WO2021094401A1

WO2021094401A1 - Device and method for maintenance planning for a train wheel

Info

Publication number: WO2021094401A1
Application number: PCT/EP2020/081802
Authority: WO
Inventors: Urs Gehrig; Sonja GASSNER
Original assignee: Schweizerische Bundesbahnen Sbb
Priority date: 2019-11-15
Filing date: 2020-11-11
Publication date: 2021-05-20
Also published as: EP4058338A1; EP4058338B1; EP4058338C0

Abstract

A computer-implemented method for maintenance planning for a train wheel, the method comprising: receiving (S1) measurement data of the train wheel; determining (S2), using the measurement data, whether the train wheel will need to undergo maintenance in the future; and predicting (S3), if the train wheel will need to undergo maintenance in the future, a maintenance point (M1), in particular a critical time point (T1) and/or a critical travel distance (D1) until maintenance.

Description

DEVICE AND METHOD FOR MAINTENANCE PLANNING FOR A TRAIN WHEEL

FIELD OF THE INVENTION

The present invention relates to a method and system for maintenance planning for a train wheel. Specifically, the present invention relates to a method, a computer, a maintenance system, and a computer program product for maintenance planning for a train wheel by predicting a maintenance point for the train wheel.

BACKGROUND OF THE INVENTION

Trains are highly complex systems that are subject to harsh operating and environmental conditions. To ensure safe and efficient operation, regular maintenance is required. Train wheels in particular are a crucial component of the rolling stock and, as train wheels are subject to high stresses and wear and tear, they must be well maintained to minimize potential safety hazards and to increase the longevity of the train wheels. Train wheels can become worn through long term use and/or become damaged due exposure to harsh environmental and/or operating conditions. In particular, causes of damage to train wheels are due to drag breaking down steep inclines or emergency braking, which results in so- called "flat spots" on the train wheels.

Types of wear seen on train wheels include abrasive wear, which occurs by several different mechanisms such as microcutting, microploughing, microfatigue, or microcracking. Abrasive wear occurs when a relatively harder surface, e.g. of the wheel, has slipping movement against a relatively softer surface, e.g. of the rail, or when there are fine particles such as sand between the train wheel and the rail. Another type of wear is adhesive wear, which is produced with non-ideal contact surfaces resulting in sliding and/or slipping, in particular when the rail is curved leading to a shift in the contact point between train wheel and rail. Other types of wear include delamination wear, tribochemical wear, fetting wear, surface fatigue wear, impact wear, etc. The manner in which wear progresses in train wheels, ultimately resulting in maintenance or replacement of the train wheels being required, is not yet fully understood. Traditionally, trains, including the train wheels, were inspected using a time-based method, such that the train or train carriage would be inspected according to a pre determined maintenance schedule. This was not particularly efficient, as it led to situations where a particular train which did not see much use was being inspected and maintained at an unnecessarily high rate. Therefore, mileage based methods were introduced, in which the trains were inspected according to the distance covered, with checks occurring depending on the usage of the train. This, however required accurate and voluminous record-keeping and became less cumbersome only with the introduction of widespread electronic record keeping.

In the modern era, with widespread electronic telecommunications networks and the ability to deploy networked sensors, condition based monitoring has become feasible. In condition based monitoring, sensors are deployed and configured for automatic inspection of various components of the trains, and maintenance checks are triggered if a sensor reading indicates that a particular component requires inspection and/or maintenance. However, this may still result in some components being inspected and/or maintained too frequently, i.e. before it would actually be necessary, for example if the maintenance check is triggered with a large safety margin. Alternatively, condition based monitoring may also result in components being inspected and/or maintained too infrequently, i.e. after it would actually have been necessary, for example if automatic inspection occurs at large time intervals and the maintenance check is triggered with a relatively lower safety margin. SUMMARY OF THE INVENTION

It is an object of this invention to provide a method, a computer or generally computer system, a maintenance system, and a computer program product for maintenance planning for a train wheel, which method, computer, and computer program product do not have at least some of the disadvantages of the prior art. In particular, it is an object of the present invention to provide a method, a computer, a maintenance system, and a computer program product for a train wheel by improved predicting of a maintenance point for the train wheel.

According to the present invention, these objects are achieved through the features of the independent claims. In addition, further advantageous embodiments follow from the dependent claims, claim combinations and the description including the figures.

According to the present invention, the above-mentioned objects are particularly achieved in a computer-implemented method for maintenance planning for a train wheel, the method comprising receiving, in a processor, measurement data of the train wheel; determining, in the processor, using the measurement data, whether the train wheel will need to undergo maintenance in the future; and predicting, in the processor, if the train wheel will need to undergo maintenance in the future, a maintenance point, in particular a time interval or a time point or a critical time and/or a distance or critical distance or number of kilometers until maintenance. The maintenance point is a point in the future, defined by a time that the train wheel may still be in operation for or a specific time point in the future, and/or a distance the train wheel may travel until maintenance is required.

In an embodiment, the method comprises receiving, in a processor, measurement data of the train wheel and predicting, in the processor, a maintenance point, in particular a time interval or a time point or a critical time and/or a distance or critical distance or number of kilometers until maintenance. According to this embodiment, a step of determining whether the train wheel will need to undergo maintenance in the future is expressly not required nor essential. It is only required that a maintenance point is predicted.

At this point it is noted that the maintenance point is the end-point of a maintenance interval in which maintenance must be performed on the train wheel, at least within a reasonable tolerance interval of still allowable or residual time or travel distance. In addition to the maintenance point for the train wheel as described in the present disclosure, local jurisdictions often set guidelines and regulations which impose maintenance schedules onto rail operators. These guidelines and recommendations can, for example, impose a maximum period between inspection and/or maintenance, a maximum mileage before inspection and/or maintenance, and/or a maximum level of wear and/or damage to a particular component which must trigger maintenance or replacement. In an embodiment, the methods described herein are designed to account for such additional external constraints by planning appropriate inspection intervals and maintenance intervals. In particular, the processor is configured to receive the train wheel maintenance schedule comprising a maximum maintenance distance, i.e. the maximum distance the train wheel is permitted to travel before maintenance, and/or a maximum maintenance time, i.e. the maximum time the train wheel is permitted to operate before maintenance, and compare the predicted maintenance point with the maximum maintenance distance and the maximum maintenance time.

In an embodiment, if the predicted maintenance point lies beyond the maintenance interval, then the processor is configured to determine an adjusted maintenance point in conformity with the maintenance schedule, i.e. falling within the maximum maintenance distance and the maximum maintenance time. In an embodiment, receiving the measurement data comprises receiving, in the processor, measurement data relating to a roundness error of the train wheel. The roundness error describes a deviation from roundness or out of roundness (OOR) of the train wheel and may be described by a ratio between inscribed and circumscribed circles. The roundness error may relate to a circularity error of the train wheel. The roundness error may also relate to a surface roughness of the train wheel. Other criteria for the roundness error are also possible, such as e.g. eccentricity of the wheel, dynamic behavior of the wheel, variables representative for roundness defects, other wheel defects, etc.

In an embodiment, the measurement data comprises optical data, for example image data of the train wheel. In another embodiment, the optical data comprises data from one or more laser scans carried out by using a camera to capture reflected light from a laser source, the light being reflected from the train wheel.

In an embodiment, the measurement data comprises data collected using a dial gauge physically touching the train wheel as the train wheel is rotating on a rotating fixture, such as to measure a roundness error along the outer circumference of the train wheel. The dial gauge is a polar recording instrument which produces a measurement trace, which measurement trace is included as part of the measurement data.

In an embodiment, the measurement data relates to data collected using accelerometers, for example accelerometers attached to a rail and configured to measure the vibration induced by a train wheel as it rolls over the rail. In another example, an accelerometer may be affixed to the train wheel, train axle, bogie, or train carriage and may be configured to record the vibration of the train wheel. In an embodiment, receiving the measurement data comprises receiving, in the processor, recorded historical measurement data of the train wheel. The historical measurement data are measurement data recorded in the past, for example measurement data recorded weekly for the past year. The historical measurement data are preferably associated with a particular wheel, but may also be aggregated to comprise measurement data of a set of wheels, for example a pair of train wheels attached to a common axle, or a set of train wheels attached to a common bogie. The historical measurement data may be retrieved from a cloud-based computing system. The measurement data therefore refers to measurement data, present and past, associated with the train wheel. In an embodiment, the measurement data includes additional meta-data relating to the train wheel, such as an identifier which identifies the train wheel, historical measurement data which has undergone processing, such as to remove outliers, and values derived using measurement data.

