Data-Driven Condition Monitoring of Fixed-Turnout Frogs Using Standard Track Recording Car Measurements

Abstract

Turnouts are critical components of railway infrastructure, ensuring operational flexibility but also representing a significant share of track maintenance costs. The frog, as the most vulnerable part of a turnout, is subject to severe wear and degradation, requiring frequent inspection and maintenance. Traditional manual inspection methods are costly, labour-intensive, and susceptible to subjectivity. This study explores a data-driven approach to condition monitoring of fixed-turnout frogs using standard track recording car measurements. By leveraging over 20 years of longitudinal level and rail surface signal data from the Austrian track-recording measurement car, we assess the feasibility of using existing measurement data for predictive maintenance. Six complementary approaches are proposed to evaluate frog condition, including track geometry assessment, ballast condition analysis, rail surface irregularity detection, and axle box acceleration-based monitoring. Results indicate that data-driven monitoring enhances maintenance decision-making by identifying deterioration trends, reducing reliance on manual inspections, and enabling predictive interventions. The integration of standardised measurement data with advanced analytical models offers a cost-effective and scalable solution for turnout maintenance.

1. Introduction

Turnouts are critical elements of railway infrastructure, providing network flexibility and operational safety. However, their structural complexity and exposure to concentrated dynamic loads make them particularly vulnerable to accelerated deterioration [1,2,3]. As a result, turnouts account for a disproportionate share of maintenance costs—up to a third of total annual track maintenance expenditure on some networks [1,4]. By continuously monitoring critical parameters such as material degradation, condition-based maintenance offers a data-driven alternative to traditional inspection methods. This approach not only extends asset life but also supports proactive intervention before major failures occur.

Structurally, a turnout is typically divided into three sections: the switch panel, the closure panel, and the crossing panel [5,6,7]. Turnout inspections focus on two key areas: the switch panel, which includes the movable switch rail and locking mechanism, and the crossing panel—and, in particular, the common crossing or “frog”—which is subject to the highest mechanical loads [8]. Due to a discontinuity in the running surface, this area experiences repeated wheel impacts as vehicles transition over the crossing nose [9,10], resulting in strong vertical forces and high-impact energies [7]. These stresses lead to localised wear phenomena such as plastic deformation and rolling contact fatigue (RCF) [10]. Consequently, fixed frogs represent one of the most failure-prone and maintenance-intensive components within the railway system. Deterioration in fixed frog areas is often due to localised anomalies such as voided sleepers, ballast contamination or inadequate substructure support [11]. These irregularities increase wheel–rail interaction forces, creating a vicious cycle that accelerates component failure. Early detection of such defects is therefore essential to break this cycle and optimise maintenance timing.

Conventional maintenance of turnouts, including fixed frogs, relies predominantly on scheduled inspections supported by manual assessments [11]. At present, turnout maintenance follows a fixed schedule based primarily on visual inspections. This method is not only resource-intensive and costly but also poses significant safety risks as it requires personnel to work directly in the danger zone of the track. In Austria, for example, these inspections are carried out in track closures, temporarily interrupting rail operations and thereby reducing network capacity and availability [12].

While periodic inspections remain essential to ensure operational safety [13], they lack reproducibility and provide limited predictive insight into occurring failures. To address these limitations, the integration of condition monitoring [14] and predictive maintenance strategies has emerged as a critical objective for infrastructure managers. Data-driven maintenance approaches offer a path to increased reliability, cost efficiency, and asset longevity [15,16]. Recent advances in this field have also explored hybrid data-driven and physics-informed methods. For example, Men et al. proposed an adaptive style-transfer framework for small-sample-bearing fault diagnosis, demonstrating how digital-twin-based and transfer-learning approaches can enhance robustness under limited data conditions [17]. Real-time monitoring of degradation allows operators to shift from reactive maintenance to proactive asset management, targeting maintenance activities based on actual needs rather than fixed schedules.

Condition monitoring can be implemented using two complementary strategies: (1) the use of specialised measurement systems adapted to the asset, or (2) the reuse of existing data sources not originally designed for turnout analysis. In the first case, specifically designed sensor systems and manual measurement devices have been developed for turnouts, including portable instruments capable of capturing detailed switch and crossing geometry. However, these sensor-based approaches currently cover only a limited number of turnouts across the network, and the resulting data cannot yet be generalised for network-wide evaluation. Similarly, manual measurement devices carry the same operational risks as traditional inspections, requiring track closures and exposing personnel to safety risks. In contrast, the second approach using standard track-recording cars is already well-established. No additional investment is required and there is extensive historical data, enabling time series analyses and the development of evaluation methods for assessing turnout condition on the observed track sections. The first approach provides high accuracy but is associated with significant costs and limited scalability. The second, while more analytically demanding, is cost-effective and scalable across the network. Standard track-recording cars offer a foundation for implementing such strategies. These vehicles routinely capture key geometric parameters—including longitudinal and lateral level, gauge, twist and rail surface irregularities, and axle box accelerations—across entire networks. Of these, longitudinal level is widely considered to be the primary indicator of track condition and degradation progression [13,18,19,20,21], but the potential of any available measurement signal needs to be evaluated.

