Methodology to Detect Rail Corrugation from Vehicle On-Board Measurements by Isolating Effects from Other Sources of Excitation

Abstract

Detecting track geometry and rail surface defects using on-board vehicle monitoring systems is a key issue for rail infrastructure managers to increase availability and reliability while reducing the costs associated with monitoring and maintenance. Rail corrugation is one of the most common rail surface defects which grows in almost all metro, conventional and high-speed lines. This paper focuses on the development of a methodology to detect rail corrugation using axle box acceleration measurements acquired on an in-service high-speed vehicle. The main purpose of the proposed methodology is to distinguish the effect of rail corrugation on the accelerations from the other excitations that can be observed in the same wavelength range. For this purpose, the accelerations are analysed by calculating the fast Fourier transform and the spectrogram. Based on the characteristics of each excitation, the effects of modes of vibration, resonances, bridges, switches, and wheel defects are identified. From the remaining effects, which have congruent characteristics, a hypothesis of rail corrugation is formulated. The hypothesis is consolidated with multibody dynamics simulations and by comparing the corrugation indicators provided by the railway infrastructure company.

1. Introduction

Wheel/rail interaction is the main cause of the rise and growth of irregularities in vehicles and tracks. These irregularities result in unwanted vibrations and dynamic loads, which affect safety and the ride comfort for passengers and lead to component damage. Among all track irregularities, rail corrugation can be defined as the quasi-sinusoidal vertical defects on the top rail surface with wavelengths shorter than 1 m. In addition to dynamic loads and vibrations, very short wavelength irregularities are also the dominant source of rolling noise. Rail corrugation, like other track irregularities, has a significant impact on the inspection and maintenance operating costs for rail network managers. The current approach for the monitoring of these irregularities relies on the use of very sophisticated and expensive visual and optical measurement systems. For these reasons, much effort has been made in recent decades to develop new methodologies to monitor track condition using inexpensive sensors mounted on in-service vehicles, such as accelerometers and microphones. This allows continuous monitoring of the infrastructure, increasing the availability of transport service and reducing maintenance costs. Railway network managers are interested in methods to detect rail corrugation and new approaches to separate the effect of rail corrugation from other excitations arising in the same wavelength range. For this reason, the aim of this paper is to analyse axle box acceleration measurements acquired on in-service vehicles to detect rail corrugation following an exclusion process.

The paper is structured as follows: in Section 2 the state of the art about rail corrugation is presented; in Section 3 the methodology to detect rail corrugation is described in detail; in Section 4 the results of the methodology applied to the use case are presented; and finally, the conclusions are given in Section 5.

2. State of the Art

Rail corrugation has been studied extensively over the past decades, and the knowledge on its formation and characteristics has increased. However, the phenomenon of rail corrugation is still not fully understood and additional investigations are needed to address the open questions left in the literature. Some of these open questions are, for example: what determines the track sections where corrugation develops and what complex characteristics of the vehicle/track interaction determine its wavelength and amplitude.

Since corrugation refers to a wide variety of defects caused by many different factors, a universal definition and classification have not been assigned yet. In [1,2,3,4] the authors attempt a rail corrugation classification based on the root cause and define six categories: pinned–pinned resonance (or roaring rails), rutting, heavy load, light rails, special track form and other P2 resonance. Grassie was the first to describe rail corrugation as a wavelength-fixing mechanism that results from the vehicle/track dynamic interaction or as a damage mechanism that is caused by plastic deformation, rolling fatigue or wear. Another common classification of rail corrugation found in the literature is based on the wavelength range: short pitch (20–100 mm) and long pitch (>100 mm) [5]. In the German language rail corrugation occurring in a tangent track is called “Riffel” (translated as ripple) and in a curved track “Schlupfwellen” (translated as slip waves). These two definitions can be considered similar to roaring rails and rutting, respectively [6].