In an embodiment, receiving the measurement data further comprises receiving, in the processor, a previous maintenance point, in particular a time since last maintenance and/or a travel distance since last maintenance, wherein the last maintenance comprises a repair, in particular reprofiling, or a replacement of the train wheel.

In an embodiment, receiving the measurement data comprises receiving, in the processor, vertical load data of the train wheel as the train wheel rolls across a load measuring station, in particular vertical load measuring station, arranged at the rail. The vertical load data is a series of data points comprising the vertical component of the force due to gravity (vertical load) which thetrain wheel exerts on the rail. The vertical load measuring station is installed along a section of the rail and configured to measure the vertical load as the train wheel W rolls across the section of the rail. A round train wheel free of defects rolling along a rail exerts a constant vertical component of the force. When the train wheel is on the section of the rail where the vertical load measuring station is installed, the vertical load data is accurately measured. However, when the train wheel is situated on the rail either side of the vertical load measuring station, the vertical load measured at the vertical load measuring station is diminished. The vertical load data may be pre-processed using a window function, preferably a rectangular window function, such that vertical load data associated with data points where the train wheel is situated on the vertical load measuring station are retained, whereas vertical load data associated with data points where the train wheel is situated outside the vertical load measuring station are discarded. In an embodiment, the vertical load data comprises a plurality of vertical load time-series measurements, each measurement being made by a different one of a plurality of sensor units, which comprise vertical load sensors, arranged at the load measuring station. The method can comprise identifying, in the processor, from the vertical load data a stable measurement time range in each of the vertical load time-series measurement, which corresponds to the train wheel being within a stable measurement distance of a given sensor unit. The stable measurement distance can be pre-determined. Alternatively, or in addition, the stable measurement time range can be identified using a vertical load threshold of the vertical load data, such that a start point of the time range is identified when the vertical load exceeds the threshold for the first time, and an end point of the time range is identified when the vertical load data falls below the threshold for the last time. The method can comprise pre-processing, in the processor, the vertical load data by removing, from each vertical load time-series measurement, data points lying outside the identified stable measurement time range. Pre-processing the vertical load data may comprise using a window function, for example a rectangular window function. In an embodiment, the method further comprises calculating, in the processor, a dynamic coefficient, which dynamic coefficient is a ratio of a maximum dynamic load to a static load according to the following relation:

wherein the max. (QForce) is a maximum value of the vertical load of the train wheel determined across the plurality of vertical-load time-series measurements, and the static(QForce) is an average value of the vertical load of the train wheel determined across the plurality of vertical-load time-series measurements.

The dynamic coefficient for a round train wheel free of wear, damage, or defects will be close to one, asthe maximum value of the vertical load of thetrain wheel will be essentially the same as the average value of the vertical load data of the train wheel. Train wheels which are worn and/or damaged will roll in an uneven manner resulting in the maximum value of the vertical load exceeding the average value of the vertical load, and the dynamic coefficient therefore being larger than one. It is observed that as wear and/or damage increases, the dynamic coefficient grows larger. If the dynamic coefficient for a train wheel exceeds a critical dynamic coefficient threshold, then the train wheel must undergo maintenance, maintenance including further inspection, repair, and/or replacement of the train wheel. The critical dynamic coefficient threshold is in a range of1 .2 to 6, preferably in a range of 1 4to 4, more preferably in a range of 1 .6 to 2.0, and most preferred is 1 .8. The critical dynamic coefficient threshold can vary depending on the type of train wheel and/or the type of bogie and/or the type of carriage the train wheel is installed on. For example, the critical dynamic coefficient threshold is lower for train wheels installed on a passenger train carriage than for train wheels installed on a cargo train carriage, as the comfort of passengers is a relevant factor for deciding at which point a train wheel must undergo maintenance.

In an embodiment, the method further comprises receiving, in the processor, recorded train mileage data related to the distance traveled overtime by a train car to which the train wheel is attached. For example, the train mileage data comprises a number of kilometers traveled. In another example, the train mileage data comprises time series data relating the number of kilometers traveled totime. The method comprises generating, inthe processor, using the recorded historical measurement data and the recorded train mileage data, historical dynamic coefficient data of the train wheel, comprising a plurality of dynamic coefficients as a function of one or more of: a plurality of corresponding measurement time points, a plurality of corresponding distances travelled over time by the train wheel, in particular since a last repair, one or more previous maintenance points, a combination thereof. The historical dynamic coefficient data enables tracking of the dynamic coefficient data for a particular train wheel. In an embodiment, the method further comprises pre-processing, in the processor, the dynamic coefficient data of the train wheel by removing those dynamic coefficients from the dynamic coefficient data that correspond to time points and/or travelled distances prior to a last maintenance point, in particular a last maintenance time point or a last maintenance travel-distance point. Because the maintenance typically includes reprofiling or replacing the train wheel, dynamic coefficient data prior to the last maintenance point is less relevant. However, such data could give additional information about any influence of previous repair or reprofiling actions on actual or future damage behavior and/or on a number of total or future allowable repair or reprofiling actions. In an embodiment, the method further comprises identifying, in the processor, a discontinuity time point in the dynamic coefficient data, if a difference between a particular later dynamic coefficient and a previous dynamic coefficient is negative and exceeds a pre defined difference threshold, and pre-processing, in the processor, the dynamic coefficient data for the train wheel by removing those dynamic coefficients from the dynamic coefficient data that correspond to time points and/or travelled distances prior to the discontinuity time point. In particular, the discontinuity time point may be detected, if the dynamic coefficient drops by the pre-defined difference threshold and/or drops to a value close to one. Identifying discontinuity time points may be used to detect, in the processor, the last maintenance time point. This is, because the dynamic coefficient of a train wheel after maintenance is close to one.

In an embodiment, determining whether the train wheel will need to undergo maintenance comprises generating, in the processor, using the dynamic coefficient data and a forecasting model, forecasted dynamic coefficients. Determining whether the train wheel will need to undergo maintenancefurther comprises predicting, inthe processor, using the forecasted dynamic coefficients, the maintenance point by determining the critical time point and/or the critical travel distance at which the forecasted dynamic coefficients exceed a critical dynamic coefficientthreshold, in particular wherein the critical dynamic coefficient threshold has a value in a range of 1 .2 to 6, preferably in a range of 1 .4 to 4, more preferably in a range of 1 .6 to 2.0, and most preferred is 1 .8.

The precise value of the critical dynamic coefficient can depend on a number of factors, for example the particular type of train wheel, bogie, and/or carriage, in particular whether the train wheel is installed on a passenger train carriage or a cargo train carriage. The maintenance point is, for example, the remaining distance which the train wheel may travel before maintenance is required. The maintenance point may also be expressed as a time point, for example a date in the future, or a number of days remaining until maintenance.

In an embodiment, the forecasting model comprises on or more of: a linear regression model, a dynamic linear model (DLM), an exponential smoothing model, an ARIMA model, a dynamic linear model, or a combination of these models. The forecasting model can also comprise modifications and/or combinations of such models.

In an embodiment, the processor is configured to use the linear regression model to determine whether the train wheel will need to undergo maintenance by fitting a straight line onto the dynamic coefficient data and checking whether the straight line will exceed the critical dynamic coefficient threshold at a point in the future. Additionally, or alternatively, the processor is configured to use the linear regression model to determine the maintenance point as the point in the future at which the straight line exceeds the critical dynamic coefficient threshold. The advantage of using the linear regression model is that it requires only a small dataset of dynamic coefficient data, and that the predictions are readily comprehensible. A potential disadvantage is that, if the dynamic coefficient data has a large uncertainty and/or error involved, that the predictions may be unreliable.

