Estimation of car body vibration from axle box acceleration by combining machine learning and model-based methods

Kurzfassung

The objective of this study was to develop a machine learning model to predict the vertical acceleration transfer between the axle box and the body of a train. The modelling was based on a data set from a project on predictive maintenance of modern railway infrastructure systems and included recordings from a on-board multi-sensor system with axle box acceleration (ABA) sensors and an inertial measurement unit (IMU). The data set was resampled, filtered, scaled, synchronised and split into training, validation and test sets in order to prepare the data for model training. Three neural network types were considered: conventional neural networks (NNs), convolutional neural networks (CNNs) and recurrent neural networks (RNNs). The models were implemented in Python and parameter optimisation was carried out, for example for the model width and model depth. The different approaches were evaluated and compared with each other. The study demonstrated that the development of a reliable prediction model is associated with considerable challenges. All model types exhibited severe underfitting during training and eventually reached a local minimum, resulting in suboptimal parameterisation and predictions that were close to zero. No parameter combination could be identified through optimisation that demonstrated superior prediction quality; instead, the configurations differed in the speed at which the plateau was reached. The suboptimal model quality can be attributed to the use of noisy and inconsistently correlated signals for training, which prevented the underlying dynamics from being accurately captured. Potential improvements include the inclusion of additional data, the investigation of multi-channel solutions with multiple sensors and the investigation of larger models with more computing power. Furthermore, modelling approaches that incorporate partial differential equations could represent promising avenues for future research.