High-Precision Autonomous Positioning of an Unmanned Locomotive Based on the Integration of a Kolmogorov — Arnold Network with an Extended Kalman Filter

the paper considers the problem of improving the accuracy of autonomous navigation-parameter estimation for an unmanned locomotive under unavailable or limited satellite positioning. A positioning algorithm based on the integration of an extended Kalman filter and a Kolmogorov — Arnold network is proposed. The extended Kalman filter estimates the travelled distance and velocity using a nonlinear model of longitudinal locomotive motion and measurements from autonomous sensors. The KAN module generates a predictive state estimate taking into account the track gradient and curvature. The scientific novelty consists in using the neural-network prediction to form an additional a priori estimate and in introducing a prediction-error covariance determined from validation data. The algorithm is studied by simulation under measurement noise, wheelset slip, short-term measurement outages, random longitudinal accelerations, variable gradient, and track curvature. The Railway-Precise-Localization dataset, containing precise GNSS/RTK and IMU measurements recorded onboard a railway vehicle, is used to form a reference railway trajectory. If standard odometer and tachometer channels are absent from the selected data, their measurements are generated by adding prescribed errors to the reference distance and velocity. The results demonstrate a reduction in position and velocity root-mean-square errors compared with a conventional extended Kalman filter. The results characterize the performance of the method within the domain represented in the training data and require additional validation on independent routes with different dynamic and geometrical characteristics.