融合注意力机制的航空发动机推力估计方法研究  

Research on aircraft engine thrust estimation method incorporating attention mechanism

作　　者：邹雨杭 赵永平[1] ZOU Yuhang;ZHAO Yongping(College of Energy and Power Engineering,Nanjing University of Aeronautics and Astronautics,Nanjing 210016,China)

机构地区：[1]南京航空航天大学能源与动力学院,南京210016

出　　处：《航空工程进展》2025年第2期40-51,共12页Advances in Aeronautical Science and Engineering

基　　金：国家科技重大专项(J2019-Ⅰ-0010-0010);中央高校基本科研业务费(NS2022027);航空发动机及燃气轮机基础科学中心项目(P2022-B-V-002-001)。

摘　　要：准确预测航空发动机推力大小对直接控制发动机推力具有重要意义。为了提升航空发动机推力估计模型的准确性和实用性,针对时间序列预测构建融合LSTM和注意力机制的多任务LSTM-Attention模型;针对不同飞行条件下推力估计的问题,运用Fine-tune和改进DANN的迁移学习方法以增强模型对多工况条件下的适应性。结果表明:LSTM融合注意力机制可以有效地对长时间序列数据进行建模,修正了LSTM在全局建模能力上不足的问题,同时也克服了注意力机制难以捕捉相对位置信息的局限;多任务学习策略能显著提高模型在油门杆突变节点处的预测精度,进一步提高了模型的准确性;当目标域数据较少时应当选择Fine-tune,而在目标域数据充足的情况下使用改进DANN方法将得到准确性更高的模型。Accurately predicting the thrust of aircraft engines is of great significance for directly controlling engine thrust.This study aims to enhance the accuracy and practicality of thrust estimation models for aero engines.The re-search first constructs a multi-task LSTM-Attention model that integrates Long Short-Term Memory(LSTM)and attention mechanisms for time series forecasting.Additionally,to address the issue of thrust estimation under different flight conditions,this paper employs Fine-tuning and an improved Domain-Adversarial Neural Network(DANN)transfer learning method to strengthen the model's adaptability to multiple operational conditions.The re-sults demonstrate that LSTM combined with the attention mechanism can effectively model long time series data,rectify LSTM's insufficiency in global modeling capabilities,while also overcome the limitation of the attention mechanism in capturing relative position information.The multi-task learning strategy can significantly improve the model's prediction accuracy at the abrupt changes in the throttle levers,further enhance the model's accuracy.The study of thrust prediction under different conditions based on transfer learning methods indicates that Fine-tuning should be selected when there is limited target domain data,while the modified DANN method will yield a model with higher accuracy when there is sufficient target domain data.

关 键 词：推力估计 时间序列预测 LSTM 注意力机制 多任务学习 迁移学习 

分 类 号：V23[航空宇航科学与技术—航空宇航推进理论与工程] TP183[自动化与计算机技术—控制理论与控制工程]