Session C4: Positioning Technologies and Machine Learning

Artificial Intelligence and Machine Learning for Inertial Measurement Unit Noise Estimation and Denoising
Andrew Isaacson and Garrett Payne, Safran Federal Systems
Date/Time: Thursday, Sep. 19, 2:12 p.m.

The growing demand for position, navigation, and timing (PNT) products has led to better development of inertial navigation systems (INS), incorporating advanced and novel technologies to compete at the cutting edge of innovation. The employment of machine learning (ML) artificial intelligence (AI) is no different. Much research has been conducted on how to apply ML and AI to navigation technologies, ranging from applying AI/ML to data fusion to using AI/ML for navigation error estimation. One specific topic of interest for AI/ML is for noise estimation of Inertial Measurement Units (IMUs), where noise can significantly affect the integration of acceleration and gyroscopic values for positioning. Safran Federal Systems (SFS) is evaluating applying ML and AI methods for calibration of gyroscope and/or acceleration measurements and error modeling. SFS is evaluating these methods for application to current navigation hardware systems that include MEMS-based IMUs. MEMS based IMUs typically drift overtime due to sensor noise and biases that cannot be easily modelled analytically. However, ML and AI have been applied to problems that have been historically hard to solve with noteworthy improvements. SFS has begun implementing ML/AI methods based upon Neural Networks (NNs) including Light Gradient Boosting Machines (LGBMs), Gated Recurrent Units (GRUs), Recurrent Neural Networks (RNNs), Long Short-Term Memory (LSTM) models, and k-nearest neighbor (kNN) models for IMU noise estimation and denoising. The outputs from these models can also be further denoised through tuning and using Convolutional Neural Networks (CNNs). The denoised IMU measurements are then intended to be included in filtering methods such as Kalman Filters, where different versions such as the Extended Kalman Filter (EKF) and Invariant Extended Kalman Filter (IEKF) are being evaluated. The ML/AI models are being trained, tested, and evaluated with real-world collected data from test events that SFS has attended and captured data for evaluation and post-test analysis. The collected data provides a simple way to quickly adapt and test/train models using realistic data. Once results are tested and verified, these methods and models can be implemented within SFS hardware systems for improved performance. After testing the above methods, SFS can propose a novel fusion engine to contend with navigation-grade INS systems while using lower cost hardware. The ML/AI methods will be detailed and the results of these methods using simulated data will be shown and discussed.