PREDICTION OF CONCRETE MESO-MODEL COMPRESSION STRESS-STRAIN CURVE BASED ON “AM-GOOGLENET + BP” COMBINED DATA-DRIVEN METHODS
-
摘要:本文结合GoogLeNet卷积神经网络和BP神经网络分别在图像数据挖掘和数据分析方面的良好性能, 采用“AM-GoogLeNet + BP”联合数据驱动方法, 对混凝土细观模型(含砂浆、骨料及孔隙)的单轴压缩应力−应变曲线进行了有效预测. 通过引入力学参量对图像数据驱动的训练结果进行优化, 从而提升了神经网络的物理可解释性. 基于Python语言实现混凝土细观模型在Abaqus中的自动建模及细观图像生成过程, 并将生成的细观图像数据库与相应的压缩应力−应变曲线作为训练数据集. 在GoogLeNet中分别引入SENet, ECANet和CBAM三种代表性注意力机制并对三种注意力机制的性能进行对比和分析, 以自适应方式提升神经网络对混凝土各相组分的分析能力, 并以此得到混凝土细观模型的初步应力−应变预测曲线; 将骨料体积分数、孔隙率及初步峰值应力等物理参量作为输入引入BP神经网络以改善峰值应力的预测精度, 并与将物理参量直接引入卷积神经网络输入层的方法进行了对比, 最后定量给出了骨料体积分数和孔隙率对峰值应力的影响权重. 结果表明, 对于不同骨料体积分数及孔隙率的混凝土细观模型, 该方法均展现了较高的预测精度. 本文采用的“AM-GoogLeNet + BP”联合数据驱动预测模型从统计角度解决了传统方法对细观尺度参量分析的复杂性, 为复合材料的跨尺度力学行为研究提供了新思路.Abstract:This paper uses the “AM-GoogLeNet + BP” combined data-driven methods to predict the uniaxial compression stress-strain curve of the concrete meso-model (including mortar, aggregates, porosity) effectively by combining the good performance of GoogLeNet convolutional neural network and BP neural network in image data mining and data analysis, respectively. The physical interpretability of the neural network is improved by introducing the mechanical parameters to optimize the image data-driven training results. The automated modeling of the concrete meso-model in Abaqus and microscopic image generation process are realized by Python language, and the generated mesoscopic image database and the corresponding compression stress–strain curves are used as the training dataset. Three typical attention mechanisms, SENet, ECANet and CBAM, are introduced into GoogLeNet respectively to enhance the analysis ability of neural network for each phase of concrete in an adaptive manner and the performance of the three attention mechanisms is compared and analyzed. The initial stress-strain prediction curves of the concrete meso-model are obtained with this method; In order to improve the prediction accuracy of the peak stress, the physical parameters such as aggregate volume fraction, porosity and initial peak stress are introduced into BP neural network as inputs. It is also compared with the method of introducing the physical parameters directly into the convolutional neural network input layer. At the same time, the weight of influence of aggregate volume fraction and porosity on peak stress is given quantitatively. The results show that this method has high prediction accuracy for the concrete meso-model with different aggregate volume fraction and porosity. In this paper, the “AM-GoogLeNet + BP” combined data-driven prediction model is used to solve the complexity of the traditional method in the analysis of mesoscale parameters from the statistical point of view, which provides a new idea for the study of the cross-scale mechanical behavior of composite materials.