This study pursues the second approach. Using 20 years of data from the Austrian standard track-recording car, we investigate the feasibility of condition monitoring for fixed frogs. The aim is to identify patterns of deterioration, assess the effectiveness of past maintenance interventions, and explore the potential for predictive modelling based on existing data, allowing for low-cost retrospective analysis and model development without the need for additional instrumentation. By focusing on this critical but under-monitored asset class, we contribute to the development of scalable, data-driven maintenance strategies for railway turnouts.

2. Methodology

The Austrian track-recording car monitors the main railway network several times a year and describes multiple aspects of the track condition. The data sources used for this study are the longitudinal level, the rail surface signal, and axle box accelerations measurements. Details of these signals are given in the relevant subchapters of the paper. Whilst the track-recording car is designed for track, the data also covers turnouts. Due to the limited spatial extent of turnouts and even more so of the frog, a data positioning process is essential to make the data from the track-recording car usable for the purpose of this study. Data positioning is described in detail in [22,23] and includes relative positioning (alignment of measurement runs) and absolute positioning (alignment of data with actual track positions). For a recent study, we utilised the fine-positioned data for condition monitoring of entire turnouts. Details can be found in [24]. Approximately 200 turnouts were analysed in the manner described in the publication. In this study, we use the same databases to monitor the condition of turnout frogs.

The condition of frogs cannot be adequately described by a single quality index, as different aspects are relevant to ensure a safe and cost-effective frog performance. Poor-quality behaviour has its roots in different parts of the system, either related to the metal part or to the bedding of the frog. For poor quality of the metal part, it is to a certain extent possible to restore the intended frog geometry by maintenance operations such as grinding and repair welding, but at some point, replacement of the frog is advisable. If the bedding of the frog is poor, tamping might solve the problem. However, if the ballast is already worn, tamping is not sustainable, and a ballast replacement is the only sustainable option. Within this research, six different approaches are proposed which can provide a good overview of the condition of the frog. Figure 1 depicts an overview of these approaches.

Approach 1 deals with cracks, breakouts, and rolling contact fatigue, as these effects are present relatively often in the area around frogs. High-definition images in combination with image recognition algorithms are a possibility to detect effects of this kind. Refs. [25,26] provide examples on the possibilities of the technology. However, image recognition alone is not expected to be sufficient, as not all cracks are visible from above, so ultrasonic testing will still be required in the future. While all the other proposed quality indicators will be evaluated in detail, image recognition is not part of this study. The six approaches proposed are interrelated and need to be analysed together. However, each approach requires different data sources and data processing. Therefore, each approach is described in detail and finally summarised in a common datasheet. Since track geometry is the most widely used indicator in track quality assessment, we selected it as the first approach.

2.1. Track Geometry Assessment

Due to the dynamic impact loads, the frog area is susceptible to track geometry failures [11]. Track geometry in vertical direction is best described by the longitudinal level in the D1 range as defined by EN 13848-1 [27]. Most infrastructure managers worldwide measure this parameter regularly using dedicated measurement cars either based on an inertial unit or a chord-based system in order to guarantee the safe geometrical condition of the track. These measurements are also available in turnouts. Following appropriate data alignment, they can be used to monitor isolated track geometry faults at the frog. Figure 2 illustrates the raw signal of longitudinal level and its deterioration over time for one of the analysed turnouts.

The upper half of the figure shows the longitudinal level expressed spatially. The frog area is highlighted by a grey rectangle. An isolated failure of the track geometry (local settlements) is clearly visible in this area. The lower half shows the deterioration of the longitudinal level of the highlighted area. To perform time series analysis, we utilised maximal zero deviations of the longitudinal level as the quality indicator, as defined by regulatory guidelines. Local settlements are increasing over time, as the indicator demonstrates. Historical maintenance actions are illustrated as vertical dashed lines. Tamping in 2011 had a positive effect on the longitudinal level of the entire turnout, but as there were no severe isolated failures in the frog area, tamping had no effect on the frog. By the end of 2017, no significant reduction in failures had been achieved, but the system had stabilised to some extent, with a lower rate of deterioration after tamping than before. In 2021 and 2022, only local measurements were carried out on the frog instead of tamping the whole turnout. No data is available on the type of maintenance carried out, but this example is consistent with experience and shows that local measures are rarely sustainable.

