Machine Learning for Multi-IMU Sensor Fusion
ABSTRACT
We will present an overview of our SEMI PNT Phase I program where we developed and applied machine learning (ML) models for fusion of multiple industrial-grade, MEMS gyroscopes with varying performance characteristics to obtain a single, superior gyroscope measurement. Sensor errors in MEMS devices can be highly complex and nonlinear when compared to high-grade sensors such as ring laser gyroscopes making them difficult to calibrate. ML has been shown to be a powerful tool when it comes to learning correlations from data that can be difficult to capture in conventional, physics-based error modeling. Our ML-based sensor fusion algorithm fuses together multiple coincident MEMS gyros into a single gyro measurement with significant improvements over other state of art fusion algorithms in both angle random walk and bias instability across the operating temperature range of the sensors. In addition, the outputs of the ML model were augmented with a predictive uncertainty metric that quantified the relative confidence in its predictions. These improvements lead to better overall inertial navigation system performance, especially in GPS-denied environments which is a key technology need for future autonomous systems.
BIOGRAPHY
Andrew holds a B.S. in Physics and M.S. in Aerospace Eng. from the University of Minnesota and has worked for 15 years in software engineering, algorithm development, and simulation. He is the technical lead for programs applying AI, computer vision, and autonomy technologies to small UAS and UAM applications.