Modern rail vehicles, whether locomotives or wagons, are now equipped with comprehensive built-in sensors and continuously collect a range of information about themselves and their operating environment. However, the maintenance of rail vehicles is, also due to legal requirements, by and large, still based on the decades old concepts of maintenance intervals and periodic inspections. Predictive maintenance is now often promoted as a restorative approach to change and revolutionize all this. In this article I would like to give an insight into the necessary technical, methodical and organizational steps we take to advance towards the holistic approach of predictive maintenance.
The approach begins with the (physical) acquisition of values and states of the vehicle. These can be either built-in or integrated systems from vehicle manufacturers or supplementary or retrofitted IoT systems that bring older vehicle material into Generation 4.0. Some limiting factors here are, on the one hand, the necessary certification for railway suitability, which limits the possible manufacturers and equipment suppliers and, on the other hand, the special challenge of freight cars not usually having a power supply (for sensor technology). In addition to this information about on-going operations, material and component catalogues, workshop and work information as well as estimates about the intensity of use of the vehicle and the train composition are necessary to successfully complete the following steps.

The amount of data generated per vehicle in this way varies from a few kilobytes to gigabytes per day, depending on the manufacturer, vehicle generation and sampling interval. Since a railway company operates a large number of vehicles, for example Rail Cargo Austria drives about 20,000 of its own freight cars, the amount of data is multiplied very quickly and inevitably requires a big data system in order to deal adequately with the accumulating data volume. We use our own in-house analysis platform based on Hadoop, Spark and Kafka.

The ideal case we would wish for predictive maintenance is access to the raw values of the sensor data. Although this can rapidly generate enormous amounts of data (e.g. two times 4000 measuring channels every 20 milliseconds), it allows the best possible modelling. Due to a lack of available bandwidth, however, such amounts of data can no longer be transmitted from the vehicle unprocessed. This calls for hardware-oriented developers who can use intelligent algorithms to compact and process the data and transmit it depending on the vehicle situation and available bandwidth. The last step in this cascade is the back-propagation of selected statistical models in order to enable intelligent data processing on the vehicle itself and to be able to decide which information should be transmitted when and in what density.

Another challenge for rail vehicles lies in the type of data transmission – after all, these vehicles are operated throughout Europe and Turkey. In most cases, GSM is used with SIM cards capable of roaming throughout Europe. Here, too, good developments are in demand, as continuous network coverage is inherently unavailable (tunnels) and is also more frequently railway routes run through areas with poor or non-existent network coverage. The cost structure (e.g. roaming costs in Switzerland) also requires consideration in terms of transmission behavior.

In many cases, vehicles generate alarms or status messages in their own firmware from the raw data of the sensors according to certain rules of the manufacturer, which are often not transparent to the user and which can be read out and transmitted. These individual messages are only conditionally suitable for further steps without additional processing and content enrichment. At this point, even before modelling or evaluation, an aggregation by means of rules (persistent specialist knowledge) or the derivation of systematic, complex time-linked states, is necessary in order to be able to make good predictions.
The modelling of the way in which the change in the condition of rail vehicles takes place in order to predict maintenance needs is currently a major challenge. Even if a rail vehicle resembles a car or truck (it has 4 or 8 wheels and moves on them), the dependency mechanisms of the influence quantities and patterns or developments are clearly different. Unfortunately, this means that the findings of the automotive industry can only be reused conditionally and to a very limited extent. It even goes so far that individual construction editions and model ranges – if, for example, they have been adapted for individual railway companies – can differ fundamentally and require the development of new algorithms and models even for externally identical vehicles.

The complexity is exacerbated by the fact that the vehicles are always on the move in different combinations (as trains) and that the individual vehicles can also have a positive and negative influence on each other. A final challenge in this step lies in the problem of so-called “rare events”. When predicting which models to deliver, the element to be predicted is occurring damage. However, this is very rare in the initial data due to the already good maintenance situation (thank goodness). This eliminates many of the usual statistical analysis and forecasting methods.

However, the development of predictive models alone is not enough. The predictive maintenance procedure only shows its strength if the entire maintenance process is reversed conceptually. In our case it means, as the first step, optimization by carrying out timely repairs and the preventive replacement of parts during regular maintenance, for which the vehicle has to go to the workshop anyway. A really significant effect is achieved if regular maintenance is not based rigidly on mileage or operating hours but is determined at least roughly by predictive models. In order to be able to do this, however, the legal framework conditions and company specifications must ultimately change. The scenario can then be applied to ideally predetermine and schedule the necessary workshop visits automatically in advance and to plan and automate them based on the availability of spare parts (or their delivery times), workshop capacity and the employees available and their qualifications.

Of course, the accuracy and trustworthiness of the prediction models is currently a limit. Estimates always have an element of uncertainty in the end, and especially with new, as yet not validated algorithms that have to be developed as described above, you don’t know where you, as a vehicle owner, stand at first. Therefore, it makes sense to start a parallel operation in which the machine suggestions are evaluated and, if necessary, activated by maintenance specialists. Gradually (with sufficient trust) more and increasingly extensive decisions and advance planning are left to the machine until it has reached the desired degree of automation.

The path to predictive maintenance is – as it turns out – much longer than expected and more intensive than some people would wish and yet now is the right time to begin.

Author: Gerald Schinagl, Digital Innovation Manager OEBB