Although the literature is not uniform in its definition and classification, many publications can be found about the formation and development of rail corrugation. In [7], rail corrugation found on railway networks around the world is described with causes and treatments. The influencing track characteristics are also analysed: rail material, rail support, rail surface irregularities, track stiffness, etc. In [8], the formation of short pitch rail corrugation, found on a metro line, is analysed with experiments, and two mathematical models are formulated to evaluate the appropriate possible remedies. In [9], the causes of rail corrugation development and the corresponding control measures are analysed using a finite element model of the track structure and the dynamic model of the vehicle/track interaction.

Based on these analyses and investigations, it is found that the formation and growth of rail corrugation are related to the characteristics of the interaction between the vehicle and the track. Metro lines and sharp curves are generally more prone to short pitch rail corrugation [10,11,12]. However, corrugation has also been widely observed on conventional and high-speed lines [13,14]. The vehicle/track interaction is the main focus of the analysis of rail corrugation: in [15], the effect of rail corrugation on the dynamic behaviour of the vehicle is analysed with simulations. A Kalker’s non-Hertzian contact model is modified and used to calculate the frictional work density on the contact area and the rail surface wear depth is determined. In [16], an extended Hertzian contact model is used to generate a numerical wear profile history and to study the influence of contact-induced wear filtering on short pitch corrugation growth. A modified Kalker’s non-Hertzian non-steady three-dimensional wheel/rail contact model is also used in [17] to calculate wear for various rail corrugation wavelengths and amplitudes.

In addition to the formation, evolution and countermeasure of rail corrugation, and besides the influence of the dynamic interaction of wheels and rails, another important topic in the literature is how to detect rail corrugation based on the dynamic response of the vehicle. Different methodologies can be found in the literature on how to process the vehicle accelerations to detect rail corrugation: the wavelet transform and the more conventional Fourier transform are the most common approaches; some publications opt for the Hilbert–Huang transform. In [18], the advantages and disadvantages of the wavelet, Fourier and Hilbert–Huang transforms are described and compared. In [19], the axle box accelerations, measured by an inspection car on a conventional line, are used to monitor rail corrugation by comparing the vehicle response with the rail surface roughness acquired by a trolley equipped with laser sensors. The rail roughness is detected in the wavelength range 40–250 mm, and the accelerations are filtered accordingly. A high correlation is observed between the standard deviation of accelerations and the standard deviation of rail roughness, calculated within windows of 25 m. The authors also investigate the accuracy achievable considering different amplitudes of the rail roughness. In [5], the same rail roughness is compared with the internal cabin noise level. By comparing the noise level in the cabin before and after grinding maintenance, it is found that the noise is significantly influenced by the rail surface defects. A high correlation between the internal noise level and the rail roughness is evaluated in terms of standard deviation. In [20], acceleration signals are acquired on a metro line and the wavelet technique is used to monitor the rail conditions. A comparison is made between the wavelet and Fourier approaches to show how better resolution can be achieved in the scalogram compared to the spectrogram. In [21], the track measurements are used to perform simulations and the resulting axle box accelerations are used to detect welds and corrugations. The analysis is performed comparing the accelerations with and without rail corrugation in terms of amplitude of the peaks observed in the PSD. In [22], the rail corrugation wavelength is evaluated analysing simulated axle box accelerations applying the ensemble empirical mode decomposition, then a support vector machine model is developed to estimate the rail corrugation depth.

The effect of rail corrugation on vehicle response measurements has been widely investigated in the literature. However, a detection method capable of distinguishing rail corrugation from other effects such as wheel defects, vehicle modes, rail modes, etc. has not yet been developed.