The time series displays a recurring pattern of tamping actions in turnouts: tamping typically enhances the track geometry of a turnout, though it is not always effective in addressing isolated track geometry failures beneath the frog. There may be two reasons for this. Either the tamping machine used is not capable of sufficiently compacting the ballast under the frog due to geometric constraints, or the ballast under the frog is already severely degraded to the point that the ballast bed cannot support the tamped geometry. A wavelength-based analysis of the longitudinal level can give an indication of the latter possibility and will be addressed in the next approach.

2.2. Ballast Condition Assessment

In his thesis, Fellinger describes how the analysis of the wavelength spectrum at the longitudinal level provides an indication of the ballast condition of turnouts [5]. Figure 3 shows the principles of the method.

The input to the calculation is longitudinal level (LL) measurements over the length of the entire turnout. The frequency content of the signal is analysed for each measurement run using the power spectral density. The power spectral density is normally used to describe the power of frequencies within a signal. Since wavelengths are commonly used in the railway industry to classify track geometry effects, the power spectral density is plotted against wavelengths. According to Fellinger, the wavelength range between 3 m and 7.6 m corresponds to ballast problems. Finally, the BCI is described as the slope of the linear regression of the power spectral density curve between the 3 m and 7.6 m wavelengths, scaled by the length of the input signal. By repeating the process for each measurement run, the BCI quality can be described over time. Validation and details of the indicator can be found in [5]. As location information is lost in the process, the level of ballast quality is averaged over the whole turnout and no indication of specific points within the turnout is possible. The index calculation must therefore be modified in order to accurately describe ballast under the frog. Figure 4 illustrates the modified calculation developed to indicate the ballast quality under the frog.

Instead of calculating the BCI over the entire turnout, it is calculated for window lengths of 13 m in a sliding fashion (every 25 cm). Window lengths must be as short as possible to represent spot effects, but cannot be reduced indefinitely, as the window must be longer than the longest effect length for mathematical reasons described in [24]. The maximum included wavelength within the 3–25 m signal is 25 m. However, given the fact that the wavelength spectrum between 3 and 7.6 metres is related to ballast quality, shorter window sizes are possible. Following a period of experimentation, we are currently utilising 13 m, and the results are proving to be very encouraging. By scaling the input signal with a triangle function, spot effects in the centre of the calculation window are emphasised, while more distant signal components are attenuated. The result is a spatial representation of the BCI. If suboptimal track geometry near the frog is due to a deteriorated ballast bed, the sliding BCI will give high values. High values indicate that the ballast that is already strongly worn, leaving only limited options for remediation. Reliable indicators should not only quantify the degree of wear but also highlight the main causes of ballast deterioration, which are predominantly dynamic impact loads resulting from track irregularities. One significant source of such irregularities is welded joints located near the frog. These aspects will be addressed in the next chapter.

2.3. Consideration of Weld Surface Irregularities

Frogs are commonly welded into the track and are therefore adjusted to welded joints. Irregularities in the rail surface resulting from these welds can give rise to significant dynamic impacts and therefore necessitate monitoring. The effect of weld battering causing growing rail surface irregularities at welds is described in detail in [28]. As shown by previous research, these irregularities can be monitored using the rail surface measurement system. The system uses three vertical distance lasers to build a measuring chord, thereby recording the rail head’s longitudinal profile. An actuator shifts the device laterally to ensure it aligns with the rail head’s centre position. In Austria, this measurement system is mounted onto the measurement cars and therefore provides regular measurement of the longitudinal profile of the rail head in the wavelength span between 20 mm and 1000 mm. Figure 5 pictures the welds of a frog and depicts relevant rail surface signal characteristics.

While the signal output is unstable due to the crossing nose gap, the evaluation of several turnouts shows that the signal is stable at the position of the welded joints close to the frog. Ramp angles are the best way to quantify rail surface irregularity, as they include the amplitude of an irregularity and the longitudinal expansion of the irregularity in the calculation. Ramp angles represent the sum of the linearised slopes of a V-shaped irregularity. The superiority of using ramp angles over using maximum amplitudes alone is that dynamic vehicle responses depend on a combination of amplitude and wavelength and ramp angles therefore they better represent the vehicle response induced by the irregularity [29]. By the calculation of ramp angles for every measurement run, the evolution of the ramp angle over time can be depicted (Figure 6).

