面向变刚度复合材料筒壳高效屈曲分析的变保真度迁移学习模型

Variable-fidelity transfer learning model for efficient buckling analysis of variable stiffness composite cylindrical shells

相较于传统直线铺层设计的复合材料筒壳结构,变刚度复合材料筒壳结构通过曲线纤维路径铺层,可以极大地增加复合材料的设计空间,进而获得更优的抗屈曲能力。为了准确描述曲线纤维路径,需要针对变刚度复合材料筒壳建立精细有限元模型,因此对屈曲分析和优化效率带来了较大的挑战。本论文以变刚度复合材料筒壳结构的线性屈曲及后屈曲承载力快速预测为目标,提出了一种变保真度迁移学习模型的构建方法。首先,针对变刚度复合材料筒壳结构建立合适的高保真度、低保真度模型;然后,基于大量低保真度样本数据作为源域样本集建立并训练深度神经网络,得到预训练模型;最后,以少量高保真度样本数据作为目标域样本集对最后一层神经网络参数进行微调,训练得到变保真度迁移学习模型。变刚度复合材料筒壳线性屈曲和后屈曲算例结果表明,在达到相同的预测精度水平时,变保真度迁移学习模型比直接采用高保真度样本数据构建的代理模型分别节约了47.7%和62.3%的计算成本,验证了提出方法的高效率优势。同时,与基于桥函数构建的变保真度代理模型和Co-Kriging进行比较,所提出方法在不同高保真度、低保真度样本数据组合下均具有更优精度,验证了提出方法的高精度优势。

Compared with the traditional design method of composite cylindrical shells with straight fiber laminate,variable stiffness composite cylindrical shells can greatly increase the design space of composite material and thus achieve higher buckling loads by means of the curved fiber laminate.To describe the curved fiber path precisely,it is necessary to establish high-fidelity detailed finite element model for variable stiffness composite cylindrical shells.Therefore,it brings great challenges to the efficiency of buckling analysis and optimization of variable stiffness composite cylindrical shells.In this paper,a variable-fidelity transfer learning model was proposed for the fast prediction of linear buckling load and post-buckling load of variable stiffness composite cylindrical shells.Firstly,the appropriate high-fidelity model and low-fidelity model of variable stiffness composite cylindrical shells were constructed.Then,the deep neural network was established and trained with a large number of low-fidelity samples as the source dataset,and the pre-trained model was obtained.Finally,the last layer was retained by finetuning with a small number of high-fidelity samples as the target dataset,and the variable-fidelity transfer learning model was constructed after the retraining on the pre-trained model.The example results of linear buckling and post-buckling load prediction of variable stiffness composite cylindrical shells indicate that,the computational cost of variable-fidelity transfer learning model can reduce by 47.7%and 62.3%than surrogates built by the high-fidelity samples directly when achieving similar prediction accuracy,showing the advantage of high prediction efficiency of the proposed method.Besides,compared with the variable-fidelity surrogate models built by the bridge function and Co-Kriging,the proposed method shows the best prediction accuracy with different combinations of highfidelity and low-fidelity samples,which demonstrates the advantage of high prediction accuracy of the proposed method.