3. Methodology

In this paper, vertical axle box accelerations, acquired on in-service high-speed vehicles, are analysed in the wavelength range of rail corrugation. The observed excitations are investigated, and the root causes are identified. The acceleration content that cannot be otherwise explained is considered as potential rail corrugation. The final step is to verify these track sections through visual inspections or laser optical measurements of the rail surface. The effects that have a significant impact on the vehicle response in the vertical direction at high frequency are several: vehicle and track modes of vibration, wheel defects, rail defects, construction singularities, (bridges, tunnels, viaducts, etc.), switches, welds, changes in track stiffness, etc. [23]. These can be divided into three categories according to the vibrations caused: frequency-constant (vibration mode), wavelength-constant (wheel and rail defects) and impulse excitations (stiffness changes, switches, welds, etc.) which excite a very wide range of frequencies/wavelengths. The proposed methodology can be summarised in four steps, as shown in Figure 1:

• Data pre-processing: the vehicle measurements are transformed into distance domain;
• Fast Fourier transform (FFT) analysis within track sections;
• Spectrogram analysis to exclude excitations not caused by rail corrugation;
• Hypothesis of rail corrugation for excitations not attributed to other sources.

In the data pre-processing phase, the vertical axle box accelerations are transformed into distance domain by integrating the vehicle speed and synchronised with the track data based on the curvature. The time and distance domain signals are used respectively to investigate the frequency/wavelength content observed in the accelerations. Since there is no available information about where to search for rail corrugation, the section to be analysed is selected based on the noise level. The microphone measurements are A-weighted to filter the signal in the audible range of frequencies [24], and then a high equivalent pressure level (Leq) is searched in the wavelength range of interest (ripple: 10–100 mm; slip waves: 30–300 mm). The (Leq) is evaluated as [25]:

$$Leq=10{\log }_{10}\left({\displaystyle \frac{\frac{1}{T}\underset{t1}{\overset{t2}{\int }}{p}^2\left(t\right)dt}{{p_0}^2}}\right)$$

(1)

where p is the instantaneous sound pressure and p₀ is the reference pressure which is equal to 2 × 10⁻⁵ Pa.

Once the section is selected, the FFT analysis is performed to evaluate whether the excitations found in the microphones are also observed in the accelerations. Then, some FFT comparisons are made to further understand the excitations: axle box accelerations acquired on the same side of the vehicle should have the same peaks if caused by rail corrugation. Comparing the left and right axle box accelerations, the effect of corrugation is expected to be different, especially in curves where slip waves are expected to be on the inner (low) rail. On the contrary, wheel and rail modes of vibration are expected to be the same on both sides. Furthermore, if there are runs at different vehicle speeds over the same section, an FFT comparison can be made to assess whether the observed excitations are constant in the frequency or wavelength domain. This comparison makes it possible to clearly distinguish modes of vibration from rail or wheel defects. From the FFT analysis, a preliminary hypothesis can be formulated about the excitations found in the wavelength range of interest. To finalise the hypothesis, a spectrogram analysis is carried out. This analysis consists in excluding all known excitations, so that the remaining effect is considered as potential rail corrugation. By evaluating the spectrogram in the frequency domain over a longer section (a few kilometres), known excitations can be identified: the modes of vibration appear as horizontal lines (constant frequency) over many kilometres where the vehicle speed is high enough to excite them. Bridges, switches, tunnels and other instantaneous changes in track stiffness appear as vertical lines in the spectrogram because they excite all the frequencies in a very short section of the track. Wheel defects, such as out-of-roundness, wheel corrugation and wheel flats, are wavelength-constant effects, as is rail corrugation. Therefore, they appear in the frequency domain spectrogram as excitations with the same shape as the vehicle speed, since:

$$f={\displaystyle \frac{V}{\lambda }}$$

(2)

where f is the frequency, V is the vehicle speed and λ is the wavelength.

To distinguish between wheel defects and rail corrugation, it is important to assess for how long the excitation lasts: rail corrugation is expected to be found for a few hundred metres, while wheel defects last for the entire acceleration acquisition.

