Structure-aware fusion learning and intelligent decision support based on dynamic flight parameter hypergraphs in...

Area of Science:

• Aerospace Engineering
• Data Science
• Machine Learning

Background:

• Flight test data analysis faces challenges due to coupled dynamics, non-stationary disturbances, and distribution shifts.
• High-frequency sampling of multi-source flight parameters creates bottlenecks in fusion analysis and decision support.

Purpose of the Study:

• To propose a structural perception fusion representation learning framework using a dynamic flight parameter hypergraph.
• To overcome limitations in current fusion analysis methods for complex flight test data.
• To enhance stability, reusability, and interpretability in flight state assessment.

Main Methods:

• Employs dual-stream gated multi-scale encoding for temporal dimension analysis, decoupling scales and fusing adaptively.
• Models flight parameter channels as nodes in a hypergraph for end-to-end learning of dynamic associations.
• Utilizes high-order message passing and supervised contrastive learning for cross-channel fusion and robust geometric representation.

Main Results:

• The proposed hypergraph framework outperforms sequence baselines (LSTM, TCN) in metrics like F1 and PR-AUC on complex benchmarks.
• Ablation studies confirm the importance of dynamic hypergraph structure and multi-scale encoding.
• Analysis of gated statistics and hyperedge allocation entropy reveals structural patterns linked to collaborative events.

Conclusions:

• The developed framework offers a stable, reusable, and interpretable foundation for fusion representations in flight state assessment.
• It effectively addresses bottlenecks in analyzing complex, high-frequency flight test data.
• The method demonstrates significant improvements in data fusion and decision support capabilities.