In an embodiment, the processor is configured to use a dynamic linear model (DLM) to determine whether the train wheel will need to undergo maintenance. The DLM has the advantage that it can more accurately represent long term variations in data with high variability. The dynamic coefficient data in particular may have high variability, depending on the exact type of train wheel and what types of wear and/or damage it has been or is being subjected to. In an embodiment, generating the forecasted dynamic coefficients using the forecasting model comprises fitting, in the processor, a trend curve onto the dynamic coefficient data. It comprises extrapolating, in the processor, the trend curve onto future time points and/or future distances, and determining, in the processor, using the future time points and/or future distances, a critical time and/or a critical distance, respectively, at which the extrapolated trend curve exceeds the critical dynamic coefficient threshold. The critical time can be determined by the processor as a time, for example a time point in the future or a time interval extending from the present into the future.

In an embodiment, determining whether or when the train wheel will need to undergo maintenance comprises: extracting, in the processor, features of the dynamic coefficient data; classifying, in the processor, using the features and a classifier model, the train wheel as a critical train wheel, if a dynamic coefficient of the dynamic coefficient data exceeds and/or is predicted to exceed a critical dynamic coefficient threshold within a critical time and/or a critical distance; and predicting, in the processor, forthe critical train wheel, using the features and a regression model, the maintenance point.

Studying the nature of how wear and/or damage to train wheel occurs, and how this wear and/or damage to the train wheel evolves in time has led to the conclusion that train wheels can be classified either as critical or as non-critical. Train wheels after maintenance and/or replacement are initially non-critical. Wear accumulated during normal operating conditions can result in a train wheel whose dynamic coefficient data may be variable and discontinuous, i.e. a sequence of measurements of the dynamic coefficient does not show a clear trend towards ever larger dynamic coefficients. These train wheels are classified as non-critical as the dynamic coefficient does not suggest or indicate a particular maintenance point. However, if the wear and/or damage to a train wheel is large enough it may begin to accumulate rapidly, resulting in the measured dynamic coefficients, as reflected in the historical measurement data, increasing or diverging in a trend towards the critical dynamic coefficient threshold. These train wheels can de classified as critical, as they are on a perceivable path towards requiring maintenance and/or replacement. Distinguishing whether a train wheel is critical or non-critical is non-trivial, particularly because the variability in the dynamic coefficient data of a particular wheel can be high, and it is not yet fully understood or analytically predictable, how wear and/or damage to a train wheel worsens over time.

In an embodiment, the classifier model comprises using one or more of the following algorithms: logistic regression, k-Nearest neighbors, decision trees, support vector machine, or naive Bayes.

In an embodiment, the classifier model comprises a classifier neural network configured to classify, using the dynamic coefficient data, the train wheel as a critical train wheel or as a non-critical train wheel. The classifier neural network can use as an input time series data, in particular the dynamic coefficient data as comprised in the measurement data and/or the historical measurement data, and as an output provides a classification of whether the train wheel is a critical train wheel or a non-critical train wheel.

In an embodiment, the classifier neural network is trained using a training dataset. The training dataset comprises a large amount of historical measurement data of a large number of train wheels. In particular, the training dataset comprises a large amount of dynamic coefficient data associated with mileage data. In one example, each train wheel of the training dataset includes a label of either critical or non-critical, and the classifier neural network is trained using supervised learning. In another example, the training dataset is not labelled and the classifier neural network is trained to classify train wheels into two categories using unsupervised learning, which categories are then assigned to critical and non-critical.

In an embodiment, the training dataset comprises historical measurement data of only the particular type of train onto which the train wheel W is affixed. For example, the training data can comprise only passenger trains or cargo trains. Further, a particular type of passenger train may be specified, such as the RABDe 500, RABe 51 1 , or ETR 610.

In an embodiment, the training set comprises historical measurement data divided into several classes, each class corresponding to a particular type of train or trainset. The processor is configured to train a plurality of classifier models, in particular a plurality of classifier neural networks, each classifier model being trained using historical measurement data corresponding to one particular type of train or trainset.

In an embodiment, the regression model comprises a neural network regression model configured to generate, using the features, the maintenance point comprising the critical time point and/or the critical distance until maintenance. In addition to a method for maintenance planning for a train wheel, the present invention also relates to a computer or computer system for maintenance planning for a train wheel, the computer comprising a processor configured to perform the method as described herein.

In addition to a method and a computer for maintenance planning for a train wheel, the present invention also relates to a maintenance system for maintenance of a train wheel, comprising the computer as described above and a workshop for performing maintenance of the train wheel, such as re-profiling and/or replacing the train wheel. The workshop can also comprise a computing device, for example a further computer. Depending on the embodiment, the computing device can be embodied as a portable computing device, such as a smart phone or a tablet computer. The computing device can comprise a processor configured to receive, from the computer, a maintenance request message, if the maintenance point, comprising the critical time and/or the critical distance, until maintenance, for the train wheel has been exceeded. The computing device can also be configured to display, on a display of the computing device, the maintenance request message, and transmit, to the computer, a maintenance confirmation message, once maintenance has been performed on the train wheel in the workshop. Herein, the term workshop shall broadly encompass any place where maintenance, in particular repair or reprofiling of a train wheel, can be performed

In an embodiment, the maintenance system can comprise an RFID reader. The computing device, in particular the processor of the computing device, can further be configured to receive, from the RFID reader, an RFID identifier of a particular train wheel of a train car which is present in the workshop, and display, on the display of the computing device, the maintenance request message only, if the maintenance point of the particular train wheel has been exceeded.

In addition to a method and a computer for maintenance planning for a train wheel and a maintenance system for maintenance for a train wheel, the present invention also relates to a computer program product comprising a non-transitory computer-readable medium having stored thereon computer program code configured to control a processor of a computer such that the computer performs the method as described above

BRIEF DESCRIPTION OF THE DRAWINGS The present invention will be explained in more detail, by way of example, with reference to the drawings in which:

Figures 1 A- 1 F: show diagrams and photographs illustrating different types of wear and damage of train wheels; Figure 2A: shows a diagram illustrating schematically a load measuring station for taking a vertical-load time-series measurement for a train wheel;

Figure 2B: shows a diagram and a plot illustrating a vertical-load time-series measurement curve for a train wheel as it rolls over a sensor unit;

Figure 3: shows a diagram and several plots illustrating a wheel rolling across several sensor units of a load measuring station, each sensor unit producing a vertical load time-series measurement;

Figures 4A-FH: show a series of exemplary plots of the dynamic coefficient vs. distance for each of a series of train wheels;

Figure 5: shows a block diagram illustrating schematically a computer for maintenance planning for a train wheel;

Figure 6: shows a flow diagram illustrating an exemplary sequence of steps for maintenance planning for a train wheel;

Figure 7: shows a flow diagram illustrating an exemplary sequence of steps for calculating a dynamic coefficient for a train wheel; Figure 8: shows a flow diagram illustrating an exemplary sequence of steps for receiving recorded train mileage data for a train wheel;

Figure 9: shows a flow diagram illustrating an exemplary sequence of steps for pre processing dynamic coefficient data for a train wheel; Figure 10: shows a flow diagram illustrating an exemplary sequence of steps for predicting a maintenance point for a train wheel;

Figure 1 1 : shows a block diagram illustrating a system for maintenance of a train wheel;

Figure 1 2: shows a flow diagram illustrating an exemplary sequence of steps for maintenance of a train wheel; and Figure 13: shows a flow diagram illustrating an exemplary sequence of steps for indicating a maintenance request for a train wheel.

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

Figure 1 A illustrates a train wheel W which develops damage A in the form of a flat spot. Figure 1 B illustrates how damage A to the train wheel W can involve, in basic terms, removal of material B, or deposit of material C. Figure 1C shows a photograph of a train wheel W onto which material has been deposited. This can occur, for example, during emergency braking when the friction between the train wheel W and the rail is high enough to cause depositing of rail material onto the train wheel W. Figure 1 D shows a photograph of a train wheel W which has been damaged by chipping. A small hole has formed in the middle of the rolling surface of the train wheel W. Figure 1 E shows a photograph of a train wheel W with lamella-like material imperfections across the entire circumference of the running surface. Figure 1 F shows a photograph of a train wheel W which has developed a flat spot.

