire effective theoretical-practical knowledge relating to the main components of the railway infrastructure, to the innovative tools and methodologies of SHM, as well as understanding the direct implications of transforming the maintenance of railway infrastructure, making it safer, more efficient, reliable, and sustainable. For this reason, the opinions and point of view of the authors will be reported, suggesting current and future research developments and challenges.
Railway infrastructure, as described in Section 2, is a combination of technologies that takes shape in the diversity of its constituent elements. All the components of the railway seat support the entire transport system in a complementary and functional way. Knowing and assessing the condition of each element is an essential safety and maintenance management obligation to restore optimum operating conditions. In view of the continuous improvement that the whole railway system has been undergoing in recent years, the monitoring currently in place is a fundamental element in the inevitable development of ever more advanced and cutting-edge technologies. The installation of electronic sensors, particularly on the railway superstructure, would make it possible to guarantee safe operation by preventing faults and optimising maintenance operations. In particular, diagnostic systems allowing the continuous monitoring of certain track components would make it possible to obtain and process information useful for assessing the state of degradation, thus contributing to the transition from predetermined cyclical maintenance to PdM maintenance. In fact, with this approach, it will be possible to precisely select degradation thresholds, above which a specific intervention will be prepared. 161 This requires close collaboration between those developing the monitoring system, experts in data management and analysis, and infrastructure managers. Through this collaboration, it is possible to calibrate appropriate models that consider the stochastic nature of the risk and the degradation process in the real world, to estimate the time at which a failure is likely to occur, and to adapt maintenance interventions consequently. 162 In this way, it is also possible to avoid disruptions that could delay or even interrupt train services by scheduling maintenance at the most appropriate time, thus reducing costs (Figure 7).
How to realise predictive maintenance.
To support this assumption, it is possible to consider the maintenance operations that characterise bonded insulated joints. Until now, this component of the railway superstructure has been characterised by periodic visual inspections to check its correct operation. Through the installation of special monitoring instruments, in this case positioned in the head of the two adjacent rails, 163 makes it possible to assess the condition of the joints during railway operation. These devices make it possible to assess the deformations of the joint due to thermal variations and the transit of trains, thus making it possible to calibrate predictive models that allow maintenance work to be planned without compromising the regular operation of the railway. 164
The way people move, exchange information and purchase products and services online are constantly evolving, moving from high-capacity telecommunications networks to multimodal integration mobility platforms, and the growth of online commerce increases the number of delivery vehicles which stress the traditional logistics system. For all these reasons, many countries have already started investing in the transformation of the railway infrastructure, the contribution of which is fundamental to supporting the future evolution of their different communities. 165 The investigation of elasticity indicators of a railway track and geometric monitoring parameters of railway structure condition or combinations of numerical indicators obtained by innovative techniques of data processing, promises to revolutionise current monitoring strategies by allowing the variation of irregularity parameters to be predicted over time. 166
As detailed in Section 3, the growth in the number and performance of SHM tools and innovative Data-Driven models, particularly those based on AI, is rapidly making classical approaches obsolete. Although scientific, technological, and economic interest has led numerous public and private research groups to invest and actively dedicate themselves to the transformation, growth, and innovation of railway infrastructure monitoring, 30 there are still numerous challenges and possibilities for development in this field which, unlike the industrial world, is only at the beginning of its digital 4.0 revolution. 167 Below we report some of the critical challenges identified by the authors that the railway maintenance sector will face in the coming years.
The extension in the railway world of the use of advanced AI methods, such DL, multisensory and ensemble methods, reinforcement learning approaches, which have already demonstrated their qualities in other industrial sectors, bringing advantages in terms of safety, reliability, and sustainability.
A challenge related to modern AI techniques used for PdM and decision making is to increase adaptability and scalability to be able to use these techniques in a more natural and immediate way in the SHM sector despite the complexity and differences of the parts investigated, as already happens in the industrial world.
Data-Driven AI models require high-quality labelled data to achieve high performance. Furthermore, due to the large amount of data required, classic manual labelling methods are inefficient, so it is necessary to develop unsupervised learning models capable of identifying appropriate clusters that allow the labelling problem to be addressed automatically.
Since in the railway sector a large amount of documentation is produced in the form of textual reports, the development of textual interpretation techniques capable of automatically extracting, classifying, grouping, and managing information, becomes a valuable source of data for Data-Driven models employed in modern PdM strategies.
AI models not only need large amounts of data, but for good learning they require significative data, that is, balanced data to cover all classes of interest. This aspect can constitute an obstacle in the railway sector where the measurements, particularly those relating to fault models, are highly unbalanced. To make up for these shortcomings, Digital Twins capable of simulating the behaviour of the infrastructure or some of its parts in a digital environment are gaining interest, also allowing you to obtain large quantities of data for the simulated components in a short time and in generally safer and more sustainable conditions.
