Kernel-based digital twins for evaluating the running safety of freight wagons during emergency braking

The paper defines a computational framework for the derivation of digital twins of multibody (MB) simulations. The framework integrates longitudinal train dynamics (LTD) and MB codes. The digital twins are based on kernel regressions and classifications, and they can predict the running safety of freight trains during emergency air brake operations. The digital twins include two types of models, namely scalar regressions for the evaluation of the running safety indexes and a binary classifier for the detection of unsafe running conditions. The inputs for the digital twins can be easily obtained from common LTD codes. With 400 training samples, the scalar regression models have an average R2 score around 0.8, while a basic binary classifier can correctly label 90% of data. A weighted classifier can minimise misclassification of unsafe conditions, but this comes at the cost of worsened classification of safe behaviour. Overall, good accuracy can be achieved with the advantage of a small dataset. Once trained, the evaluation of the digital twins is 6 orders of magnitude faster than launching multibody simulations. The combination of good prediction capabilities and outstanding computational times can thrust the use of kernel regressions as surrogates of MB simulations.