Figure 2A illustrates schematically a top down view of a load measuring station 2 arranged at a rail 3 with sleepers 5. Vertical load sensors L1 , L2 are arranged ontop of, or embedded into, the rail 3 and are configured to measure a vertical load (Q-Force) overtime as a train wheel W rolls across the rail 3. In particular, the vertical load sensors L1 , L2 are configured to measure a vertical load time series and transmit the recorded vertical load data to a computer system.

Figure 2B shows a diagram and a plot illustrating a vertical-load time-series measurement curve for a train wheel W as it rolls over vertical load sensors L1 , L2. In particular, it can be seen that the measured Q-Force (on the abscissa) over time (on the ordinate), i.e. the vertical load force or Q-Force the train wheel W exerts on the rail 3, is at a very low baseline value when the train W has not yet passed over the vertical load sensors L1 , L2. Once the train wheel W is sufficiently close to the vertical load sensors L1 , L2, it can be seen that the Q-Force rises rapidly and remains approximately constant until the train wheel W passes beyond the vertical load sensors L1 , L2, whereupon the Q-Force drops rapidly back to the baseline value. The relatively smooth and flat Q-Force recorded as the train wheel W is between the two vertical load sensors L1 , L2 is indicative of at least part of the train wheel W not being worn or damaged. An ideal round train wheel W moving at a constant velocity would exert a constant vertical load force onto the rail 3. Because the vertical load sensors L1 and L2 are placed closer together than the circumferential length of the train wheel W, they are only able to measure the Q-Force along part of the circumference of the train wheel W. Figure 3 shows a diagram and underneath several plots illustrating a train wheel W rolling across several sensor units U 1 , U2, U3, U4 of a load measuring station 3, each sensor unit U1 , U2, U3, U4 producing one of a series of vertical-load time-series measurements P1 , P2, P3, P4, respectively. The sensor units U 1 , U2, U3, U4 each comprise vertical load sensors, as exemplarily shown above in Figure 2A and 2B. The sensor units U 1 , U2, U3, U4 are arranged in sequence such that the Q-Force which the train wheel W exerts on the trail is measured essentially over the entire circumference of the train wheel W. The plots showing P1 , P2, P3 and P4 each have a horizontal time axis in arbitrary units and a vertical Q-Force axis, also shown in arbitrary units ranging from 0 to 1 5, as a function of the time variable T. It can be seen in the individual vertical-load time-series measurements P1 , P2, P3, P4 that a central part of the vertical-load time-series P 1 , P2, P3, P4 has been windowed. In particular, each vertical-load time-series measurement P1 , P2, P3, P4 is pre- processed to discard measurement points associated with the train wheel W not being sufficiently close to a respective sensor unit U 1 , U2, U3, U4, by windowing a central portion of each vertical-load time-series measurements P1 , P2, P3, P4 such that measurement points outside the window are discarded. It can be seen that within the window of the first plot P1 the Q-Force displays a much greater variability (as indicated by the vertical extent of the window), than in the remaining vertical-load time-series measurements P2, P3, P4. This is indicative of wear and/or damage to the train wheel W at that particular section of the outer edge of the train wheel W, which has come in contact with the rail 3 at the sensor unit U 1 . A dynamic coefficient DC of the train wheel W is computed by dividing the maximum Q-Force measured across the plurality of vertical-load time-series measurements P1 , P2, P3, P4, with the average Q-Force being measured across the plurality of vertical-load time-series measurements P 1 , P2, P3, P4. Figures 4A-4H each show a plot of the dynamic coefficient DC as a function of distance for several train wheels W. The plots are generated using historical measurement data, in particular historical dynamic coefficients DC and associated mileage data of a plurality of train wheels W. Figures 4A - 4D show plots of train wheels W which are critical in that the dynamic coefficient DC has exceeded a critical dynamic coefficient threshold at least once. Figure 4A, for exam pie, shows dynamic coefficient data of a train wheel W which is highly variable yet does not show a clearly increasing trend for many thousands of kilometres, until at a certain distance the critical dynamic coefficient threshold begins to rise dramatically. This is indicative of sudden damage occurring to the train wheel, for example during emergency braking, and the dynamic coefficient DC therefore increasing suddenly. It can be seen that the train wheels W of Figure 4B-4C, for example, have variable dynamic coefficient data with an increasing trend. This is indicative of slowly accumulating wear leading to correspondingly larger values of the dynamic coefficient DC. Figure. 4D shows a similar behaviour of increasing wear, however with a larger scatter of the Q-Force measurement data.

Figures 4E - 4H show plots of train wheels W which are non-critical in that the dynamic coefficient DC has not yet exceeded the critical dynamic coefficient threshold, nor is there a trend which would indicate that the critical dynamic coefficient threshold is predicted to be exceeded at a determinate point in the future. Figure 4E - 4G, for example, show dynamic coefficient data of train wheels W which have very low variability in the dynamic coefficient DC and remain approximately constant as a function of distance. Figure 4H, for example, shows dynamic coefficient data of a train wheel W which has a variable trend. Figure 5 shows a block diagram illustrating schematically a computer 1 for maintenance planning for a train wheel W. The computer 1 comprises an electronic circuit, including a processor 1 1 , a memory 1 2, and a communication interface 13. Depending on the embodiment, the computer 1 can have a human-machine interface (HMI), comprising output means and/or input means. The output means can comprise a display (touch or non-touch) and/or a loudspeaker. The input means can comprise a touch interface, keyboard, mouse, pen, etc. The computer 1 can be embodied as a laptop computer, a desktop computer, a server computer, and/or a computer system comprising several connected computers. In an embodiment, the computer 1 is connected to a cloud-based computing system 14 which provides computing resources, including processing and storage capabilities, to the computer 1 . The person skilled in the art is aware that some or all of the functionality described in relation to the computer 1 can be, depending on the embodiment, provided by the cloud-based computing system 14. The cloud-based computing system 14 is particularly well suited for providing a large memory for storing data, for example in a database.

The processor 1 1 comprises a central processing unit (CPU) for executing computer program code stored in the memory. The processor 1 1 , in an example, can include more specific processing units such as application specific integrated circuits (ASICs), reprogrammable processing units such as field programmable gate arrays (FPGAs), or processing units specifically configured to accelerate certain applications. The memory 1 2 can comprise one or more volatile (transitory) and or non-volatile (non-transitory) storage components. The storage components can be removable and/or non-removable, and can also be integrated, in whole or in part with the computer 1 . Examples of storage components include RAM (Random Access Memory), flash memory, hard disks, data memory, and/or other data stores. Depending on the embodiment, the memory 1 2 comprises a database, which database may be implemented locally on the computer 1 itself or remotely, for example on a cloud-based computer system. The memory 1 2 has stored thereon computer program code configured to control the processor 1 1 of the computer 1 , such that the computer 1 , performs one or more method steps and/or functions as described herein. Depending on the embodiment, the computer program code is compiled or non-compiled program logic and/or machine code. As such, the computer 1 is configured to perform one or more method steps and/or functions. The computer program code defines and/or is part of a discrete software application. One skilled in the art will understand, that the computer program code can also be distributed across a plurality of software applications, which software applications can be distributed and executed on a plurality of computers 1 . The software application is installed in the computer 1 . Alternatively, the computer program code can also be retrieved and executed by the computer 1 on demand. In an embodiment, the computer program code further provides interfaces, such as APIs (Application Programming Interfaces), such that functionality and/or data of the computer 1 can be accessed remotely, such as via a client application or via a web browser. In an embodiment, the computer program code is configured such that one or more method steps and/or functions are not performed in the computer 1 , but in an external computing system or computing device, for example a mobile phone, and/or a remote server at a different location to the computer 1 , for example in the cloud-based computing system 14.

The various components of the computer 1 are interconnected using a connection line. The term connection line relates to means which facilitate power transmission and/or data communication between two or more modules, devices, systems, or other entities. The connection line can be a wired connection across a cable or system bus, or a wireless connection using direct or indirect wireless transmissions. Furthermore, the computer 1 is connected to one or more networks, including local networks such as a local area network, and other networks, such as the Internet, using a communication interface 13. The communication interface 1 3 is configured to facilitate wired and or wireless data transmission between the computer 1 and one or more further computers, computer systems, in particular the cloud-based computing system 14 either directly or via an intermediary network. In particular, the communication interface 13 is configured to receive the measurement data from the load measuring station 2 (not shown).

