Scholarlink

<p>To address the bottlenecks in fusion analysis and decision support caused by strongly coupled dynamics, significant non-stationary disturbances, and cross-condition distribution shifts under high-frequency sampling conditions of multi-source flight parameters in flight test experiments, this paper proposes a structural perception fusion representation learning framework based on a dynamic flight parameter hypergraph. The method employs dual-stream gated multi-scale encoding in the temporal dimension to achieve scale decoupling and adaptive fusion, which respectively characterize transient mutations and low-frequency trends while suppressing noise-induced spurious correlations, thereby generating robust channel-level representations. In the structural dimension, flight parameter channels are modeled as nodes, enabling end-to-end learning of hypergraph associations that vary with window contents, and achieving cross-channel collaborative fusion through high-order message passing and normalized propagation across nodes, hyperedges, and nodes. A structural fingerprint is also constructed to support traceable analysis of key channel groups and collaborative mechanisms. At the geometric representation level, supervised contrastive learning constraints are introduced to compress intra-class variance and enlarge inter-class separation, enhancing metric consistency and cross-domain robustness of the embedding space, allowing the same representation to simultaneously adapt to state recognition, similar segment retrieval, and retrieval-enhanced prediction. Evaluated on complex engineering benchmarks and multiple random seeds, the proposed method outperforms sequence baselines such as LSTM and TCN in metrics such as F1 and PR-AUC. Ablation experiments further validate the critical contributions of dynamic hypergraph structural modeling and multi-scale encoding; changes in gated statistics and hyperedge allocation entropy reveal structural reconstruction patterns triggered by collaborative events. Results indicate that this framework provides a stable, reusable, and interpretable foundation for fusion representations in flight state assessment.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

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

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年4月14日

查看全文
  • Tianchang Liu,
  • Chen Zhao

摘要

To address the bottlenecks in fusion analysis and decision support caused by strongly coupled dynamics, significant non-stationary disturbances, and cross-condition distribution shifts under high-frequency sampling conditions of multi-source flight parameters in flight test experiments, this paper proposes a structural perception fusion representation learning framework based on a dynamic flight parameter hypergraph. The method employs dual-stream gated multi-scale encoding in the temporal dimension to achieve scale decoupling and adaptive fusion, which respectively characterize transient mutations and low-frequency trends while suppressing noise-induced spurious correlations, thereby generating robust channel-level representations. In the structural dimension, flight parameter channels are modeled as nodes, enabling end-to-end learning of hypergraph associations that vary with window contents, and achieving cross-channel collaborative fusion through high-order message passing and normalized propagation across nodes, hyperedges, and nodes. A structural fingerprint is also constructed to support traceable analysis of key channel groups and collaborative mechanisms. At the geometric representation level, supervised contrastive learning constraints are introduced to compress intra-class variance and enlarge inter-class separation, enhancing metric consistency and cross-domain robustness of the embedding space, allowing the same representation to simultaneously adapt to state recognition, similar segment retrieval, and retrieval-enhanced prediction. Evaluated on complex engineering benchmarks and multiple random seeds, the proposed method outperforms sequence baselines such as LSTM and TCN in metrics such as F1 and PR-AUC. Ablation experiments further validate the critical contributions of dynamic hypergraph structural modeling and multi-scale encoding; changes in gated statistics and hyperedge allocation entropy reveal structural reconstruction patterns triggered by collaborative events. Results indicate that this framework provides a stable, reusable, and interpretable foundation for fusion representations in flight state assessment.