Tracking analytics

Tracking analytics

Understanding the wear of locomotive wheels is a key ingredient to accurate predictions of degradation and maintenance needs. Swiss Federal railways (SBB) and ABB have jointly launched a data collection and analysis activity to better understand wheel wear and the degradation of wheelsets on Re 460 locomotives.

Subscribe to ABB Review

Andrea Cortinovis, Lucas Schmid, Christian Huber ABB Traction Turgi, Switzerland,,; Robert Birke Corporate Research Baden-Daettwil, ­Switzerland

To transport passengers or cargo, a locomotive must apply a tractive force. This force can only be exerted if the wheels “grip” the rail. In other words, the tractive force that accelerates the train must be matched by adhesive forces between wheel and rail. Should the tractive force exerted rise above the adhesive limit under normal conditions, this adhesion is lost, the wheels spin without gripping the rail, and the train cannot accelerate. This phenomenon is known as wheelslip. The wheelslip is determined by multiple factors, including the weight pressing down on the wheel-rail interface, the gradient of the rail, the behavior of a locomotive driver, and ambient conditions (such as moisture, weather conditions, contamination of rail surface) limiting the tractive effort that can be utilized.

Wheelslip is not just a nuisance, but is also a cause of wear and degradation to both rail and wheel. Wear and degradation are also affected by other aspects of train-track interaction including torsional oscillations and acceleratory and braking forces.

Interaction of wheel and rail
The performance, safety, operating costs and reliability of railway operations depend to a large extent on the mechanical interaction at the interface between wheel and rail. This is described in [1]. A detailed overview of prognosis and health management is provided in the referenced article [2]. Over the past decades, numerous research projects have added to the understanding of mechanical properties of wheels and rails, the maintenance of wheel and rail profiles, the mechanical properties of bogies (rigidity, etc.) and the damping properties of rail pads.

In practice, however, the interactions between wheel and rail at the contact surface are highly complex and are significantly influenced by variable adhesion conditions, which cause interactions with the powertrain, especially for driven axles. Deeper understanding of these effects is needed to be able to optimize the drive train and adhesion control and to improve the overall system.

Swiss Federal Railways (SBB) and ABB partnered up in a co-creation activity using digitalization to gain a better understanding and new data-driven insights of wheel wear in the Re 460 fleet [3].

The 119 locomotives of the Re 460 class were originally built in the 1990s, with electrical equipment supplied by ABB [4]. The fleet has recently undergone a major refit program, implemented jointly by SBB and ABB, including the fitting of new state-of-the-art converters [5].

Among the fields of investigation included in this activity are the behavior of traction control under challenging conditions of adhesion, and the use of the sling brake, emergency braking and sanders. In connection with the Re 460 retrofit, special attention was paid to understanding the torsional oscillations of the wheel sets [6].

Collecting the data
The five class Re 460 locomotives used in this study were fitted with a railway-compatible industrial PC (IPC) reading read data from the multifunction vehicle bus (MVB), via an MVB Reader →01.

01 Collecting data on the Re460 locomotive.
01 Collecting data on the Re460 locomotive.

The MVB Reader is configured to read all relevant signals for wheel-rail contact at a sampling rate of 20 Hz, and transmit them to the IPC. Captured signals include, for example, the tractive effort, the air pressure in the main brake pipe and brake cylinders, the intensity of torsional oscillations of the axles, the rotational speed of the four driven axles and the vehicle speed.

A second data channel captures the geographic position of the locomotive every second. This data is provided by a mobile router, a component that is installed as standard in every Re 460 locomotive. The same router also provides a secure data transmission channel between the IPC and the Traction Cloud by way of the mobile network. This connection is used to transmit the collected data →02.

02 Data flow from raw data to customer benefit.
02 Data flow from raw data to customer benefit.

Data to knowledge
After the data ingestion, the Traction Cloud processes and stores the data using data pipelines. Typical tasks of data pipelines are merging of data sources (eg, offline wheel-wear data), aggregation of data, calculation of different statistics and performing data analytics. The customer benefits are finally transferred using visualizations, generated reports, data interfaces, and serve as the basis for analytics workshops in a co-creation setup.