The last step of the methodology is to formulate a hypothesis about the excitation in the wavelength range of the rail corrugation, that cannot be attributed to other causes. The hypothesis can be supported by multibody dynamics (MBD) simulations, which can also be useful to evaluate the vehicle elastic modes of vibration and compare them with the horizontal lines in the spectrogram. In the present work, an MBD model (in SIMPACK) of a commuter train, whose parameters are confidential information, was equipped with an elastic wheelset to analyse the influence of these short wavelength excitations. In the simulation model a co-running track model is implemented. The track model has two layers—under rail pad and under sleeper pad—and is parameterised as a mid-stiffness-like track. The vehicle has a standard S1002 wheel profile, and the rail profile is 60E1 with a standard 1:40 rail cant. The operating speed of the train on the given network is about 160 km/h and, considering the required excitation wavelength of about 30 mm to 300 mm, this means that the elastic model of the wheelset should cover high frequencies, approximately up to 3000 Hz. The finite element model of the wheelset was built from first and second order solid elements and reduced using component mode synthesis (CMS-Method) [26] and incorporated into the vehicle model (Figure 2).

The general rail measurement data could not cover the required wavelength range for these investigations, so the excitations were synthetised from axle box accelerations. As not only the rail can have such short wavelength defects, but also the wheel can have corrugation and OOR defects, the question arises as to whether it is possible to distinguish between them. In order to answer this question, several simulation scenarios were carried out. Firstly, rail corrugations only, then the wheel corrugation/OOR only and then both defects combined.

4. Results

This section presents the results obtained applying the methodology described in the previous section. The available data are measured on an in-service vehicle travelling on a high-speed line in Germany (maximum speed 200 km/h). The non-motorised vehicle under consideration is equipped with eight axle box accelerometers and eight microphones. The measurement locations are shown in Figure 3. The sampling frequency of the data is 8 kHz.

From the acquired data, a section is identified based on the microphone measurements searching for the maximum value of the (Leq) calculated every 100 m.

A section is identified and highlighted in relation to the curvature in Figure 4. The noise level is significantly higher on the left side than the right one, for both the first and second wheelset. As the section represents a left-hand curve, the noise level is in accordance with the definition of slip waves: rail corrugation that occurs on the inner/low rail in curves. Therefore, the potential rail corrugation is expected on the left side and the following investigation is focused on the left axle box accelerations (acc_left_1, acc_left_2).

Firstly, it is important to verify that the noise is not caused by a bridge, switch or tunnel. The location of these construction singularities is shown in Figure 5 in combination with the axle box acceleration measurements. No singularities are found in the highlighted section and therefore the FFT analysis is carried out.

To verify that the axle box accelerations are consistent with the microphone measurements, the FFT analysis is performed on the left side as shown in Figure 6. The acceleration measurements clearly show peaks within the wavelength range of the rail corrugation. These peaks can be observed on the left side on both the first and second wheelset of the front bogie (main acceleration peaks at wavelengths: 65, 60, 47 mm).

The decision to analyse the left side is based on the Leq and the curvature in the selected section. However, to support the preliminary hypothesis, it is important to compare the accelerations on both sides. To do this, the amplitude spectrum in the 100 m section is compared between the left and right axle box accelerations of the first wheelset in Figure 7. Looking at the three main peaks, there is a higher excitation on the left side, where slip waves are expected, but the peaks can also be seen on the right side with much lower amplitudes. This can be explained by the vibration transfer path of a rigid wheelset.

At this stage of the analysis, the characteristics of the peaks found in the 40–70 mm wavelength range in the 100 m section are consistent with the rail corrugation hypothesis. However, these excitations may be due to other effects. To further investigate the cause, it is important to understand whether the peaks are constant in frequency or in wavelength. To do this, three passages at different vehicle speeds over the section are compared. Figure 8a shows the three vehicle speeds (42.3, 30.2 and 18.2 m/s) considered for comparison over the 100 m section highlighted. In Figure 8b,c, the passages are compared in terms of FFT in the frequency and wavelength domain, respectively. It is clear from the comparison that the peaks analysed are constant in the wavelength domain and not in the frequency domain. This means that the cause could be rail or wheel defects. Therefore, constant frequency excitation, such as vehicle and track modes of vibration, can be excluded. The rail corrugation hypothesis is still valid, but the spectrogram analysis is required to distinguish between wheel and rail defects. If runs at different vehicle speeds over the same section are not available, this result can also be achieved with the spectrogram analysis, as shown in the following.

