There is no magic in INS technology. The most fundamental aspect of any filter used in an INS is that "crap in will result in crap out." There is no getting around that.
It is important to understand that not all measurements can aid a navigation solution. Consider a navigation solution constructed from the INS data provided by a 1200kHz DVL, with good bottom lock on flat bottom, aiding the data provided by a high-grade FOG IMU. This solution will have low noise, low error, and will be highly stable. The accuracy will depend on the quality of the initial position and time-based drift. Aiding this solution with a USBL-based position may help the overall accuracy and may help manage drift if the USBL data is accurate, low noise, and most importantly, stable. If the USBL measurement in unstable or inaccurate, it can significantly degrade the final state estimate and result in a solution worse than if it was not aided by the USBL. Understanding how the state estimate works and how aiding measurements are fused is critical to understanding the expectations of an INS.
An INS can provide a navigation and positioning solution that is far superior to any solution provided by a discrete measurement from a single sensor. It can also provide a completely unusable and chaotic solution if it is not configured correctly. Understanding the application, the installation, the quality of the core inertial sensor, and the available aiding sensors is critical to configuring the INS well.