For the first turnout (upper graph in Figure 6), there is a clear linear deterioration trend for both welds from 2012 to 2020. The rate of deterioration is the same for both welds; however, the weld on the side of the wing rail started with a larger ramp angle in 2012. In 2020, the frog and therefore the welds were replaced, resulting in a significant reduction in the ramp angles. After the replacement, the weld on the side of the frog started to deteriorate again. To date, no deterioration has been observed in the wing rail side weld. The second example (lower graph in Figure 6) again shows linear growth of the ramp angles of both welds, but with different growth rates. In 2020, a repair weld was performed which resulted in a reduction in angle for the wing side weld but not for the frog side weld. The worker performing the repair weld apparently only treated one of the welds during the grinding that is part of the repair weld process. In 2023, the frog was replaced, resulting in smooth welds on both sides.

The use of the rail surface makes it possible to assess the welds around the frog. Welds in poor condition are dynamic in themselves, but they can also interfere with a smooth wheel transition. Keeping these welds in good condition can therefore contribute to better frog performance. However, compared to the impact caused by the frog itself, the impact caused by these welds is significantly lower. Therefore, methods need to be established that include frog imperfections. Measurements of axle box acceleration can contribute to this assessment.

2.4. Axle Box Accelerations—Direct Assessment

Frogs in poor condition lead to higher vehicle reactions and consequently higher system forces. As axle box accelerations include information on unsprung (unfiltered) wheel vibrations, which are highly dependent on infrastructure condition, they have the potential to provide insights into the condition of passed frogs. There are two further issues to consider when using axle box acceleration data: (1) The measured data contain not only information about the track, but also about the condition of the vehicle, in particular, the geometry of the wheel. Simulations performed by [30,31] show that wheel geometry is one of the dominant predictors of accelerations and forces. As the data used here is from the accelerator mounted on the track-recording car, it can be assumed that the wheels are in good condition, as there are stricter limits and regular inspections for the recording car. (2) Speed has a significant effect on vehicle response. To use axle box accelerations as an indicator of frog condition, it is necessary to eliminate the influence of speed. For this purpose, the method developed by Joanneum Research—based on comparing accelerations from multiple turnout passes at different speeds [30]—proved most effective. Their study employed a custom-built sensor setup, and the resulting formula was adapted in this work to process acceleration data from Austria’s standard recording car. A self-made sensor concept was used. The formula obtained was adapted by the authors of this paper for the acceleration data of the standard recording car in Austria.

$$a_{norm}=a\ast {\left({\displaystyle \frac{1}{v}}\right)}^{1.6}\to a_{norm}=a\ast {\left({\displaystyle \frac{100}{V}}\right)}^{1.6}$$

(1)

The formula from Joanneum Research is adjusted in two ways for this study. The unit of speed is changed from m/s to km/h, as the latter is more commonly used in the railway sector. In addition, a reference speed of 100 km/h is chosen for values in a realistic range. By applying this formula, the influence of speed can be reduced to an acceptable level. Using acceleration data from the standard recording car has the advantage of covering the entire network over approximately 20 years. However, a major limitation arises specifically for the Austrian recording car and does not necessarily apply to recording cars in other countries. As it is processed within the recording car into an indicator used to detect corrugation and then stored, raw data is not available. Figure 7 provides an example of the stored accelerations for one turnout in the Austrian network.

The two newest measurement runs are represented in order to demonstrate the stability of the signal characteristic. However, it is clear that high frequency parts of the signal are lost due to the sliding average calculations performed on the recording car. Nevertheless, clear signal characteristics are visible at relevant points. In the transition zone at the front of the turnout, an insulated rail joint and a change in track stiffness led to relatively high accelerations. The effect of welds is also visible in the data. The highest accelerations are caused by the frog. As the signal has already been processed and does not allow for more sophisticated index calculation methods, root mean square (RMS) and maximum amplitude were analysed as potential quality indicators using different window lengths. The most stable results are obtained using RMS values with a window length of 5 m. In order to ascertain the extent to which the calculated indicator demonstrates good or poor quality, we conducted statistical evaluations. For 56 turnouts, information about frog replacements is provided by OeBB Infrastrukur AG. By assuming a poor quality before and a good quality after the exchange of the crossing, quality areas can be defined based on the historical data.

The left box in Figure 8 represents the quality before frog exchanges, which is relatively poor, while the right box represents the quality after frog exchanges, which is relatively good. The respective frogs of the same turnout are connected by a dashed red line. The data shows that almost every replacement resulted in a significant improvement in frog quality, which in turn resulted in lower axle box accelerations. A relatively high deviation can be seen for the quality indicators before the frog replacement, while only a few high values remain after the replacement. There are three possible explanations for this: (1) It is possible that several frogs were replaced due to rail defects and cracks. While these are valid reasons for replacement, they are not necessarily the cause of increased accelerations. (2) The data points represent the last measurement run prior to replacement. It is common practice to carry out repair welding before replacing a frog. This may result in slightly lower accelerations for frogs where repair welding is effective. (3) Replacing the frog without addressing the bedding quality may result in increased accelerations after frog replacement.

