TimesFM GAT Multisource Fusion Model for Trajectory Correction of Reentry Vehicles

Source abstract

Trajectory correction during the re-entry phase has problems due to changes in the noise level and time resolution of inertial measurement units, radar tracking, aerodynamic estimation and thermal-state indicators. A TimesFM-GAT multisource fusion model is proposed in this paper to combine foundation-model time-series embeddings with a graph attention correction layer for estimating guidance residuals of lifting reentry vehicles, with the main contribution positioned as residual correction around an existing guidance loop rather than replacement of the physical propagator. First, align the asynchronous multi-source sequences to a common state window, then extract long-context temporal descriptors through a TimesFM-style forecasting backbone, and finally propagate cross-source residual evidence via an attention graph with nodes for navigation, aerodynamics, environment and control channels. A six-degree-of-freedom simulation set with 4,800 perturbed reentry cases was employed to test cross-range, down-range, altitude and heating-rate correction. Compared with the extended Kalman filter and the temporal convolutional network/LSTM-GAT baseline, the proposed model reduced the terminal position error from 146.8m to 72.4m, increased the accuracy of peak-heating prediction by 31.6%, and kept the median inference latency at 11.8ms per correction cycle. According to ablation studies, removing the graph attention increased the down-range error by 22.7%, and substituting TimesFM embeddings increased the altitude residual error by 18.9%. Based on the above results, the foundation time-series representation and graph-based residual coupling can provide an effective correction scheme for reentry guidance under multisource uncertainty.