Figure 6 shows a flow diagram illustrating an exemplary sequence of steps for maintenance planning for a train wheel W. In a preparatory step SO (not shown), the computer 1 receives measurement data from the load measuring station 2 (not shown). The load measuring station 2 measures the train wheel W, in particular the vertical load force (Q- Force) which the train wheel W exerts on a rail 3 as the train wheel W rolls along the rail 3. The vertical load force (Q-Force) data, in particular one or more vertical load time-series measurements P 1 , P2, P3, P4, may be pre-processed and used to calculate a dynamic coefficient DC as described below in the description of Figure 7. The measurement data comprises at least one measurement of the dynamic coefficient DC of the train wheel W. The measurement data is stored locally in the memory 1 2, or stored remotely, for example in a database of the cloud-based computing system 14. In step S1 , the computer 1 , in particular the processor 1 1 , receives measurement data of the train wheel W. The measurement data is received either from the memory 1 2 of the computer or is received from a remote database, for example implemented on a cloud- based computing system 14. In step S2, the computer 1 , in particular the processor 1 1 , determines, using the measurement data of the train wheel W, whether the train wheel W will need to undergo maintenance. The processor 1 1 determines, using the measurement data, in particular the dynamic coefficient DC data, if the train wheel W has been worn and/or damaged sufficiently to require maintenance and/or replacement.

In an embodiment, the processor 1 1 receives, in addition to the measurement data, historical measurement data of the train wheel W, which historical measurement data comprise past measurement data of the train wheel W. The processor 1 1 is configured to determine whether the train wheel W will need to undergo maintenance now or at some determinate time-point in the future.

In step S3, the computer 1 , in particular the processor 1 1 , predicts a maintenance point M1 of the train wheel W. The maintenance point M 1 is defined as a critical time T1 and/or a critical distance D1 until maintenance. In particular, the maintenance point M 1 is the end point of a maintenance interval extending from the present distance which the train wheel W has covered and/or the present day up to the maintenance point M 1 , the maintenance interval being an interval in which the train wheel is to undergo maintenance. Maintenance includes manual inspection, and/or repair (in particular reprofiling), and/or replacement. The predicted maintenance point M 1 can be recorded to the memory 1 2 and/or can be transmitted, by the processor(s) 1 1 , via the communication interface 13 to the cloud- based computing system 14.

Figure 7 shows a flow diagram illustrating an exemplary sequence of steps for calculating the dynamic coefficient DC for the train wheel W. In step S4, the computer 1 , in particular the processor 1 1 , receives vertical load data from the load measuring station 2 (not shown). The vertical load data comprises a plurality of vertical load time-series P 1 , P2, P3, P4, each measuring the vertical load of the train wheel W as it rolls over a particular section of the rail 3 as described above. In step S5, the processor 1 1 identifies a stable measurement time range in each vertical load time series P1 , P2, P3, P4, the stable measurement time range corresponding to a time range in which the train wheel W is bearing down directly on the particular sensor unit U 1 , U2, U3, U4. The processor 1 1 can identify the stable time range using a vertical force threshold which corresponds approximately to the weight of a train wheel W. The processor 1 1 can also identify the stable time range using an end point of a steep rise in the vertical load time series P 1 , P2, P3, P4 and using a start point of a steep fall in the vertical load time series P1 , P2, P3, P4. In step S6, the processor 1 1 pre-processes the vertical load time series P1 , P2, P3, P4 by removing those data points in the vertical load time series P 1 , P2, P3, P4 which do not fall within the stable time range, for example using a rectangular window function as shown in Figure 3. In step S7, the processor 1 1 calculates the dynamic coefficient DC using following relation:

wherein the max. (QForce) is a maximum value of the vertical load of the train wheel W determined by taking into account all the vertical load time-series measurements P1 , P2, P3, P4, and the static ( QForce ) is an average value of the vertical load of the train wheel W determined across the vertical-load time-series measurements P1 , P2, P3, P4 (taking into account potential differences due to calibration). The dynamic coefficient DC is approximately equal to one for train wheels W which do not have any wear or damage, as the train wheel W rolls smoothly across the rail 3 with constant vertical force. If the train wheel is not well balanced, this may also result in a non-constant vertical force. As the train wheel W accumulates wear and/or damage, the resulting surface or other imperfections result in the train wheel W no longer rolling smoothly across the rail. Instead, the surface imperfections result in small vertical displacements of the train wheel W relative to the rail 3, which are measured as corresponding variations in the vertical load of the train wheel W on the rail. The dynamic coefficient DC is a dimensionless value which represents a measure of wear and/or damage of the train wheel W. The dynamic coefficient DC an depend on the type of train wheel W and/or the bogie and/or the train carriage the wheel is affixed to; maintenance can typically be required once the dynamic coefficient DC of the train wheel W exceeds a particular critical dynamic coefficient threshold. However, in principle a critical dynamic coefficient could also be defined in such a manner that maintenance is required only after exceeding the critical dynamic coefficient more than once.

Figure 8 shows a flow diagram illustrating an exemplary sequence of steps for receiving additional data and generating the dynamic coefficient data. In step S8, the processor 1 1 receives recorded historical measurement data of the train wheel W either from the memory 1 2 and/or from the cloud-based computing system 14. In particular, the processor 1 1 receives past measurements of the dynamic coefficient DC, each past measurement being associated with a particular point in time. In step S9, the processor 1 1 receives a last maintenance point. The last maintenance point indicates a time of last maintenance and/or a distance since last maintenance DO, i.e. how long it has been, or how far the train wheel W has travelled since last being reprofiled or replaced. The processor 1 1 receives the last maintenance point M0 either from the memory 1 2 or from the cloud-based computing system 14. In step s10, the processor 1 1 receives recorded train mileage data. The train mileage data indicates the distance which the train onto which the train wheel W is affixed has travelled in the past, in particular it comprises train mileage time-series data relating a distance covered to a particular time-interval, i.e. how many kilometres the train travelled on each of a set of days.

In step 10, the processor 1 1 generates, using the recorded historical measurement data, the last maintenance point and the time of last reprofiling, measurement data time-series of the train wheel W since the last maintenance point. In particular, the processor 1 1 generates the dynamic coefficient data which associate the dynamic coefficient DC of the train wheel W to the distance covered by the train wheel W since the last maintenance point.

Figure 9 shows a flow diagram illustrating an exemplary sequence of method steps for pre- processing the dynamic coefficient data. Depending on the embodiment, these method steps may be used in conjunction with, or instead of, the method steps as described above in relation to Figure 8. In step S1 2, the processor 1 2 sanitizes the dynamic coefficient data, in this case comprising a dynamic coefficient time-series of the train wheel W, by removing outlier data points. Outlier data points are detected and removed by the processor 1 1 , for example by detecting whether they exceed a particular dynamic coefficient value, or whether they lie away by a particular number of standard deviations from the mean dynamic coefficient of the dynamic coefficient time-series. In particular, three types of outlier data points can be identified. The first type relates to a dynamic coefficient data point above the critical dynamic coefficient threshold which has a value exceeding a previous dynamic coefficient data point value by 0.3 or larger. The second type relates to a dynamic coefficient data point which has a much higher value, for example over twice as high, as the median value of the data points in a surrounding interval. The third type relates to a dynamic coefficient data point which exceeds the critical dynamic coefficient threshold and which lies within a small interval of data points, e.g. 5-10 data points, and outside the small interval only a limited number, for example four, of dynamic coefficient data points exist which exceeds a defined intermediate dynamic coefficient threshold lying between 1 and the critical dynamic coefficient threshold.

In an embodiment, the processor is configured to identify a dynamic coefficient data point as an outlier data point if the dynamic coefficient data point matches one or more types of outlier data points as described above. The processor can delete the outlier data points or can replace the outlier data point with an interpolated value.

