eVTOL飞行状态识别通过HITL模拟和双流注意力网络

To address the lack of realistic high-dynamic data and the low accuracy of complex flight condition identification in the energy management of hybrid-electric flying cars, this paper constructs a Human-in-the-Loop (HITL) co-simulation platform based on Unreal Engine and Simulink. By introducing stochastic gust models and obstacle avoidance resistance fields, this platform achieves high-fidelity flight data acquisition that incorporates the random operational intent of pilots. Based on this, a Multi-Dimensional Feature Dual-Stream Attention Network (MDFD-SA) is proposed. Discarding the compromises of traditional single-stream networks, this model innovatively decouples and extracts transient mutations and long-term smooth features through a 1D-CNN and BiLSTM dual-stream architecture. Furthermore, it incorporates the use of a self-attention mechanism as "dynamic routing" for the adaptive weighted fusion of cross-scale features, effectively overcoming spatio-temporal scale conflicts and long-tail effects. Experimental results demonstrate that the proposed method achieves a balanced macro-accuracy of 87.32% across five complex dynamic conditions, including emergency obstacle avoidance, thereby laying a reliable data-driven foundation for intelligent eVTOL energy management.

当前仅有来源摘要，站内完整正文尚不可用。请打开原文查看完整内容。