建立飞机结构应变载荷关系模型时，地面标定试验数据保真度高、但工况范围及数量受限，有限元仿真工况范围覆盖广、但数据保真度低，导致单独依据地面标定试验数据和有限元仿真数据建立的应变载荷关系模型难以兼顾适用范围和预测精度。对此，本文提出了映射式与补偿式两种融合“试验-仿真”虚实数据的多级神经网络架构，开发了基于子学习器方差的模型认知程度度量方法，形成了精度高、适用性广、能够预警不可靠输出结果的机翼应变载荷关系神经网络模型，并采用缩比机翼对上述模型进行了验证。本研究表明：虚实数据融合的神经网络模型能够更好地描述机翼的应变载荷关系，且补偿式模型的预测效果优于映射式模型；本文提出的模型认知程度度量方法，能够在不影响模型预测精度的前提下，有效判别出神经网络模型认知程度差的数据样本，对神经网络的不可靠输出做出预警。

When establishing a strain-load relationship model for aircraft structures, ground calibration tests can obtain high-fidelity data but are trapped with limited test ranges, while finite element simulations are not limited by test ranges but the data fidelity is low. This leads to difficulties in achieving win-win situation of accuracy and applicability based solely on either ground calibration test data or finite element simulation data. To address above issue, two multi-level neural network models fusing real data and virtual data were put forward, respectively called mapping-based model and a compensation-based model; a method for measuring the model's cognitive degree based on the variance of sub-learners is established and embedded into the compensation-based model. A neural network model with high accuracy, wide applicability, and the capability to forewarn of unreliable prediction results was therefore developed. This developed model was validated using a scaled-down wing. Compared with complete reliance on real data from ground calibration tests, the load models based on fusion of multi-source data demon-strate superior capabilities, and the compensating-based model is better than the mapping-based one. Moreover, the compensating-based model can effectively identify data samples with poor cognitive degree of the load model after embedding the base learner variance method and thereby provide warnings for unreliable prediction results.