Despite the variations observed, it is clear that there is a distinction between the boxes. These results are used to define quality ranges (right side of the plot). Four quality ranges are defined, representing ‘very good’, ‘good’, ‘moderate’, and ‘poor’ quality. Values up to 2.89 are indicative of very good quality, as only frogs that have been replaced have reached this level. The range for good quality (2.89–5.09) is defined by the median of the accelerations before frog replacement. Values up to the third quartile of pre-frog-change accelerations are defined as moderate quality, while higher values are defined as poor quality. It should be noted that poor quality does not necessarily equate to safety issues, but it may be advisable to avoid accelerations of this magnitude in the long term, as high dynamics can lead to faster deterioration of the system. In Figure 9, a time series of the described indicator is depicted. The quality areas defined are included for an easier interpretation.

As axle box acceleration data is available for both rails, the indicator is calculated for the frog side and the guard rail side of the turnout. Darker crosses represent the frog side, which shows higher values than the guard rail side. This is to be expected, as the wheel passing the frog is more directly affected by the frog geometry. While there is a large scatter despite the harmonisation of the measurement speed, there is a trend of increasing accelerations over time from 2012 to 2020. In 2020, the frog was repair welded, which significantly reduced the acceleration (from the red bad area to the green good area). However, due to further wear of the crossing, the accelerations increased again since 2020, indicating the need for further maintenance in the near future. While condition information is contained in the pre-processed acceleration data from the Austrian track-recording car, deep evaluations are limited, and detailed conclusions cannot be drawn. Since some track-recording cars and standard locomotives measure and store raw acceleration data, approaches using these data are meaningful and will be discussed in the next chapter.

2.5. Axle Box Accelerations—Assessment Based on Dynamic Loads

In Approach 5, a quality indicator for the frog is derived directly from axle box acceleration measurements. Approach 6, however, uses axle box accelerations to approximate the vertical wheel trajectory while passing a crossing and incorporates the calculated ramp angle into an analytical model for computing dynamic impact loads. For this purpose, we use axle box acceleration measurements from two regular locomotives equipped with vertical accelerometers. As noted previously, the standard track-recording car used in Austria does not provide raw acceleration measurements. Nevertheless, there is an increasing trend among infrastructure operators to record axle box accelerations using standard measurement vehicles, making raw axle box accelerations representative of typical measurement data. By integrating axle box accelerations and implementing a suitable filter, the vertical movement of the wheel over a given distance can be determined. Filtered accordingly, these movements approximate longitudinal level [32]. Two aspects must be considered. (1) In order to nullify the impact of the measurement speed (in this case the speed of the regular locomotive), double-integrated axle box acceleration has to be dived by the squared measurement speed. (2) As integration results in long-wave drifts of the signal due to measurement noise, a method must be implemented which deals with the noise. This phenomenon is well known and often discussed in the literature [33]. While different approaches for eliminating drifts are suggested in the literature, a simple filter method is used here. After the first integration step, a Butterworth filter is used, allowing only short wavelengths. The second integration uses the filtered signal as input and returns the displacement of the wheel, which approximates longitudinal level. For validation, the double integrated axle box acceleration signal is filtered to the D1 wavelength range (3–25 m) and compared to longitudinal level measurements from the track-recording car (Figure 10).

The black line represents the longitudinal level as measured by the track-recording car using an IMU. The coloured lines represent the double-integrated axle box accelerations, filtered to the wavelength range of the longitudinal level (3–25 m). Four measurements from two vehicles are available for comparison. The comparison shows that three of the acceleration signals fit the longitudinal level quite well after double integration. One measurement, the red signal, shows some unexplained deviations. Further applications of the method show that double integration is a reliable approach to approximating the longitudinal level. However, as demonstrated in Figure 10, there are exceptions that warrant further investigation in future research.

The advantage of using double-integrated axle box acceleration data instead of longitudinal level is that it contains a wide frequency spectrum. Filtering to 3–25 m can be used for validation purposes, but the inclusion of shorter wavelengths can provide additional information on vertical track geometry. This is particularly useful for the frog area, as the inclusion of shorter wavelengths makes it possible to visualise the actual trajectory of the wheels. Figure 11 compares longitudinal level with double-integrated axle box accelerations, including shorter wavelengths.

The red line represents the double-integrated acceleration data, including wavelengths from 1 to 25 m. Due to the inclusion of shorter wavelengths, a deeper peak is depicted in the data. This peak corresponds to the short-wave wheel drop in the frog area and is not detectable in the longitudinal level as it includes longer wavelengths only. The use of double-integrated axle box accelerations therefore adds additional information that can be used for evaluation. However, the results are affected by the wavelength range chosen. This is illustrated in Figure 12.

