关键词: 应变载荷模型, 飞行载荷实测, 飞机结构, 多步注意力机制, 神经网络模型

Abstract: Establishing an accurate, reliable and robust strain-load model is of great significance to the advanced development of aircraft structures, such as in-flight load measurement and aircraft health monitoring. However, existing models neither weight the importance of each strain bridge onto the load prediction nor consider the varying importance among various load conditions, which leads to limited accuracy and robustness. To address this problem, this paper proposes to embed multi-step attention mechanism into neural network-based strain-load model. The multi-step attention mechanism realizes adaptive identification and weight assignment for each strain bridge, which is validated by the data of a dual-beam wing structure in a ground load calibration test. Compared with the multiple linear regression model and fully connected neural network model, the proposed model succeeds in reducing the root mean square error of bending moment, shear force and torque prediction by more than 30%, demonstrating a great improvement for the prediction accuracy. Moreover, it also enables autonomous weight allocation and validation judgement for all strain bridges under diverse load conditions, ensuring the robustness of strain-load models in complex scenarios.

Key words: Strain-based loads model, Flight load measurement, Aircraft structure, Multi-step attention mechanism, Neural network model?