With the spectrogram analysis along the track, a complete overview of the axle box acceleration frequencies can be easily visualised to complete the exclusion process. The axle box acceleration and vehicle speed in the time domain can be seen in Figure 9a. The axle box acceleration data are transformed into the frequency domain using a short-time Fourier transform (STFT), see Figure 9b. To compute the STFT the software Python (SciPy 1.10.1 signal package) is used, and the following parameters selected:

• Moving window length: T = 0.5 s;
• Tukey window with shape parameter: p = 0.25;
• Moving window step size: dt = 0.01 s.

In Figure 9b,c, the increasing amplitude is represented with a colour change from dark blue to bright yellow. This representation allows a specific analysis of the excitations and the corresponding cause:

• The vehicle and rail modes of vibration are constant excitations within a frequency range equal to the eigenfrequencies. In the spectrogram they can be seen as horizontal lines over many kilometres where the speed is high enough to excite the modes, see Figure 9b. The eigenfrequencies can also be estimated using the MBD simulation including the elastic model of the wheelsets. These values are shown in Figure 9c as horizontal pink lines. The estimated modes of vibration can guide the spectrogram analysis in the frequency domain to better identify the horizontal lines. However, it is important to notice that they are not necessary to apply the methodology. While searching for rail corrugation, any frequency-constant excitations can be discarded, whether they are identified as modes of vibration or not.
• The bridges and switches shown in Figure 5 appear as impulsive vibrations that excite all frequencies in a very short time. Therefore, they can be seen as vertical lines in the spectrogram in Figure 9b and as highlighted vertical sky-blue lines in Figure 9c. Again, there is a good agreement between the construction positions and the resulting effects on the axle box accelerations. However, not all of them have a significant effect on the accelerations, especially where the vehicle is running at lower speed.
• The wheel defects in the wavelength range of interest are mainly wheel corrugation (30–60 mm) and wheel out-of-roundness (OOR) within 140 − 2πR mm, where R is the wheel radius [27]. These defects are constant in the wavelength domain, and, unlike rail corrugation, should be observed along the entire acceleration signal. Being constant in the wavelength domain means, for the frequency domain spectrogram, that they have the same shape as the vehicle speed. Some wheel defects can be observed in the spectrogram in Figure 9b and are highlighted with white dashed lines in (c).

Once modes of vibration, construction and wheel defects have been excluded from the spectrogram, the remaining excitation is considered to be rail corrugation if it is constant in wavelength and lasts for a few hundred metres. Observing the section identified at the beginning of the analysis, some excitations can be seen in the 500–1000 Hz range. To better visualise the wavelength range of interest, the acceleration spectrogram is evaluated in the distance domain with respect to wavelength, see Figure 10. For the spectrogram in the wavelength domain a logarithmic scale is used for better visualisation.

As observed in the FFT, the peaks are constant in wavelength and appear as horizontal lines in the spectrogram. Observing Figure 9b, in the highlighted section these peaks are all slightly decreasing as the vehicle speeds up. This means they are not constant in frequency and therefore they cannot be excited modes. In the spectrogram it is also possible to analyse how long each excitation lasts. Peak “A” is visible in bright yellow throughout the signal and is therefore very likely to be a wheel defect. Peaks “B” and “C”, on the other hand, are only very intense for about 200 m, which is the expected length of rail corrugation. From this analysis, the final hypothesis can be formulated: rail corrugation is found in the identified 100 m section, with main wavelength contents in the range of 47–60 mm.

4.1. Simulations to Support the Formulated Hypothesis

Optical measurements data of the rail surface are not available to verify this hypothesis, but some further investigations can be carried out to support it. MBD simulations can be used to study the interaction between wheel defects and rail corrugation, and some rail corrugation indicator signals are available to be compared with the results obtained.

