Wheel Load and Ballast Stiffness Estimation from MFC Signals Using a 1D Convolutional Neural Network

Continuous monitoring of railway-track support conditions remains challenging because the measured rail response is governed by coupled vehicle-track interaction, local support properties, and operational variability. In particular, ballast stiffness strongly influences wheel-rail forces, track deterioration, and long-term maintenance demand, yet it is difficult to estimate directly under real traffic conditions. This paper presents a framework for the simultaneous estimation of wheel load and ballast stiffness from voltage signals generated by a Macro Fiber Composite (MFC) sensor using a one-dimensional convolutional neural network. The network is trained on synthetic data produced by a physics-based two-layer discrete railway-track model, in which the rail is represented as an Euler-Bernoulli beam on discrete viscoelastic supports. The model includes a tensionless-foundation correction and a simplified sensor-response model to transform the simulated rail response into voltage signals corresponding to the MFC measurement. The simulated signals are transformed into the spatial domain and used together with train speed as inputs to the neural network. The trained model achieved good accuracy on the synthetic test set and was further assessed using long-term measurements from a monitored railway line in the Czech Republic. Although the estimates obtained from measured data were affected by unmodelled dynamic effects associated with local discontinuities and defects, the method still captured stable load levels and meaningful relative changes in inferred stiffness over time. The results indicate that the proposed framework is promising primarily as a condition-monitoring and trend-detection tool rather than as a method for precise identification of absolute physical parameters under complex field conditions.