Dynamics-informed causal inverse filtering for vertical track irregularity identification from multi-level vehicle accelerations

Abstract

Track irregularity affects vehicle dynamic performance, ride quality and maintenance planning, yet its identification from vehicle responses remains challenging because the geometric excitation is embedded in coupled wheel–rail interaction and suspension dynamics. This study proposes a dynamics-informed causal inverse filtering (DICIF) method for reconstructing vertical geometric irregularity from vehicle accelerations. The irregularity is formulated as a moving spatial input, while carbody, bogie, and axlebox accelerations are synchronized in the track coordinate and assembled into channel-lag histories that retain vehicle-response memory and measurement hierarchy. A regularized spatial inverse filter reconstructs the current profile, with segment-wise feature construction preventing lag connections across independent profile segments. Coupled vehicle–track simulations at 40, 60, and 80 km/h yield median RMSE values of 0.449, 0.173, and 0.118 mm and Pearson correlations of 0.981, 0.997, and 0.999, respectively. Relative to a gated recurrent unit model, DICIF reduces RMSE by 73.0%, 86.1%, and 89.1% at 40, 60, and 80 km/h, respectively. Response-level analysis shows that combining carbody, bogie and axlebox accelerations substantially improves reconstruction over axlebox-only input, while spatial-spectral analysis confirms recovery of the dominant profile content. The results demonstrate the feasibility of reconstructing prescribed vertical track geometry from structured multi-level vehicle responses under the evaluated simulation conditions.