Inertial Navigation System Drift Reduction Using Scientific Machine Learning

Abstract

Inertial Navigation Systems (INS) are crucial for accurate navigation in GPS-denied environments, but they suffer from drift errors that accumulate over time. This thesis introduces Scientific Machine Learning (SciML) as an innovative approach to mitigate INS drift by integrating physical models with machine learning algorithms. The proposed SciML architecture leverages neural networks to learn complex error patterns and relationships from simulated IMU data, outperforming conventional techniques like Kalman filtering. Utilizing a simulation-focused approach with the Julia programming language and the HighPerformance Inertial Navigation Development Repository (HIDR) library, the research generates realistic datasets encompassing diverse trajectories, sensor errors, and operational conditions. The SciML methodology incorporates data generation, INS mechanization, error modeling using neural networks, and a filtering framework that integrates the Extended Kalman Filter (EKF) with batch filtering techniques. Experimental results demonstrate the superior performance of the SciML-based INS in reducing position, velocity, and attitude errors compared to a baseline Kalman filter. This pioneering approach of fusing SciML with INS physical models holds promise for revolutionizing drift error mitigation and advancing the field of navigation systems, paving the way for more accurate, reliable, and resilient navigation in GPS-denied environments, with potential applications in aviation, robotics, and autonomous vehicles.