The figure illustrates the vertical wheel trajectory obtained through double integration of axle box accelerations, with lower wavelength limits varying between 0.01 m and 3 m. As the minimum wavelength decreases, the depth of trajectory peaks increases, although this effect is limited. A focused analysis of three frogs, shown on the right side of Figure 12, reinforces this observation, indicating that wheel drop amplitudes tend to grow with shorter wavelengths. However, this trend levels off within the minimum wavelength range of 0.1 to 0.25 m. Shorter wavelengths than 0.1 m seem not to contribute to the wheel drop and are therefore not relevant for the load modelling. While this is not the case for the data analysed, data instabilities are more likely for signals including shorter wavelength fractions. Given the lack of additional information within wavelengths below 0.1 m and the increased risk of data instability, data is filtered to 0.1–25 m.

The severity of the wheel drop provides an indication of the condition of the frog. Worn frogs result in more severe wheel drops. More severe wheel drops are associated with higher dynamic impact loads that damage components. Monitoring and limitation of wheel drop to a certain level could therefore lead to more sustainable frog performance. By the incorporating track design and properties of passing vehicles, condition monitoring can be further specified for the respective line. For that, we utilise the analytical approach for the calculation of dynamic impact loads based on Jenkins [29]. Impact loads are typically characterised by two force peaks named P₁ and P₂ peak. While P₁ acts directly after the wheel is stimulated by an irregularity (the crossing nose dip) and with very high frequencies around 1000 Hz, P₂ acts several milliseconds later and with frequencies between 20 and 100 Hz. As P₂ transfers more energy into the system and comes with frequencies which lead to an increase ballast degradation and track geometry deterioration, it is the more critical one.

$$P_2=P_0+2\alpha V\ast \sqrt{{\displaystyle \frac{m_u}{m_u+m_t}}}\ast \left(1-{\displaystyle \frac{C_t\ast \pi }{4\ast \sqrt{K_t\ast \left(m_u+m_t\right)}}}\right)\ast \sqrt{K_t\ast m_u}$$

(2)

P₀  Static wheel load

2α Ramp angle of the irregularity [rad]

V   Vehicle speed [km/h]

m_u Unsprung mass per vehicle wheel [kg]

m_t     Effective vertical track mass per vehicle wheel [kg]

C_t  Effective track damping per vehicle wheel [Ns/m]

K_t  Effective track stiffness per vehicle wheel [N/m]

The formula comes with input parameters of track (blue) and vehicles (red) which have to be filled with representative values for the respective track vehicle combination. For track, component-specific input data must be defined. For vehicle parameters, the Austrian standard universal locomotive is used as the reference vehicle. Vehicle speed is defined by the permitted speed of track. 140 km/h was chosen for the frog analysed. The ramp angle 2α is derived from two data sources: (1) double-integrated axle box accelerations measurements, as described above, and (2) geometry information from handheld tools. Data was provided by a project partner and gained by a rail profile measurement executed several times over the longitudinal extension of the frog. The measurement principle is described in detail in [34]. Ossberger also developed a method to simulate the wheel trajectory of an ORE-S1002 wheel passing the measured frog. This method is best described in [34]. Figure 13 provides a 2D representation of three frog geometry measurements available for this study.

The left (falling) lines represent the geometry of the wing rail, the right (rising) lines represent the geometry of the frog. The upper curves refer to the first measurement after the turnout was renewed, so the frog can be assumed to be in good condition. During operation, the geometry deteriorates, resulting in the lower curves. It is clearly visible that the ramp angle has increased due to the deterioration process. The frog was replaced after 2018, resulting in the geometry shown by the middle curves. Again, the geometry is more symmetrical and the transition of the wheel is smoother. For further interpretation, the geometries need to be linearised. This can be achieved by using linear regressions described by the coloured lines in Figure 13.

For one turnout in the Austrian network, both data sources are available, therefore this turnout is analysed in detail. Additionally, one of the regular locomotives which was equipped with axle box acceleration sensors was also equipped with a measurement wheel as described in [35] for the measurement of Q forces. Q forces are important for vehicle homologation and therefore well known in the railway sector. Typically, measurements are limited to a frequency of 20 Hz, which is an important aspect in result interpretation. Figure 14 combines the described methods and depicts calculated and measured forces arising when a locomotive passes the analysed frog.

