Railway Systems: Decision basis for predictive maintenance based on a Triple Hybrid Approach
BERND LUBER, Key Researcher Digital Operations – Rail Systems
JOSEF FUCHS, Lead Researcher Digital Operations – Rail Systems
Condition-based and predictive maintenance strategies for components of railway vehicles and tracks require a reliable decision basis. Therefore, the current und future conditions of components must be determined. The trend for solving this task is moving more and more towards pure application of machine learning algorithms. But, is it enough to consider the vehicle/track-system as a black box? Our research at VIRTUAL VEHICLE has proven that system knowledge is essential for achieving an optimal decision basis. This is realised by a context-dependent combination of different approaches – the so-called “Triple Hybrid Approach” (THA).
The starting point for the development of a methodology is a component-dependent issue. Exemplary issues regarding vehicle components are:
- On-board fault diagnosis of spring and damper elements of conventional railway vehicles or
- Continuous estimation and wear condition prediction of wheel profiles of freight wagons,
as well as issues regarding track components such as:
- Prediction of future track irregularities under given operating conditions or
- Estimation of geometry and condition of switch components based on data from conventional track recording cars.
Basic methodology “Triple Hybrid Approach”
The context-dependent knowledge such as the dynamics of vehicles and tracks under varying operating conditions enables us to combine different methodologies into an overall approach. The so-called “Triple Hybrid Approach” shown in Figure 1 builds up on three critical pillars:
- Knowledge Generation – Combination of overall system simulations and real operational measurements or test facilities,
- Data Acquisition – Combination of data from on-board and wayside monitoring systems,
- Algorithm Development – Combination of model-based and data-driven approaches.
Figure 1: Triple Hybrid Approach – Context-dependent combination of different sub-approaches
A key factor for developing a methodology is knowledge generation. This means that the knowledge of the correlation between the faulty components and their effects on the dynamic behaviour of the system under different operating conditions has to be investigated. For this purpose, simulation models of vehicle and track with adequate model depth are required.
The generation of appropriate simulation scenarios requires knowledge from real operation, which is derived from measurement data from operation or test facilities. Based on this hybrid approach (simulation and measurement data), the relevant influencing factors in combination with suitable operating conditions are determined by means of sensitivity and robustness analysis. The resulting knowledge forms the well-founded basis for the development of algorithms, which is illustrated in Figure 2 for the example of a wheel profile classification methodology.
The increasing number of continuous on-board and wayside monitoring systems provide large amount of data. Additional information from the data acquisition process needs to be gained to generate value from it. The properties of the respective sensors up to the behaviour of the measurement systems has to be understood and modelled. The data from on-board and wayside monitoring systems combined into a hybrid approach reduces the measurement uncertainties and minimize the risk of missing information. As an example, the combination of acceleration data from a switch and from an axle box of a conventional track recording car can be used to estimate the health state of the switch frog.
The development of algorithms is the key for generating a decision basis. Depending on the issue or target component, suitable model-based and data-driven methods are used or combined as a hybrid approach. Model-based methods presume physical knowledge of the monitored system, as in the case of estimation of spring and damper parameters of a rail vehicle. Established observer methods such as nonlinear Kalman Filters, but also novel methods like Sliding-Mode-Observers for robust estimations, are used if noisy measurement data in nonlinear systems occur.
For issues such as the detection of the wheel profile condition, the understanding of the physical behaviour and the resulting dynamic effects are fundamental for the algorithm, especially for the required feature selection. The algorithms themselves are based on data-driven approaches (machine learning algorithms) due to complexity of the system behaviour. Supervised learning algorithms for fault diagnosis based on regression or classification methods and unsupervised learning algorithms for anomaly detection of vehicle and track components are applied.
Figure 3 shows the results for the detection of the wheel profile condition based on the discussed approach. The results are achieved with only two on-board features calculated from lateral accelerations of the bogie and the car body. This example clearly underlines the advantages of the hybrid data acquisition approach: on-board monitoring data combined with data from conventional track recording cars provides a more robust estimation with a higher quality (Fig. 3c).
Figure 2: Investigation of physical effects to generate knowledge for machine learning algorithms.
Figure 3: Machine learning classification of new and worn profiles depending on available environment information – (a) known track irregularities, (b) unknown track irregularities, (c) track irregularities and rail profile conditions known.
Condition-based and predictive maintenance strategies require a reliable decision basis. A decisive added value for the development of such a decision basis provides the ‘Triple Hybrid Approach’ developed at VIRTUAL VEHICLE. This approach includes not only model-based and data-driven algorithms, but also knowledge generation from vehicle-track interaction simulations enhanced with measurement data from on-board and wayside monitoring systems. Novel methods will be continuously integrated into this concept making a significant contribution to the deployment of condition-based and predictive maintenance strategies.