Timothy Glinski, LSAS Tec
https://doi.org/10.64861/QZQA5129
Keywords: SSA, Data Fusion, STM, Orbit Accuracy, Calibration, Automation, Bias Models, Force Models, Covariance Realism
Abstract:
Space situational awareness systems increasingly depend on fusing observations across heterogeneous sensor networks, but that heterogeneity is what makes fusion difficult: sensors differ in precision, and their biases vary in time. An Extended Kalman Filter and smoother with tuned process noise and bias models, displaces the problem into calibration. Those models require frequent recalibration to keep both the orbit solution and its covariance credible, and that cost does not scale. Calibrating the bias model for a single sensor, or the force models for a single satellite, can take an experienced analyst hours to days. A system drawing on hundreds of sensors and tracking tens of thousands of objects therefore incurs multiple analyst-years per year simply to remain operational.
Where recalibration is infeasible, operators commonly inflate process noise, deweighting measurements well below their true precision to preserve covariance realism at the cost of orbit accuracy and of the ability to observe satellite behavior. Automating the alternative has proven difficult because the error sources are confounded: satellite dynamics, bias dynamics, sensor noise and anomalous malfunction, and error in the orbit solution itself all project into the same residuals, and reference solutions against which to separate them are sparse or absent.
We present a system that recursively calibrates both sensor bias models and satellite force models within the orbit determination process itself. Observability is assessed in two stages. Rank and conditioning of the information matrix screen candidate parameters online during recursive estimation, and empirical sensitivity of the consistency statistics to controlled parameter variation validates those determinations. Parameters failing either test are held rather than absorbed into inflated process noise, so corrections are applied only where the data supports them. The system operates with or without reference solutions, allowing it to function on new objects and newly integrated sensors where observation conditions are poor. Calibration quality is assessed throughout using McReynolds filter-smoother consistency, forward-backward filter consistency, and residual ratio statistics.
We evaluate the system on both simulated and operational data. The simulated cases validate the structural determinations empirically and establish which parameters the metrics can observe errors in. Where controlled variation of a parameter away from truth produces no response in the consistency statistics, that parameter is unobservable in the data at hand, and its estimate could be arbitrarily wrong while every system diagnostic continues to pass. The system therefore holds such parameters rather than estimating them, so they cannot drift, and labels them accordingly in the solution. This yields a per-parameter distinction between estimates the data observes and estimates whose reported uncertainty is optimistic or ungrounded, allowing an operator to be confident in the solution presented. The operational cases then demonstrate that the system reports tighter formal covariance than the uncalibrated baseline while continuing to satisfy all three consistency criteria, so the improvement reflects better models rather than optimistic noise. We report bias models, force models, and orbit solutions for a case study processed with and without the system, achieved with no operator input.
Date of Conference: September 15-18, 2026
Track: Astrodynamics