Local measurements are available between 2012 and 2023; therefore, dynamic forces are presented over time. The blue points represent calculated P₂ forces when a locomotive passes through the geometry provided by local measurements. As clearly visible, due to wear of the frog and the wing rail, the linearised ramp angle grows over time, leading to P₂ forces becoming higher over time. In 2020, a repair welding action repairs the frog geometry and therefore reduces the P₂ force level. In mid-2020, in addition to local measurements, ramp angles gained from double-integrated axle box accelerations from two locomotive types with two measurements were available and depicted as red points. Again, the Jenkins model is used for P₂ force calculation. As the two locomotives come with similar masses, the P₂ forces are comparable. Also, in mid-2020, Q forces provided by a measurement wheel set were available and visualised as black points. As the vehicles passed the frog with a speed of 140 km/h, P₂ force calculation also assumed the same velocity. When comparing the force level of the three groups, significant differences appear. Although they all represent dynamic forces, a direct comparison is not technically possible. The lowest force levels (~200 kN) are represented by the measured Q forces. The reason for this is that the measurement principle (wheel set measurement) is limited to frequencies up to 20 Hz. This results in a force signal that filters out high frequency force peaks and therefore reduces the amplitudes of the force peaks, as the highest amplitudes are typically caused by high frequencies on top of lower frequencies. The P₂ forces of the Jenkins model represent the 20–100 Hz frequency range. Therefore, Q forces from wheel set measurements clearly underestimate the force peaks relevant to ballast damage.

The intermediate force level (~270 kN) is the result of ramp angles derived from manually measured geometry data and incorporated into the Jenkins formula. Although the critical frequency spectrum is represented, the results still underestimate the actual force level as the manually measured geometry data only represents the unloaded frog geometry. The highest and most realistic force level (~450 kN) is achieved by using double-integrated axle box accelerations as input. In this case, the actual wheel trajectory represents the loaded ramp angle of the frog and includes geometric properties of the metal part and bedding effects.

Finally, the developed indicators are applied in two complementary ways. First, a datasheet integrating the time series of all indicators is presented to provide a detailed assessment of the condition of a single turnout. Such a datasheet can serve as a practical tool for decision-making in operational asset management. Second, indicators from multiple turnouts are analysed to determine typical index values and to examine the distribution of the indicators across a random sample of the main network, allowing for the identification of general trends and patterns.

3. Results

The approaches described provide information on the condition of the frog based on different input data and different methodologies. Different aspects of frog condition are represented. The indicators are most useful when combined in a hybrid model. For this reason, a datasheet has been developed that includes several indicators and their evolution over time (Figure 15).

The datasheets are based on standard html files, providing an interactive interface with time on the x-axis (2014–mid-2024 in the demonstrated case). The display is divided into five sections that collectively illustrate turnout condition: track geometry quality at the frog, indicating bedding condition; sliding ballast condition index, averaged over ±5 m around the frog, reflecting load distribution; irregularities caused by welded joints near the frog, which can increase dynamic loads; axle box acceleration-based indicators, highlighting measurements on the frog side; and dynamic forces derived from geometry data, double-integrated accelerations, and measuring wheels, available for a limited number of turnouts.

The depicted turnout was renewed in 2012. Since then, isolated track geometry failures have developed at the frog due to local settlements. Tamping interventions in 2016, 2018, and 2020 were only partially effective, likely due to hollow sleepers and the complex frog geometry. The frog exchange in 2023 corrected the bedding, resulting in notable improvements in both geometry and ballast condition. Long-term monitoring will show whether these improvements are sustainable or if further settlement occurs over time. Analysis of ramp angles reveals that weld irregularities connecting the frog to the standard rail gradually increased since 2012. The 2020 repair addressed the wing rail but left the frog weld unchanged; after the 2023 frog exchange, both welds are now in good condition, and no further intervention is expected. Repair welding in 2020 also significantly reduced axle box accelerations, stabilising them at low levels. Interestingly, the 2023 frog exchange slightly increased accelerations, reflecting a “run-in phase” during which the new frog geometry smooths under traffic. P₂ forces derived from manually measured geometry data corroborate the trends seen in the indicators, showing both deterioration and improvement, although limited data prevent a full evaluation of maintenance effectiveness. Figure 16 presents a statistical summary of the indicators available across the network.

The statistical analysis is based on randomly selected turnouts that were not part of the index development process. Subplots (a–c) show a dataset of 176 turnouts, including all measurements passing each turnout since 2001. For the acceleration-based index, only measurements from 2012 onward are considered, as earlier signals were unstable. Due to data limitations, weld ramp angles could not be extracted from the same turnout dataset. Instead, a larger sample containing all welds in turnouts (227 welds) was used. Since the dataset does not allow differentiation between weld types, all welds are included, which may slightly bias the values in either direction.

