Model incremental learning of flight dynamics enhanced by sample management

Abstract

In this study, our focus is on investigating the utilization of real-time collected flight data to enhance the accuracy of dynamical models that are traditionally dominated by offline first-principles. The necessity and practicality of data-driven model correction/learning has been recognized in an increasing number of aerospace and industrial control scenarios. However, the issue of learning stability should be specifically emphasized. Thus, we propose an online incremental learning method for dynamical models that incorporates first-principle knowledge and data-driven correction mechanisms. To the best of our knowledge, this is the first work in which the convergence conditions of closed-loop learning systems involving Gaussian Process (GP) models are provided. Moreover, we propose an online sample management algorithm to optimize the spatial and temporal distribution of dataset samples for model training, thereby improving the model's ability to fit globally on the whole sample space. Finally, we provide three simulation examples to demonstrate the effectiveness of the proposed techniques, resulting in a data-driven model incremental learning algorithm with promising potential applications in adaptive control, optimal control, and model-based reinforcement learning.