On Performance Evaluation of Inertial Navigation Systems: The Case of Stochastic Calibration

In this work we address the problem of rigorously evaluating the performances of an inertial navigation system during its design phase in presence of multiple alternative choices. We introduce a framework based on Monte-Carlo simulations in which a standard extended Kalman filter is coupled with realistic and user-configurable noise generation mechanisms to recover a reference trajectory from noisy measurements. The evaluation of several statistical metrics of the solution, aggregated over hundreds of simulated realizations, provides reasonable estimates of the expected performances of the system in real-world conditions. This framework allows the user to make a choice between alternative setups. To show the generality of our approach, we consider an example application to the problem of stochastic calibration. Two competing stochastic modeling techniques, namely, the widely popular Allan variance linear regression, and the emerging generalised method of wavelet moments, are rigorously compared in terms of the framework’s defined metrics and in multiple scenarios. We find that the latter provides substantial advantages for certain classes of inertial sensors. Our framework allows to consider a wide range of problems related to the quantification of navigation system performances, such as the robustness of integrated navigation systems (such as INS/GNSS) with respect to outliers or other modeling imperfections. While real world experiments are essential to assess to performance of new methods, they tend to be costly and are typically unable to lead to a sufficient number of replicates to provide suitable estimates of, for example, the correctness of the estimated uncertainty. Therefore, our method can contribute in bridging the gap between these experiments and pure statistical consideration as usually found in the stochastic calibration literature.