Estimating track geometry irregularities from in-service train accelerations using deep learning

Timely identification of Track Geometry Irregularities (TGIs) is essential for ensuring the safety and comfort of high-speed rail operations. Existing inspection methods rely on costly Track Recording Vehicles (TRVs) and manual trolleys, resulting in infrequent and expensive inspections. This paper proposes a data-driven approach for estimating TGIs using a Convolutional Neural Network with Multi-Head and Multi-Layer Perceptron (CNN-MH-MLP) architecture. A comprehensive vehicle-track-embankment-ground Finite Element (FE) model incorporating geometric wheel-rail nonlinearity is developed to generate the in-service train acceleration data used for training the network. The CNN-MH-MLP network demonstrates strong performance in estimating TGIs, exhibiting robustness to noise. Optimized sensor placement with three sensors achieves the best trade-off between accuracy and efficiency. Furthermore, the network's transferability highlights the significance of detailed numerical models in producing virtual databases. This work is expected to facilitate the development of intelligent systems for real-time TGI monitoring, improving inspection efficiency and reducing labor costs.