Median values represent the “typical” value for each index. However, as indicated by the whiskers, outliers, and violin plots, the distributions are strongly right-skewed rather than normal. This skewness is most pronounced for the ballast condition index, with a mean of 0.71 compared to a median of 0.4. Such distributions are characteristic of asset quality indicators, where most assets are in good condition, but a small proportion exhibits significantly elevated values as deterioration occurs.

4. Discussion/Outlook

The quality indicators outlined in this paper employ existing data sources, typically provided by standard track-recording cars and/or standard vehicles equipped with axle box acceleration (as used by on-board monitoring concepts), to provide meaningful information about the condition of the frog. By integrating the results with the approaches presented in [36], the toolbox for turnout monitoring can be expanded, leading to a more comprehensive understanding of turnout condition and quality behaviour. The quality assessment is thus both component-specific for fixed-turnout frogs and applicable to the entire main network (as measured by the track-recording car), representing a significant advancement over existing studies. The toolbox can compare the quality behaviour of turnouts and highlight critical areas, but it cannot yet suggest maintenance plans. To do so, intervention limits must be defined based on the quality indicators. While this is straightforward for regulated safety critical quality indicators such as longitudinal level, defining technical–economic (preventive) limits is a more complicated task. Statistical evaluations performed in this study represent a first step toward defining meaningful intervention limits based on the status quo. However, since the status quo is unlikely to reflect the system optimum, additional methodologies should be considered to identify more effective strategies. A potential methodology is provided by [37]; however, the number of turnouts must be extensively extended to generate a respective descriptive model, which we see as an essential part of future research. Another promising research direction concerns the integration of all approaches into a unified health assessment index suitable for high-level condition evaluation.

While the track-recording car data is capable of describing the turnout (frog) condition to a certain extent, further data sources may be a significant development in turnout/frog monitoring. Therefore, future research aims to incorporate additional data sources and compare the feasibility and limitations of each data source individually, as well as the benefits of combining monitoring concepts. The intention is to analyse four concepts in detail: (1) usage of track-recording car data as outlined in this study, (2) a turnout-recording car dedicated for turnout inspection, (3) locally installed sensors for continuous turnout monitoring, and (4) usage of data generated by tamping machines for ballast condition evaluations.

Dedicated turnout-recording cars provide data that is currently not available from any other source. Safety critical gauge values at several points of the turnout, usually measured during manual inspection, can be provided more quickly and objectively for the loaded and unloaded state of the turnout. This enables a better understanding of the geometric condition under load and therefore better reflects the actual situation of a train passing the turnout. Also, the geometrical condition of the frog can be monitored in detail using the 3D scan provided by turnout measurement cars. In comparison to the track-recording car, data is available for the entire turnout, as opposed to just the thorough-going branch. It is anticipated that the utilisation of a turnout-recording car has the potential to extend inspection intervals and substantially enhance the capabilities of predictive maintenance.

While a turnout-recording car applies the concept of a recording car to turnouts, the usage of locally installed sensors allows turnout monitoring from a different perspective. By integrating technologies such as acceleration, acoustic, and displacement sensors, infrastructure managers can accurately monitor the real-time behaviour of turnouts. These sensors provide the potential for continuous condition monitoring, allowing early detection of potential failures. The data collected has the potential not only to improve predictive maintenance strategies, but also to support the development of advanced diagnostic methods and improve efficiency. Ongoing research aims to determine what specific parameters these sensors can capture and how they can help optimise turnout availability.

A further option to enhance turnout monitoring is to outfit maintenance machinery with sensors, subsequently analysing the track during the maintenance process. Ongoing research aims to extend the tamping machine-based ballast condition assessment methodology introduced in [38] to the more complex context of turnouts. Unlike the open track, where ballast degradation, apart from effects such as welded joints, insulated rail joints, or other irregularities, primarily results from gradual abrasion under relatively uniform loading conditions, turnouts are subject to highly heterogeneous dynamic and static loads. These loads vary significantly across the turnout structure—for instance, while the closure panel section may resemble straight track in terms of loading, the frog area is exposed to high vertical impact forces and dynamic loads due to wheel transitions. Such localised variations in mechanical stress are likely to cause distinct patterns of ballast degradation, which cannot be captured by methods calibrated on straight track alone. By leveraging sensor data collected during tamping operations as described in [39]—such as penetration resistance and vibration signals—it becomes possible to characterise the ballast condition in turnouts with greater spatial resolution and objectivity. This expanded diagnostic functionality not only provides information on the expected effectiveness of tamping interventions, but also helps to determine the need for ballast cleaning or renewal. This analysis is important for understanding the mechanical behaviour of ballast in critical turnout areas and for developing predictive, data-driven maintenance strategies for complex track infrastructure.

Initial results of the respective research projects will be presented in forthcoming scientific publications.