Adaptive Inertial Navigation System based on Continual Learning in GPS-Denied Environments

• Ano de defesa: 2025
• Autor(a) principal: Aguiar, Ricardo Gonçalves de
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: eng
• Instituição de defesa: Biblioteca Digitais de Teses e Dissertações da USP
• Palavras-chave em Português: Aprendizado contínuo; Continual learning; Drones; Inertial navigation system; Neural networks; Redes neurais; Sistema de navegação inercial
• Link de acesso: https://www.teses.usp.br/teses/disponiveis/55/55134/tde-29102025-191417/

Artificial Neural Networks (ANNs) have emerged as a promising alternative to traditional physicsbased Inertial Navigation System (INS) models, especially in Global Navigation Satellite System (GNSS)-denied environments. Unlike conventional physical models, ANNs can operate as \"black-boxes,\" learning the relationship between inertial sensor data and estimated position without relying on detailed physical models. However, ANN-based inertial navigation systems often face accuracy challenges in real-world environments due to trajectories that are not included in the initial training dataset. While some studies have explored fine-tuning models with new trajectory data, this approach can lead to catastrophic forgetting of previously learned information. To address this, we propose an adaptive inertial navigation system based on continual learning with rehearsal, allowing the neural network to better adapt to new trajectories while retaining performance on previously learned ones. Our network was trained using drone flight log data from the PX4 Community. Compared to traditional fine-tuning methods, our approach reduces position estimation error by approximately 90%. Additionally, when evaluated against a solution from the literature that uses the same training dataset and reports an average error of 35 meters, the proposed method achieved a maximum position error of 4.71 meters during a 2-minute GNSS-denied fligh