-
摘要:
随着深度学习在众多领域的成功应用与快速发展,将深度学习与传统的结构分析相融合已经成为了新的研究方向。在求解有限元单元刚度矩阵的具体问题上,研究了卷积神经网络在结构分析上的应用。以四边形平面应力单元为例,基于卷积神经网络,提出了一个求解有限元总体刚度矩阵的神经网络模型;同时分析了网络的学习效果与网络卷积核数目、训练样本数目之间的关系。计算实例表明,在一定范围内,网络的学习能力随着卷积核数目、训练样本数目的增加而不断提升。在现实应用时,可以根据具体的精度要求而设定相应的卷积神经网络。卷积神经网络训练完成后,单元刚度矩阵的计算具有实时性,且精度满足工程要求。
Abstract:With the successful application and rapid development of deep learning in many fields, the integration of deep learning with traditional structural analysis has become a new research direction. In terms of solving the finite element stiffness matrix problem, the application of convolutional neural network in structural analysis is studied. Taking the quadrilateral plane stress element as an example, based on the convolutional neural network, a neural network model for solving the finite element global stiffness matrix is proposed. Moreover, the relationship between the learning effect of the network and the number of network convolution kernels and the number of training samples is analyzed. The calculation example shows that, within a certain range, the learning ability of the network increases with the number of convolution kernels and the number of training samples. In practical applications, the corresponding convolutional neural network can be set according to specific accuracy requirements. After the convolutional network training is completed, the calculation of the element stiffness matrix is real-time, and the accuracy meets the engineering requirements.
-
表 1 平面四边形单元所需求解单元刚度矩阵元素的真实值与预测值对比
Table 1. Comparison between real and predicted values of stiffness matrix elements to be solved for a planar quadrilateral unit
单元刚度矩阵元素 真实值 预测值 k12 0.167 0.167 k13 -0.893 -0.893 k14 -0.033 -0.033 k15 -0.485 -0.485 k16 -0.167 -0.167 k17 0.408 0.407 k18 0.033 0.034 k23 0.033 0.033 k24 -0.276 -0.276 k25 -0.167 -0.167 k26 -0.241 -0.241 k27 -0.033 -0.033 k28 0.035 0.034 k34 -0.167 -0.167 k35 0.408 0.409 k36 -0.033 -0.033 k37 -0.485 -0.486 k38 0.167 0.167 k45 0.033 0.033 k46 0.035 0.036 k47 0.167 0.167 k48 -0.241 -0.242 k56 0.167 0.167 k57 -0.893 -0.894 k58 -0.033 -0.033 k67 0.033 0.033 k68 -0.276 -0.276 k78 -0.167 -0.167 注:真实值的输入为经过平移和归一化操作之后的平面四边形单元的4个节点坐标值,分别对应:x1=-0.387, y1=-1;x2=0.387, y2=-1;x3=0.387, y3=1;x4=-0.387, y4=1。 -
[1] GOLOVKO V A.Deep learning:An overview and main paradigms[J].Optical Memory and Neural Networks, 2017, 26(1):1-17. http://cn.bing.com/academic/profile?id=269ac3d9f6001d015e05a978bb2983a5&encoded=0&v=paper_preview&mkt=zh-cn [2] 孙道恒, 罗萍平, 胡俏, 等.有限元单刚矩阵计算的神经网络法[J].东北大学学报(自然科学版), 1997, 18(4):431-434. doi: 10.3321/j.issn:1005-3026.1997.04.005SUN D H, LUO P P, HU Q, et al.Neural network method for finite element single stiffness matrix calculation[J].Journal of Northeastern University(Natural Science), 1997, 18(4):431-434(in Chinese). doi: 10.3321/j.issn:1005-3026.1997.04.005 [3] 孙道恒, 胡俏.弹性力学的实时神经计算原理与数值仿真[J].力学学报, 1998, 30(3):348-353. doi: 10.3321/j.issn:0459-1879.1998.03.012SUN D H, HU Q.The principle and numerical simulation of real-time neural computing for elastic mechanics[J].Chinese Journal of Theoretical and Applied Mechanics, 1998, 30(3):348-353(in Chinese). doi: 10.3321/j.issn:0459-1879.1998.03.012 [4] 丁耀武.线弹性结构静动态有限元法[M].沈阳:东北工学院出版社, 1990:31-54.DING Y W.Static and dynamic FEM for linear elastic structural analysis[M].Shenyang:Northeastern University Press, 1990:31-54(in Chinese). [5] 蒋维城.固体力学有限元分析[M].北京:北京理工大学出版社, 1989:308-312.JIANG W C.Finite element analysis of solid mechanics[M].Beijing:Beijing Institute of Technology Press, 1989:308-312(in Chinese). [6] 姜晋庆, 张铎.结构弹塑性有限元分析法[M].北京:宇航出版社, 1990:1-10.JIANG J Q, ZHANG D.Structural elastoplastic finite element analysis method[M].Beijing:Aerospace Publishing House, 1990:1-10(in Chinese). [7] 曾攀.材料的概率疲劳损伤特性及现代结构分析原理[M].北京:科学技术文献出版社, 1993:140-141.ZEN P.Probabilistic fatigue damage properties of materials and modern structural analysis principle[M].Beijing:Science and Technology Literature Press, 1933:140-141(in Chinese). [8] YOSHIOKA T, KARITA S, NAKATANI T.Far-field speech recognition using CNN-DNN-HMM with convolution in time[C]//IEEE International Conference on Acoustics, Speech and Signal Processing.Piscataway, NJ: IEEE Press, 2015: 15361744. [9] YOSHIOKA T, OHNISHI K, FANG F, et al.Noise robust speech recognition using recent developments in neural networks for computer vision[C]//IEEE International Conference on Acoustics, Speech and Signal Processing.Piscataway, NJ: IEEE Press, 2016: 16021176. [10] GLOROT X, BENGIO Y.Understanding the difficulty of training deep feedforward neural networks[C]//Proceedings of the Thirteenth International Conference on Artificial Intelligence and Statistics, 2010: 249-256. [11] ZHANG X, ZOU Y, SHI W.Dilated convolution neural network with LeakyReLU for environmental sound classification[C]//International Conference on Digital Signal Processing.Piscataway, NJ: IEEE Press, 2017: 1-5. [12] ZHAO P, ZHANG T.Accelerating minibatch stochastic gradient descent using stratified sampling[EB/OL].(2014-05-13)[2019-03-28].https://arxiv.org/abs/1405.3080. [13] NG A Y.Feature selection, L1 vs.L2 regularization, and rotational invariance[C]//Proceedings of the Twenty-first International Conference on Machine Learning.New York: ACM, 2004: 78. [14] KINGMA D, BA J.Adam: A method for stochastic optimization[EB/OL].(2014-12-22)[2019-03-28].https://arxiv.org/abs/1412.6980. [15] 陈国荣.有限单元法原理及应用[M].北京:科学出版社, 2016:91-116.CHEN G R.Principle and application of finite element method[M].Beijing:Science Press, 2016:91-116(in Chinese).