In step S13, the processor 1 1 generates, using the dynamic coefficient time-series and recorded train mileage time-series data, dynamic coefficient vs distance data relating the dynamic coefficient DC to the distance travelled by the train wheel W. In step S14, the processor 1 1 determines a distance since last maintenance DO. In an example, the distance since last maintenance DO is detected by the processor 1 1 by determining the last maintenance DO as a discontinuity in the dynamic coefficient data DC, in particular in a trend of the dynamic coefficient DC in which the dynamic coefficient DC orthe trend of the dynamic coefficient DC falls, in particular falls to a value close to one. In another example, the distance since last maintenance is not detected but retrieved, for example as described above in step S9 of Figure 8. In step S14, the processor 1 1 uses the distance since last maintenance DO to split the dynamic coefficient data, retaining for further processing only those dynamic coefficient data which belong to the distance travelled since last maintenance DO to the present, as only these recent measurement data are relevant for predicting a next maintenance point M 1 .

Figure 10 shows a flow diagram illustrating an exemplary sequence of method steps or method elements for predicting a maintenance point M1 . The processor 1 1 , using dynamic coefficient data, predicts the maintenance point M 1 , comprising a critical time T1 until maintenance and/or a critical distance D1 until maintenance.

In step S1 5, the processor 1 1 extracts a plurality of features F1 , F2, F3 from the dynamic coefficient data. The dynamic coefficient data comprises dynamic coefficient vs distance data, and the processor 1 1 is configured to use a feature extraction model for extracting the features F1 , F2, F3 from the dynamic coefficient data. For example, the feature extraction model can comprise a window function to select a particular distance range within the dynamic coefficient vs distance data, the processor 1 1 being configured to extract the features F 1 , F2, F3 within the distance range as determined by the window function. Distance ranges between 25Ό00 km and 70Ό00 km show good results, with a distance range of 30Ό00 km showing particularly good results. The feature extraction window can be configured to use only a single window comprising a distance range of 30Ό00 km from the present, in particular such that only the last 30Ό00 km are used to extract the features F1 , F2, F3. The feature extraction model can be configured to shift (so- called "rolling") the window such that multiple overlapping distance ranges of equal length are analysed. Further, depending on the embodiment, the feature extraction model can use a plurality of shifted windows of different length such that features F 1 , F2, F3 present in different distance-scales are extracted by the processor. Typically, for each window of a given length, a dozen or more features F1 , F2, F3 can be extracted. For example, over 1000 features F1 , F2, F3 can be extracted. The features F1 , F2, F3 which are to be extracted can be pre-determined functions of the dynamic coefficient vs distance data, such that, for example, within a given window of given length the features F1 , F2, F3 include an average value of the dynamic coefficients DC within the distance range of the window, a minimum, a maximum, a variance, a slope, trends, etc. The features F1 , F2, F3 obtained from each window function are preferably arranged in a feature matrix, wherein the column-wise entries correspond to particular features F1 , F2, F3, and the row-wise entries correspond to windows. The features F1 , F2, F3 are used for downstream classification and/or regression.

In addition, or as an alternative, the features F1 , F2, F3 can also be determined using machine learning, in which a suitable machine learning model is used to determine the set of features F 1 , F2, F3 with the most predictive power. In an embodiment, the processor 1 1 can be configured to select a subset of features F 1 , F2, F3 with the most predictive power from a larger set of possible features using a training dataset and the machine learning model. The fraction of features F 1 , F2, F3 selected from the larger set of possible features can typically be relatively low, for example, a fraction in a range of 4% to10% has been shown to provide good generalization and predictive power for reliable determination of the maintenance point.

Depending on the embodiment, the feature extraction model can be realised in a number of different ways. For example, a machine learning model can be used. A recurrent neural network, for example a long short-term memory model (LTSM), can be used. A decision tree can be used, in particular a random forest model. Other suitable machine learning models include support vector machines, gradient boosting regressor models, and time delay neural networks (TDNN).

In step S16, the features F1 , F2, F3 are used by a classifier model to classify the train wheel W using the dynamic coefficient data into one of the following two categories: critical or non-critical. In particular, the processor 1 1 is configured to classify the train wheel W as either a critical or a non-critical train wheel W using a classifier neural network, such as a recurrent neural network (RNN), more particularly a long short-term memory (LSTM) neural network. Critical train wheels are those expected to exceed the critical dynamic coefficient threshold within a determinate time or within a determinate distance. Non- critical train wheels are those for which a time or a distance, at which the critical dynamic coefficient threshold would be predicted to be exceeded, cannot be determined. The classifier model can be realised in a number of ways, for example using logistic regression. In an embodiment, the classifier model comprises a random forest classifier model. In an embodiment, the classifier model comprises a neural network classifier model. In an embodiment, the classifier model uses gradient boosting.

In an embodiment, the classifier model is trained using machine learning. In particular, the processor 1 1 is configured to train the classifier model using machine learning and the training dataset, for example using supervised learning. The training dataset comprises dynamic coefficient data of a large number of train wheels W. The training dataset is prepared by removing outlier data points as described above. Further, the dynamic coefficient data relating to each train wheel W of the training dataset is labelled as critical or non-critical. The labelling of the data can occur manually. Critical train wheels W are those for which the critical dynamic coefficient threshold was exceeded at least once. For the critical train wheels W, the first dynamic coefficient datapoint exceeding the critical dynamic coefficient threshold is identified and all dynamic coefficients succeeding that first dynamic coefficient, i.e. lying between the first dynamic coefficient datapoint and the more recent dynamic coefficient datapoint, are removed. This results in trimmed dynamic coefficient data for the critical train wheel W such that only the latest dynamic coefficient datapoint equals or exceeds the critical dynamic coefficient threshold. For both critical and non-critical train wheels W, the features F 1 , F2, F3 are extracted for dynamic coefficient datapoints lying up to a classifier boundary. The classifier boundary is a pre-defined demarcation point, i.e. a distance or number of kilometres from the present distance the train wheel W has covered. The classifier model is trained using dynamic coefficient data points up to the classifier boundary, while ignoring more recent dynamic coefficient data points. This results in the classifier model being trained not having full possession of all the dynamic coefficient datapoints of the training set, in particular with the classifier model being trained not having the most recent dynamic coefficient data (dynamic coefficient data lying between the classifier boundary and the present). This results in the classifier model being trained such that it identifies, from previously unseen measurement data of the train wheel W, whether the train wheel W is critical or non-critical. Depending on the embodiment, the classifier boundary lies between 25k km and 70k km, preferably at 30k km. In an embodiment, the classifier model is trained by extracting features F1 , F2, F3, from a plurality of windows of a given size, for example 30k km. A stride length of between 1 k km and 30k km between each window is used. Preferably, features F1 , F2, F3 of critical train wheels W are extracted using a smaller stride length than features F 1 , F2, F3 of non- critical train wheels W. This over-sampling of critical train wheels W results in better performance of the classifier model, as critical train wheels W are underrepresented in the training set and critical train wheels W are those which must be identified most reliably.

In addition, the classifier model can be validated using a validation dataset comprising historical measurement data. In particular, the historical measurement data can be partitioned into the training dataset and the validation dataset. In an embodiment where the classifier model is embodied as a neural network classifier model, the training dataset is used to train parameters (e.g. weights and biases) of the classifier model itself using optimization methods such as gradient descent and the validation set is used to tune hyper parameters relating to the specific architecture of the classifier model itself (e.g. number of layers, connections between layers, and activation functions used), and how the classifier model is trained (e.g. the learning rate). The classifier model thus trained demonstrates greater precision and recall when used to classify the train wheel W using the measurement data, than a baseline model, the baseline model being configured to linearly extrapolate the measurement data and identify whether the linear extrapolation will exceed the critical dynamic coefficient threshold or not. In particular, for a classifier model trained with a classifier boundary of 30k km, the classifier model is typically less precise than the baseline model, with a value of 0.339 vs. the baseline model which has a value of 0.572. The recall of the classifier model, however, is significantly improved (0.684 vs. 0.425) vs. the baseline model. Precision is defined asthe number of true positives divided by the sum of the number of true positives plus the number of false negatives, i.e. the number of train wheels W identified as critical by the classifier model which are actually critical, divided by the sum of the number of train wheels W identified as critical by the classifier model which are actually critical plus the number of train wheels W identified as critical by the classifier model which are actually non-critical. Recall is defined as the number of true positives divided by the sum of the number of the true positives plus. The classifier model strikes an advantageous balance between precision, meaning the ability of the classifier model to identify only critical train wheels W, and recall, meaning the ability of the classifier model to identify all critical train wheels W. This is because it is more important for the classifier model to identify as critical those train wheels W which are critical than to not identify as critical those train wheels W which are not critical.