The MBD simulations are used to simulate different scenarios and investigate the mutual influence between the wheel defects (OOR and corrugation) and the rail corrugation (represented as sinusoidal signal). The simulations are performed on the 100 m section selected in the FFT analysis, using the same track layout and curve characteristics. First, the wheel corrugation is analysed considering six scenarios (three with wheel corrugation only, three with the same wheel corrugation and additional rail corrugation):

• wheel corrugation with a wavelength of 60 mm and an amplitude of 10 µm;
• wheel corrugation with a wavelength of 50 mm and an amplitude of 8 µm;
• wheel corrugation with a wavelength of 40 mm and an amplitude of 6 µm;
• rail corrugation with a wavelength of 61 mm and an amplitude of 5 µm + wheel corrugation 1;
• rail corrugation with a wavelength of 61 mm and an amplitude of 5 µm + wheel corrugation 2;
• rail corrugation with a wavelength of 61 mm and an amplitude of 5 µm + wheel corrugation 3.

Figure 11a shows the reference FFT of the axle box acceleration (Figure 6) with the ripple and slip waves wavelength range. Figure 11b shows the FFT of the axle box accelerations from three simulations. These three simulations are performed considering only wheel corrugation as in scenarios 1, 2 and 3. Figure 11c shows the results of wheel and rail corrugation for scenarios 4, 5 and 6.

Wheel and rail corrugation are the same type of defect that can occur on both components with similar wavelength and amplitude [27]. From the simulation results shown in Figure 11, it is clear that they have no relevant mutual influence and, when they occur at the same time, the effect is simply the sum of the two responses. This result supports the spectrogram analysis approach, which distinguishes between wheel and rail corrugation on the basis of the length of the effect observed in the acceleration.

The interaction between wheel OOR and rail corrugation can be investigated with MBD simulations following a similar approach. Different scenarios with only OOR and with both OOR and rail corrugation are considered:

• 1st order OOR (wavelength 2889 mm) with amplitude 1 mm;
• 2nd order OOR (wavelength 1444 mm) with amplitude 470 µm;
• 3rd order OOR (wavelength 963 mm) with amplitude 300 µm;
• 5th order OOR (wavelength 578 mm) with amplitude 160 µm;
• 15th order OOR (wavelength 193 mm) with amplitude 27 µm;
• 20th order OOR (wavelength 144 mm) with amplitude 10 µm;
• rail corrugation + OOR 1;
• rail corrugation + OOR 2;
• rail corrugation + OOR 3;
• rail corrugation + OOR 4;
• rail corrugation + OOR 5;
• rail corrugation + OOR 6.

In Figure 12 the reference from the measurements is shown again in (a), in (b) the scenarios considering only OOR (1, 2, 3, 4, 5, 6), are shown and in (c) the scenarios combining both OOR and rail corrugation are shown.

As mentioned above, the wavelength range of the OOR is typically between 140 mm and the wheel circumference, which is greater than the peaks observed in the axle box acceleration measurements. The scenarios presented in Figure 12b prove that only the unusual very high order of the OOR (15th–20th) can be observed in the wavelength range of the rail corrugation. However, when rail corrugation with a wavelength of 61 mm is added (c), some OOR effects can be observed around the corrugation peak.

These MBD simulation results support the hypothesis formulated in the spectrogram analysis: the peaks found in the axle box acceleration, which last only a few hundred metres, cannot be wheel corrugation due to their length and cannot be OOR if not combined with rail corrugation.

4.2. Further Analysis in Support the Hypothesis

The final analysis in support of the hypothesis is the observation of slip waves and ripple indicators provided by the infrastructure network manager. Figure 13 shows the comparison between the axle box accelerations (a), the microphone data (b), the slip waves (c) and the ripple (d) indicators within the highlighted 100 m section. Slip waves and ripple signals represent the difference between the maximum and minimum every 48 cm of the rail surface profile, filtered within 30–300 mm and 10–100 mm, respectively.

