Identification of uncertain model parameter in flight vehicle using physics-informed deep learning

Abstract

In a controller design process, Monte-Carlo simulations for investigating the effects of uncertainty produce big data. Due to the rapid development of machine learning and deep learning data analysis such as anomaly detection and parameter estimation using artificial intelligence is investigated in aerospace systems. However, these methods are criticized for their lack of explanation. From these perspectives, this paper proposes an estimation framework developed from Physics-informed neural networks by adding an integration-based loss. Compared to existing physical-based neural networks, the suggested method confirmed more stable performance through experiments and theories. Four types of uncertainty in a missile system, such as uncertainty in burnout time, rocket motor tilt angle, location of the center of pressure, and control fin bias are identified individually. From testing a hundred simulation data, the average and maximum estimation error is within 3 percent of the mean value of each uncertainty. From analysis with the traditional methodologies, it is proved that the proposed method has the robustness to noise. As long as structured uncertainty, the unknown parameters can be identified and this framework applies to other aerospace systems.