The balance and trade-off between precision and recall can be expressed as the F1 score, defined as the harmonic mean of precision and recall, namely twice the product of precision and recall divided by the sum of precision and recall. The classifier model is trained, using the training dataset, to maximise the F1 score. In particular, the processor is configured to adjust parameters of the classifier model, comprising weights and biases, such that the F1 score is maximised in the training dataset.

In an embodiment, a plurality of classifier models are used for a corresponding plurality of types of train. Each classifier model is trained using historical measurement data of a particular type of train. As a result, the values for precision, recall, and the F1 score of the classifier model may be different depending on the type of train. In particular, depending on the type of train, the classifier model can achieve a recall of up to 0.889 and a precision of up to 0.774, and an F1 score of up to 0.993. More specifically, for the RABDe 500 train type, the classifier model achieved an F1 score of 0.863, for the RABe 51 1 train type the classifier model achieved an F1 score of 0.71 5, and for the ETR 610 train type the classifier model achieved an F1 score of 0.933.

In step S17, the features F 1 , F2, F3 are used to generate forecasted dynamic coefficients of the train wheel W. In particular, for those train wheels W classified as critical, the processor 1 1 is configured to generate forecasted dynamic coefficients as a function of distance by using a regression model. The forecasted dynamic coefficients are depicted by a dashed line in Figure 10. In an embodiment, the regression model uses a random forest regression model. In an embodiment, the regression model comprises a neural network regression model, for example a recurrent neural network (RNN). The regression model can be trained using machine learning and the training dataset. The regression model is trained using only those train wheels W of the training dataset which were labelled as critical.

The performance of the regression model is compared to the baseline model. As described above, the baseline model uses linear regression, and the dynamic coefficient data of the baseline model is extrapolated, for those train wheels W which are critical, to identify a distance in the future at which the extrapolated dynamic coefficient datapoints exceed the critical dynamic coefficient threshold. The root mean squared error (RSME) regression model is compared to the baseline model for critical dynamic coefficient values of 1 .4, 1 .6 and 1 .8. Selecting a critical dynamic coefficient value of 1 .8 resulted in a regression model with a RSME value of 10k km vs.453k km for the baseline model. It is found that the RSME value varies for different train types, in particular for the RABDe 500 train type, the regression model achieves an RSME value of 9.7 kilometers, for the RABe 51 1 train type 1 2k km, and for the ETR 610 train type 5.9k km. However, forthe RABDe 500 train type, a critical dynamic coefficient value of 1 .6 proves particularly advantageous with a corresponding RSME value of 6.8k km.

In step S18, the forecasted dynamic coefficients are used by the processor 1 1 to predict the maintenance point M 1 , comprising the critical distance D1 until maintenance and/or the critical time T1 until maintenance. The critical distance D1 until maintenance D1 can be determined by the processor 1 1 as the distance at which the forecasted dynamic coefficients exceed the critical dynamic coefficient threshold for the first time, or alternatively for a predefined determinate number of times. The critical time T1 can be determined by the processor 1 1 as a time, for example a time point in the future or a time interval extending from the present into the future. The processor 1 1 can store the critical distance D1 and/or the critical time T1 in the memory 1 2 and/or can transmit the critical distance D1 and/or the critical time T1 to the cloud-based computing system 14.

Figure 1 1 shows a block diagram illustrating a maintenance system 6 for maintenance of the train wheel W. The maintenance system 6 comprises a workshop 61 , the workshop 61 comprising means for reprofiling the train wheel W, for example including a lathe. The workshop 61 may further comprise means for replacing the train wheel W. The maintenance system 6 can comprise a computing device 62, comprising a processor and a memory and a display 63. The computing device 62 can be, for example, a smart-phone, laptop computer, or tablet computer. The maintenance system 6 can further comprise an RFID reader 64. The RFID reader 64 can be implemented as part of the computing system 62. The RFID reader 64 can be configured to read an RFID tag, the RFID tag being affixed to the train wheel W, wheel bogie, or train carriage, or other. The RFID tag has stored thereon a unique RFID identifier associated with the train wheel W.

Figure 12 shows a flow diagram illustrating an exemplary sequence of method steps for maintenance of the train wheel W. The train carriage on which the train wheel W is affixed enters a carriage workshop in which inspection and maintenance of the train carriage is performed. In step S19, the maintenance system 6, in particularthe computing device 62, receives a maintenance request message. The maintenance request message is generated by the computer 1 and/orthe cloud-based computing system 14 if the dynamic coefficient DC exceeds, or is predicted to exceed within a pre-defined distance and/or time, the critical dynamic coefficient threshold. The maintenance request message is transmitted from the computer 1 and/or the cloud-based computing system 14 to the computing device 62 of the maintenance system 6. The maintenance request message indicates that the train wheel W is to undergo repair, such as reprofiling and/or replacement. In step S20, the computing device 62 displays the maintenance request message on the display 63. A technician can then inspect the train wheel visually and/or by using other instruments which can perform out of roundness (OOR) measurements on the train wheel W. The technician can then reprofile and/or replace the train wheel W as required. In step S21 , the technician uses the computing device 62 to transmit a maintenance confirmation message, confirming that the train wheel W was reprofiled and/or replaced, to the computer 1 and/orthe cloud-based computing system 14. Figure 13 shows a flow diagram illustrating a sequence of method steps useful for displaying the maintenance request message of the particular train wheel W. In step S22, the computing device 62 receives the RFID identifier associated with the train wheel W from the RFID reader 64. The RFID identifier is preferably associated with only the particular train wheel W, however the RFID identifier can also be associated with a plurality of train wheels W, for example all the train wheels W of the train carriage on which the particular train wheel W is affixed. The computing device 62 then transmits a status request message to the computer 1 in step S23, the status request message comprising the RFID identifier. The computer 1 uses the received RFID identifier to retrieve the maintenance point M 1 of the one or more train wheels W associated with the RFID identifier. As described above, the maintenance request message is generated by the computer 1 , if the dynamic coefficient DC exceeds, or is predicted to exceed within a pre-defined distance and/or time, the critical dynamic coefficient threshold. In step S24, the computing device 62 receives from the computer 1 the maintenance request message only, if the maintenance point M 1 has been exceeded. Subsequently, if the maintenance request message was received, steps S20 and S21 as described above are carried out.

It should be noted that, in the description, the sequence of the steps has been presented in a specific order, one skilled in the art will understand, however, that the computer program code may be structured differently and that the order of at least some of the steps could be altered, without deviating from the scope of the invention. Thus, the terms step or method step or method element are used as equivalents.

Throughout this application, critical time or critical time point or critical time interval are used as same or equivalent terms. Alike, distance or travel distance are used as equivalent terms. Furthermore, the term train wheel shall also encompass a tram wheel or in general any type of wheel of a track-bound vehicle. Furthermore, the computing device of the workshop mentioned herein can be embodied together with the computer, in particular in a single digital device or a combination of digital devices or in a cloud-based or otherwise distributed manner. Furthermore, the terms trend or trend curve are used as equivalents.

Claims

1 . A computer-implemented method for maintenance planning for a train wheel (W), the method comprising: receiving (S1 ), in a processor (1 1 ), measurement data of the train wheel (W); determining (S2), in the processor (1 1 ), using the measurement data, whether the train wheel (W) will need to undergo maintenance in the future; and predicting (S3), in the processor (1 1 ), if the train wheel (W) will need to undergo maintenance in the future, a maintenance point (M 1 ), in particular a critical time (T1 ) and/or a critical distance (D1 ) until maintenance.