As an example of benefit transfer, a tailor-made web portal was developed to visualize the processed data. It can provide a quick overview of the historical utilization of the locomotives. In addition, it permits the identification of unusual or special incidents or situations, so that they can be analyzed in more detail.

Beyond displaying the most important signals in the time domain, the web portal offers a selection of different views for inspecting the data. For example, two signals can be displayed in a scatter diagram (as a cloud of points) or in a two-dimensional histogram as a frequency distribution. The portal can also overlay data on a geographic map, identifying locations on which, for example, have excessive wheelslip →03.

03 Wheel-rail slip by GPS position. Low slip values are visualized in green and high slip values in red.
03 Wheel-rail slip by GPS position. Low slip values are visualized in green and high slip values in red.

The web portal offers different filtering capabilities to focus only a subset of the data. For example, the data of either one or of both bogies can be displayed. Moreover, the user can filter the data by train configuration. For example, only journeys in which the locomotive is pushing rather than pulling with a specific bogie in front.

Case study: difficult adhesion conditions
The following example illustrates how the web portal is used to better understand the locomotive’s behavior under given conditions. In →03, the wheel-rail slip is shown in color on the map for a rainy day on the route section between Sion and Geneva (Switzerland).

04 Scatter plot of wheel slip versus vehicle speed for a journey under difficult adhesion conditions.
04 Scatter plot of wheel slip versus vehicle speed for a journey under difficult adhesion conditions.

The data reveals there were difficult adhesion conditions over longer stretches of the line (visualized by the red and orange and colors highlighting increased slip conditions). The user can visualize the same journey in a scatter diagram, selecting any signal combination. For example, in →04, velocity is plotted against wheel slip. The diagram shows that the increased wheel slip on this journey occurred primarily at higher speeds. Furthermore, the data shows that the front bogie, “Bogie 1”, tends to have higher wheelslip than the following bogie. This effect can be frequently observed under difficult adhesion conditions. The traction control optimizes the overall traction of the locomotive by inducing higher wheelslip on the leading bogie, thereby conditioning the rail surface for the second bogie.

05 Comparative visualization of two trips on the same line with difficult adhesive conditions (left) and dry conditions (right).
05 Comparative visualization of two trips on the same line with difficult adhesive conditions (left) and dry conditions (right).

Another interesting comparison is shown in →05. Here the tool displays two journeys on the same railway line. The plot on the left was recorded under difficult adhesion conditions, and the plot the right under favorable conditions.

Besides detailed insights, the tool can also be used to generate quick overviews. This is done using two types of statistics: general statistics, which present overviews of locomotives, and geographic statistics, which aggregate data with respect to GPS coordinates.

The general statistics make it easy to compare data from different locomotives. It offers visualization of data and events over different time scales, such as monthly or yearly. For example, bar plots of operating hours and operating kilometers reveal basic information on deployment of different locomotives →06.

06 The tool permits statistical comparisons of the operation of the different locomotives for which data is collected.
06 The tool permits statistical comparisons of the operation of the different locomotives for which data is collected.

It is also possible to combine such data and, for example, look at different operating-speed classes.

The geostatistics are displayed as pie charts on a map, with the size of the pie showing the whole value for any given parameter, and the pie sectors showing the values per GPS coordinates. This enables the interpretation of locomotive data in relation to specific locations. It can identify, for example, areas where there is a lot of sanding →07, or where there are high torsional oscillations. It is also possible to capture correlation with weather data, for example slip caused by wet track conditions.

07 Pie charts superimposed on map showing usage frequency of sanding gear.
07 Pie charts superimposed on map showing usage frequency of sanding gear.

Wheel wear
Wheel wear is checked periodically to determine whether a wheelset needs to be replaced or re-profiled. Replacement is a tedious process, but even measurements and reprofiling of the wheel profile requires downtime (including the need to take the locomotive out of normal operation). Optimizing maintenance and wheel wear thus offer great potential for cost savings.

The aim of the data analysis on wheel wear is, on one hand, to predict wheel wear so that wheel measurements are performed only when necessary, and on the other hand, to identify the most important wheel-wear factors, and hence introduce optimizations that extend the life of the wheel.

