Ensuring Accurate Navigation Solution in GPS-Denied Scenarios with Machine Learning

Modern advancements in technology infrastructure have paved the way for widespread adoption of artificially intelligent systems, whether that be to supplement legacy systems or mimic their functionality and replace entirely. In this study we propose a machine learning system that can be leveraged to develop a navigation solution in GPS-denial scenarios. A vehicle's inertial measurement unit (IMU) produces acceleration and angular rates throughout its trajectory. This information is then used to develop the navigation solution in a software package that simulates on-board navigation for a missile system. However, the system relies on GPS input in order to perform navigation and hence the need for a supplementary machine learning agent that is capable of mapping output IMU data to the navigation solution in GPS-denial scenarios. The study leverages a dataset of approximately 500 trajectories of a generic glide vehicle with associated, model-specific IMU output data and navigation solution to train a neural network. The trained neural network has been validated against a holdout set of available navigation data in order to test performance of the model and preliminary results show promise in the approach. The successful proof of concept in this study lays the groundwork for more advanced input regimes, considering a variety of aerodynamic and environmental factors to ensure mission success in light of GPS-denial.