2. The method according to claim 1 , wherein receiving (S1 ) the measurement data comprises receiving, in the processor (1 1 ), measurement data relating to a roundness error of the train wheel (W).

3. The method according to any one of the preceding claims, wherein receiving (S1 ) the measurement data comprises receiving (S8), in the processor ( 1 1 ), recorded historical measurement data of the train wheel (W).

4. The method according to any one of the preceding claims, wherein receiving (S1 ) the measurement data further comprises receiving (S9), in the processor (1 1 ), a previous maintenance point (M0), in particular a time since last maintenance and/or a distance since last maintenance (DO), wherein the last maintenance (DO) comprises a repair, in particular reprofiling, or a replacement of the train wheel (W).

5. The method according to any one of the preceding claims, wherein receiving (S1 ) the measurement data comprises receiving (S4), in the processor ( 1 1 ), vertical load data of the train wheel (W) as the train wheel (W) rolls across a measuring station

(2), in particular vertical load measuring station (2), arranged at a rail (3).

6. The method according to claim 5, further comprising: the vertical load data comprising a plurality of vertical load time-series measurements (P1 , P2, P3, P4), each measurement being made by a different one of a plurality of sensor units (U1 , U2, U3, U4), which comprise vertical load sensors (L1 , L2), arranged at the load measuring station (2), identifying (S5), in the processor ( 1 1 ), from the vertical load data a stable measurement time range in each of the vertical load time-series measurements (P1 , P2, P3, P4), which corresponds to the train wheel (W) being within a stable measurement distance from a given sensor unit (U 1 , U2, U3, U4), and pre-processing (S6), in the processor (1 1 ), the vertical load data by removing, from each vertical load time-series measurement (P1 , P2, P3, P4), data points lying outside the identified stable measurement time range. 7. The method according to any one of the preceding claims, further comprising calculating (S7), in the processor (1 1 ), a dynamic coefficient (DC), which dynamic coefficient (DC) is a ratio of a maximum dynamic load to a static load according to the following relation: max. (QForce ) dynamic coefficient — - . - -,

^{J J J} staticiQForce ) wherein the max. (QForce) is a maximum value of the vertical load of the train wheel (W) determined across the plurality of vertical-load time-series measurements (P1 , P2, P3, P4), and the staticiQForce) is an average value of the vertical load of the train wheel (W) determined across the plurality of vertical load time-series measurements ( P 1 , P2, P3, P4).

8. The method according to claim 3 and any one of the claims 1 to 2 and 4 to 7, further comprising: receiving (S10), in the processor ( 1 1 ), recorded train mileage data related to the distance traveled over time by a train car to which the train wheel (W) is attached, and generating (S 1 1 ) , in the processor ( 1 1 ), using the recorded historical measurement data and the recorded train mileage data, historical dynamic coefficient data of the train wheel (W), comprising a plurality of dynamic coefficients (DC) as a function of at least one of: a plurality of corresponding measurement time points, a plurality of corresponding distances travelled overtime by the train wheel (W). 9. The method according to any one of the preceding claims, further comprising pre processing (S14), in the processor (1 1 ), the dynamic coefficient (DC) data of the train wheel (W) by removing those dynamic coefficients (DC) from the dynamic coefficient (DC) data that correspond to time points and/or travelled distances prior to a last maintenance point (MO), in particular a last maintenance time point (MO) or a last maintenance travel-distance point (DO).

10. The method according to any one of the preceding claims, further comprising: identifying, in the processor ( 1 1 ), a discontinuity time point in the dynamic coefficient (DC) data, if a difference between a particular later dynamic coefficient (DC) and a previous dynamic coefficient (DC) is negative and exceeds a pre-defined difference threshold, and pre-processing, in the processor ( 1 1 ), the dynamic coefficient (DC) data for the train wheel (W) by removing those dynamic coefficients (DC) from the dynamic coefficient (DC) data that correspond to time points and/or travelled distances prior to the discontinuity time point.

1 1 . The method according to any one of the preceding claims, wherein determining whether the train wheel (W) will need to undergo maintenance comprises: generating, in the processor ( 1 1 ), using the dynamic coefficient data and a forecasting model, forecasted dynamic coefficients, and predicting, in the processor ( 1 1 ), using the forecasted dynamic coefficients, the maintenance point (M 1 ) by determining the critical time (T1 ) and/or the critical distance (D1 ) at which the forecasted dynamic coefficients (DC) exceed a critical dynamic coefficient threshold, in particular wherein the critical dynamic coefficient threshold has a value in a range of 1 .2 to 6, preferably in a range of 1 .4 to 4, more preferably in a range of 1 .6 to 2.0, and most preferred is 1 .8. 2. The method according to claim 1 1 , wherein the forecasting model comprises at least one of: a linear regression model, a dynamic linear model (DLM), an exponential smoothing model, an ARIMA model, a dynamic linear model, a modification of any of these models, or a combination of such models. 3. The method according to any one of the preceding claims 1 1 and 1 2, wherein generating the forecasted dynamic coefficients using the forecasting model comprises: fitting, in the processor (1 1 ), a trend curve onto the dynamic coefficient (DC) data: extrapolating, in the processor ( 1 1 ), the trend curve onto future time points and/or future distances, and determining, in the processor (1 1 ), using the future time points and/or future distances, a critical time (T1 ) and/or a critical distance (D1 ), respectively, at which the extrapolated trend curve exceeds the critical dynamic coefficient (DC) threshold. 14. The method according to any one of the preceding claims 1 1 to 1 3, wherein determining whether or when the train wheel (W) will need to undergo maintenance comprises: extracting (S1 5), in the processor (1 1 ), features (F1 , F2, F3) of the dynamic coefficient data, classifying (S16), in the processor ( 1 1 ), using the features (F1 , F2, F3) and a classifier model, the train wheel (W) as a critical train wheel (W), if a dynamic coefficient (DC) of the dynamic coefficient data exceeds and/or is predicted to exceed a critical dynamic coefficient (DC) threshold within a critical time (T1 ) and/or a critical distance (D1 ), and predicting (S18), in the processor (1 1 ), for the critical train wheel, using the features (F1 , F2, F3) and a regression model, the maintenance point (M 1 ) .

1 5. The method of claim 14, wherein the classifier model comprises a classifier neural network configured to classify, using the dynamic coefficient data, the train wheel (W) as a critical train wheel or as a non-critical train wheel.

16. The method of any one of the preceding claims 1 1 to 1 5, wherein the regression model comprises a neural network regression model configured to generate, using the features (F1 , F2, F3), the maintenance point (M 1 ) comprising the critical time (T1 ) and/orthe critical distance (D1 ) until maintenance. 17. A computer (1 ) for maintenance planning for a train wheel (W), the computer (1 ) comprising a processor (1 1 ) configured to perform the method according to any one of the claims 1 to 16.

18. A maintenance system (6) for maintenance of a train wheel (W), comprising the computer ( 1 ) according to claim 17 and a workshop (61 ) for performing maintenance of the train wheel (W), such as re-profiling and/or replacing the train wheel (W), the workshop (61 ) comprising a computing device (62) configured to: receive (S19), from the computer ( 1 ), a maintenance request message, if the maintenance point (M 1 ), comprising the critical time (T1 ) and/or the critical distance (D1 ) until maintenance, for the train wheel (W) has been exceeded, display (S20), on a display (63) of the computing device (62), the maintenance request message, and transmit (S21 ), to the computer (1 ), a maintenance confirmation message, when the maintenance has been performed on the train wheel (W) in the workshop (61 ).

19. The maintenance system (6) of claim 18, further comprising an RFID reader (64) and wherein the computing device (62) is further configured to: receive (S22), from the RFID reader (64), an RFID identifier of a particular train wheel (W) of a train car which is present in the workshop (61 ), transmit (S23), to the computer ( 1 ), a status request message comprising the RFID identifier, and receive (S23), on the display (63) of the computing device (62), the maintenance request message only if the maintenance point (M 1 ) of the particular train wheel (W) has been exceeded.

20. A computer program product comprising a non-transitory computer-readable medium having stored thereon computer program code configured to control a processor ( 1 1 ) of a computer ( 1 ) such that the computer ( 1 ) performs the method according to any one of claims 1 to 16.