In the performed analysis, a list of train signals that potentially influence wheel wear is created from the recorded data and from prior knowledge. These signals are then fed into a feature selection process to identify the best model inputs. First, constant signals are removed. Then the model is trained with decreasing subsets of the remaining signals. On each pass, the signal whose important score is the lowest gets removed. The reduced model is then evaluated using its determination coefficient →08. The determination coefficient states how much variance in the data is explained by the model (the higher the better, maximum 1.0). Based on the determination coefficients, the smallest group of signals representing the best compromise in terms of model performance is selected (five in the illustrated case). These signals are used as inputs to the data-driven model.

08 Determination coefficient for decreasing number of signals.
08 Determination coefficient for decreasing number of signals.

Wheel wear is often predicted using simple linear regression models based on mileage. Using the additional data and exploring different machine learning algorithms, it is possible to build a more accurate model. To give an idea of the model accuracy, →09 plots measured values (x-axis) against predicted values (y-axis). Ideally, all points should lie on the diagonal, implying that the predicted value is equal to the given value.

09 The random forest model (right) improves the mean absolute error by 22.8 percent, compared to the simple linear regression model (left).
09 The random forest model (right) improves the mean absolute error by 22.8 percent, compared to the simple linear regression model (left).

The random forest model (right) uses the five signals from →10 to improve the mean absolute error by 22.8 percent compared to the simple linear regression model (left).

10 The five signals from the selection process.
10 The five signals from the selection process.

Actionable results
Thanks to the data collected and analyzed in this project, the understanding of numerous physical interactions and the transparency of the drive train and wheel-wear mechanisms have been increased. This permitted software adjustments to be implemented in the area of the drive chain, leading to increased mileage of wheelsets.

In addition to the fields of investigation initially defined in the collaboration, it was found that the recorded data could also be used in further areas of interest. Examples of this are supporting the proof of axle fatigue strength with respect to torsional oscillations, the detection of sensor problems prior to test runs, the understanding of pantograph bounces with fixed overhead lines, and more. This shows how data can be used in a versatile way to unlock hidden potential can be exploited when the data is made available in a processed and structured form. The knowhow gathered using this project will also be used in future converter designs and converter controls.

Finally, no major project runs without a few “lessons learned”. Especially in the early phases of the project, significant effort was required until the data could be reliably recorded and well structured in a database. The data itself also presented several challenges: from consolidating very slow and very fast time series to data quality issues – not to mention avoiding the ubiquitous trap of mixing correlation with causality.

The project was also a valuable showcase for co-creation, in which the product was not simply a deliverable but in which ABB and the customer jointly set goals, shared observations and stewarded the development from beginning to end. 

[1] R. Lewis, U. Olofsson, Wheel-Rail Interface Handbook, Woodhead Publishing, 2009.
[2] P. Dersin, A. Alessi, B. Lamoureux, M. Brahimi, O. Fink, “Prognostics and health management in railways”, 27th European Safety and Reliability Conference (ESREL 2017), Portoroz, Slovenia, 18–22 June 2017.
[3] A. Cortinovis, R. Birke, L. Schmid, T. Wymann, C. Fehrenbach (2022). Radverschleißuntersuchung im Co-Creation-Ansatz: Von Rohdaten zum besseren Verständnis der Radabnutzung , ZEVrail, 2022, issue 08 (publication pending)
[4] J. Luetscher, J. Schlaepfer, The ‘Locomotive 2000’ – a new generation of high-speed, multipurpose locomotives – 99 main-line locomotives with AC propulsion for Swiss Federal Railways, ABB Review 10/1992, pp. 25 – 33.
[5] T. Hungerberger, IGBT converters extend life of Re460 locomotives, ABB Review 1/2017, pp. 65 – 69.
[6] “ABB increases the energy efficiency of the ‘Lok 2000’ for even more sustainable operation in the future”, ABB Switzerland press release,July 14, 2020, [Accessed April 11, 2022].


Contact us


Share this article

Facebook LinkedIn Twitter WhatsApp