A further factor of imbalance in terms of costs and safety is linked to the different consequences caused by a model error. The accuracy of most current models for detecting and predicting track defects is measured by the false alarm prediction, which occurs when the actual safe condition is incorrectly predicted causing unnecessary intervention and with false safe prediction which occurs when the actual fault is mistakenly believed to be safe, creating a potentially dangerous situation which can result in large economic losses and more. It is therefore clear that from the point of view of railway managers it is a priority to reduce the percentage of false safe predictions compared to false alarm ones to reduce costs and improve the safety of the infrastructure.
A limitation of current advanced AI techniques, particularly those related to Deep Neural Networks, is poor interpretability, indeed, these tools process the data automatically providing solutions that allow the implementation of correct and effective maintenance strategies, typically without clearly explaining the reasons for their choices and evaluations made. Improving interpretability will allow experts of railway monitoring, to better understand the solutions proposed by automatic models and possibly to propose modifications, integrate or directly update the model following the evolutions of the system.
The development of modern Wireless Sensor Networks, essential to support and implement innovative SHM strategies, requires accurate, robust, and efficient sensors potentially capable of integrating energy harvesting functions. A technology suitable for these needs is represented by MEMS and NEMS devices.168,169
Virtual Reality, that is, computer technology to simulate a 3D environment, and Augmented Reality which allows you to interact with virtual information relating to objects in the real world represent innovative tools which combined with IoT based models and UAVs will revolutionise the SHM enabling simpler, continuous, real-time, and efficient inspection of remote infrastructure.
Conclusion
This review has comprehensively addressed the critical issue of railway infrastructure monitoring, emphasising the importance of the development and progressive deployment of modern Artificial Intelligence techniques in the railway field and highlighting their significant scientific, technical and economic benefits. From this analysis, it should be clear that investing in Predictive Maintenance helps to create a safer, more efficient and reliable network, capable of preventing critical failures and reducing minor ones, as well as reducing downtime for maintenance or repairs, thus minimising costs.
In order to implement an innovative monitoring strategy, it is necessary to clarify the main characteristics and technical performance parameters that indicate the health conditions of the railway infrastructure in real time. Traditional monitoring systems only partially meet the high demand for information on which modern PdM tools are based. This is why, over the years, Structural Health Monitoring has seen a continuous development of new sensors, measuring instruments capable of collecting and transmitting large amounts of data with accuracy, availability and sustainability, characteristics that are highly sought after in the railway sector, which sees a fundamental opportunity for evolution and improvement in the ongoing digital transformation. The use of new tools based on the Internet of Things (IoT) and Wireless Sensor Networks (WSNs) for condition monitoring makes it possible to assess the actual status of railway components and, based on this information, use Data-Driven models to effectively plan subsequent maintenance work.
The opening of the railway sector to the Big Data Era confirms the need to strengthen the collaboration between developers of monitoring systems, experts in data management and analysis, and infrastructure managers, in order to favour the realisation of a modern network equipped with accurate and sustainable monitoring systems, from various points of view, that is, implementation, environmental and economic, also with regard to the provision of resources for the maintenance of works and the improvement of safety. All these reasons confirm that now is the time to invest resources and knowledge in the innovation of the SHM of new and old infrastructure, a transformation favoured by the recent development of new algorithms and techniques based on Artificial Intelligence. Such tools have already started to improve this area and will quickly establish themselves as a new paradigm, offering greater reliability, efficiency and simplicity compared to traditional instruments, by automating the management and processing of the vast amount of data required to implement a robust and modern PdM strategy.
Modern tools include Digital Twins, which can simulate the behaviour of the infrastructure or parts of it in a digital environment, providing large amounts of data for the simulated components quickly and under generally safer and more sustainable conditions, anticipating failures and possible system evolutions. Strong correlation between physical and virtual systems underpins the success of future innovative monitoring strategies. On the one hand, modern infrastructures adapt their behaviour based on feedback generated in real time by DT and AI technologies, while on the other, IoT sensors connected by increasingly dense networks collect data from the real world and pass it on to digital systems for processing, detecting any anomalies and making decisions on how to behave.
In conclusion, we expect that innovation in SHM will lead to an increase in the resilience and efficiency of mobility systems, with solutions and services for public and private transport, creating personalised and accessible mobility, enabling new social inclusion and strengthening the supply chain, national competitiveness and international visibility.
Footnotes
Handling Editor: Sharmili Pandian
Declaration of conflicting interests
The author(s) declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.
Funding
The author(s) disclosed receipt of the following financial support for the research, authorship, and/or publication of this article: Costs of this research have been partially covered by the National Recovery and Resilience Plan (NRRP), Mission 4 Component 2 Investment 1.4—Call for tender No. 3138 of 16/12/2021 of Italian Ministry of University and Research funded by the European Union—NextGenerationEU. Award Number: Project code CN00000023, Concession Decree No. 1033 of 17/06/2022 adopted by the Italian Ministry of University and Research, CUP D93C22000400001, ‘Sustainable Mobility Center’ (CNMS). Spoke 4—Rail Transportation.