In the selected section, where the hypothesis of rail corrugation is formulated, both indicator signals present relevant peaks. This is further evidence to support the hypothesis. The validation of our hypothesis is made with the available data provided by the infrastructure manager company. The ripple and slip waves indicators are derived from visual inspection data. However, the signal processing applied has not been disclosed.

As already mentioned, the methodology is considered fully validated only when compared with visual inspection and rail corrugation measurements. However, the hypothesis made is strongly supported by all the analyses shown.

5. Discussion

In this paper, a new approach to detect rail corrugation from axle box accelerations is described. First, a section is identified based on the pressure level measured by the microphones. The section is then analysed by calculating the fast Fourier transform (FFT) of the axle box accelerations. The wavelength range of interest is observed, and the excitations found are further investigated by comparing runs at different vehicle speeds over the same section. This comparison allows the FFT contents to be divided into frequency-constant and wavelength-constant excitations. The spectrogram analysis distinguishes between the different sources of excitation: vehicle and rail modes of vibrations are identified as frequency-constant contents; bridges and switches are identified as impulsive effects that excite a wide range of frequencies within a very short time/distance span. Wheel and rail defects, both of which are constant in the wavelength domain, are distinguished based on the length of the response found in the acceleration data: wheel defects remain the same over the entire acquisition for many kilometres, while rail corrugations only affect the accelerations for a few hundred metres.

The proposed methodology is then applied to data acquired on an in-service vehicle traveling on high-speed lines in Germany. The section of interest is identified using the microphone measurements by searching for a high value of the equivalent pressure level. The section is then analysed by computing the FFT: some wavelength-constant peaks are observed within the rail corrugation range. These peaks are then investigated in the spectrogram analysis to distinguish between wheel defects and rail corrugation based on the length of the excitations. One of the three peaks corresponds to an excitation that lasts for many kilometres and is therefore considered to be a potential wheel defect. The other two peaks, on the other hand, are only visible in the spectrogram for a few hundred metres and so the hypothesis of rail corrugation is formulated.

The present work is considered a low technology readiness level (TRL) study to investigate the possibility to detect rail corrugation among many hundreds of kilometres using only measurements acquired on-board the vehicles. Visual and optical measurements could also be used to detect rail corrugation. However, it is important to acknowledge that the laser optical sensors are expensive and cannot be mounted on a complete fleet of vehicles. They are reliable, although they can be affected by environmental conditions such as dirtiness or light effects. On the contrary, accelerators and microphones can be easily installed on many vehicles, providing a large dataset of measurements. This allows both to continuously monitor the track and several vehicles to pass on the same track section. Comparing the acquisition from different vehicles on the same track section makes it possible to distinguish between the effects due to the vehicle component irregularities and to the rail irregularities.

Further steps in the methodology validation are required before the implementation in real-world conditions. The automatisation of the methodology is an important aspect to be considered when implementing the methodology in practical settings. To this purpose the current manual steps should be simplified and adapted for an automatised implementation. These modifications can be performed without affecting the quality of the results achieved.

6. Conclusions

In this paper, a new methodology is presented to distinguish the effect of rail corrugation on the vehicle response from the other sources of excitation. This is achieved with the prior knowledge of their impact on the vehicle accelerations: modes of vibrations and resonances are constant in the frequency domain, constructure singularities have an impulsive effect, wheel defects are constant in the wavelength domain and their excitation lasts for the entire acquisition, etc. However, open issues remain for the detection of rail corrugation when combined with one or more other excitation sources and for the exclusion of other sources with unknown effects on vehicle accelerations. The goal of the present work was to investigate the possibility to detect corrugation when the only available data are axle box accelerations and acoustic measurements acquired on in-service vehicles provided with onboard monitoring systems. Labelled data are required to improve the current approach and to investigate these open points. Accelerations acquired over the same section, with and without corrugation, should be compared to separate the specific excitation of interest. Once the methodology is fully validated, it can be simplified and the spectrogram analysis automatised for application in practical settings. With these additional data, further approaches such as machine learning classification and other deep learning techniques can be implemented.