A health status estimation method based on interpretable neural network observer for HVs.

Area of Science:

• Aerospace Engineering
• Artificial Intelligence
• Control Systems

Background:

• Accurate health status estimation is vital for hypersonic vehicle safety and fault diagnosis.
• Conventional neural network methods suffer from data dependency and lack interpretability, hindering precise estimation.
• Challenges include ensuring accuracy and understanding model behavior in complex dynamic systems.

Purpose of the Study:

• To develop advanced health status estimation methods for hypersonic vehicles.
• To address data dependency and improve model interpretability in estimation techniques.
• To enhance the accuracy and reliability of online health monitoring.

Main Methods:

• Developed a block interpretable neural network model incorporating trajectory and attitude equations.
• Proposed unsupervised and supervised health status estimation methods based on the interpretable model.
• Introduced an FC-LN-Mish structure for the supervised approach to map fault residuals to fault parameters.

Main Results:

• The proposed methods demonstrated improved fitting to system mechanisms, enhanced model interpretability, and reduced data dependency.
• High estimation efficiency and precision were achieved, outperforming other models in low fault deviation scenarios.
• FC-LN-Mish structure effectively reduced missed and false detection rates.

Conclusions:

• Interpretable neural network observers accurately estimate health status parameters for rudders and Reaction Control Systems (RCS).
• The methods reduce data dependence and processing costs, offering superior performance under high uncertainty.
• Provides an effective approach for online health estimation in hypersonic vehicles.