一种气动大差异性数据多任务学习方法

A multi-task learning method for large discrepant aerodynamic data

与传统方法相比,基于深度学习的空气动力学建模方法建模速度快、精度高。但是传统深度学习采用的全连接神经网络或卷积神经网络往往没有考虑输入数据本身的差异对预测结果的影响,而飞行器的外形特征参数和飞行状态参数在数据类型上存在较大差异。在同时使用这两种参数预测气动特性时,如果忽视这些差异性,预测结果的精度势必会损失。受到多任务学习和集群网络方法的启发,提出了一种基于多任务学习的翼型外形参数与飞行状态参数联合建模方法:大差异性多任务学习网络(LD-MTL)。该方法首先将数据集划分为多个任务,随后将整个学习网络分为多个集群,分别根据不同的任务学习所预测的气动性能相关知识,最终对每个集群所学习到的相关知识进行融合,得到预测结果。通过对比实验,证明了在进行气动大差异性数据建模时,本文提出的结构能更好地反映数据差异性对模型预测精度的影响程度,有更高的预测精度,且能对此差异性进行量化分析。

Compared to traditional aerodynamic modeling methods,the deep-learning-based aerodynamic modeling has higher modeling efficiency and precision.However,previous deep learning methods with fully connected network(FCN) or convolutional neural network(CNN) ignored the influence of input data discrepancies.The shape characteristic parameters and flow state parameters of aircraft are greatly discrepant with different types,which will cause different degrees of impacts on the predicted results.When these two types of parameters are used to predict aerodynamic characteristics together,if we ignore the data discrepancy between them,the accuracy of prediction will be lost.Inspired by the multi-task learning method,we propose LargeDiscrepancy Multi-task Learning Network(LD-MTL).Our method firstly divides the dataset into multiple tasks,and then splits the whole learning network into multiple clusters to learn relevant knowledge of the predicted aerodynamic performance according to different tasks.Finally,the relevant knowledge learned from individual clusters are then fused into the final prediction result.Through comparative tests,it is shown that our method can better reflect and quantify the influence of data discrepancies and have a higher prediction accuracy when modeling with largely discrepant aerodynamic data.