-
Key words:
- concrete/
- neural network/
- data-driven/
- meso-model/
- uniaxial compression
-
表 1两种细观组分的力学参量
Table 1.Mechanical parameters of the two meso-components
Elasticity modulus
E/GPaPoisson’s ratio
υCompressive strength
Fc/MPaDensity
ρ/(t·m−3)Dilatancy angle
Ψ/(º)Eccentricity
η/%Stress ratio
σb0/σc0aggregate 43 0.23 − 2.67 − − − mortar 25 0.2 35 2.40 38 0.1 1.16 表 2试验环境的硬件和软件参数
Table 2.Hardware and software parameters of the experimental environment
Name Parameters central processing unit Inter Core i7-11800 H CPU @ 2.3 GHz memory DDR4 memory 8 GB graphics card NIVIDA GeForce RTX3060 system Windows 10 environment Python 3.6 TensorFolw 2.8.0 Keras 2.8.0 NUMPY 1.22.2 compute unified device architecture CUDA 11.2 表 3不同GoogLeNet的训练过程参数
Table 3.The training process parameters of different GoogLeNet
CNN Parameters of AM Total parameters Model size/MB Epoch Time/s Validation loss GoogLeNet − 6015577 22.95 50 6254 1.2454 SE-GoogLeNet 8352 6541145 24.95 50 6466 1.2070 ECA-GoogLeNet 5 6015582 22.95 50 6321 1.3250 CBAM-GoogLeNet 4194 6277819 23.95 50 6280 1.2053 表 4BP神经网络的权重和偏置值
Table 4.The weights and biases of BP neural network
Neurons Winput Bias Woutput Bias Peak stress Volume fraction Porosity 1 −0.00172 −0.12354 0.59741 −2.15972 −2.69027 1.21105 2 −0.17380 1.08733 −0.53564 1.61095 2.00213 3 −1.36082 0.33879 0.01926 1.44915 1.63320 4 0.33198 −1.49304 0.67236 1.67449 1.40409 5 0.03562 −0.19979 −1.14093 1.75658 1.35286 6 1.12495 −0.56972 −1.04270 1.19552 1.38380 -
[1] Tian ZY, Yan Y, Li J, et al. Progressive damage and failure analysis of three-dimensional braided composites subjected to biaxial tension and compression.Composite Structures, 2018, 185: 496-507doi:10.1016/j.compstruct.2017.11.041 [2] Yen CF, Kaste B, Chen CCT, et al. Modeling and simulation of carbon composite ballistic and blast behavior.Journal of Composite Materials, 2020, 54(4): 485-499doi:10.1177/0021998319866902 [3] 金浏, 李健, 余文轩等. 混凝土动态双轴拉压破坏准则细观数值模拟研究. 力学学报, 2022, 54(3): 800-809 (Jin Liu, Li Jian, Yu Wenxuan, et al. Mesoscopic numerical Simulation on dynamic biaxial tension compression failure criterion of concrete.Chinese Journal of Theoretical and Applied Mechanics, 2022, 54(3): 800-809 (in Chinese) [4] Zhang J, Wang ZY, Yang HW, et al. 3D meso-scale modeling of reinforcement concrete with high volume fraction of randomly distributed aggregates.Construction and Building Materials, 2018, 164: 350-361doi:10.1016/j.conbuildmat.2017.12.229 [5] Guo L, Wu J, Li JH. Complexity at Mesoscales: A common challenge in developing artificial intelligence.Engineering, 2019, 5: 924-929doi:10.1016/j.eng.2019.08.005 [6] Li X, Liu ZL, Cui SQ, et al. Predicting the effective mechanical property of heterogeneous materials by image based modeling and deep learning.Computer Methods in Applied Mechanics and Engineering, 2019, 347: 735-753doi:10.1016/j.cma.2019.01.005 [7] 瞿同明, 冯云田, 王孟琦等. 基于深度学习和细观力学的颗粒材料本构关系研究. 力学学报, 2021, 53(9): 2404-2415 (Qu Tongming, Feng Yuntian, Wang Mengqi, et al. Constitutive relations of granular materials by integrating micromechanical knowledge with deep learning.Chinese Journal of Theoretical and Applied Mechanics, 2021, 53(9): 2404-2415 (in Chinese) [8] Yang C, Kim Y, Ryu S, et al. Prediction of composite microstructure stress-strain curves using convolutional neural networks.Materials&Design, 2020, 189: 108509 [9] Seolhyun N. Analysis of gradient vanishing of RNNs and performance comparison.Information, 2021, 12(11): 442doi:10.3390/info12110442 [10] 梁雪慧, 程云译, 张瑞杰等. 基于卷积神经网络的桥梁裂缝识别和测量方法. 计算机应用, 2020, 40(4): 1056-1061 (Liang Xuehui, Cheng Yunyi, Zhang Ruijie, et al. Bridge crack classification and measurement method based on deep convolutional neural network.Journal of Computer Applications, 2020, 40(4): 1056-1061 (in Chinese) [11] Chaudhari S, Polatkan G, Ramanath R, Mithal V. An attentive survey of attention models//IEICE Transactions on Fundamentals of Electronics, Communications and Computer Sciences, 2019: 1-33 [12] Mnih V, Heess N, Graves A, et al. Recurrent models of visual attention//IEICE Transactions on Fundamentals of Electronics, Communications and Computer Sciences, 2014: 1-12 [13] 王晓玲, 李克, 张宗亮等. 耦合ALO-LSTM和特征注意力机制的土石坝渗压预测模型. 水利学报, 2022, 53(4): 403-412 (Wang Xiaoling, Li Ke, Zhang Zongliang, et al. Coupled ALO-LSTM and feature attention mechanism prediction model for seepage pressure of earth-rock dam.Journal of Hydraulic Engineering, 2022, 53(4): 403-412 (in Chinese) [14] Ye S, Li B, Li QY, et al. Deep neural network method for predicting the mechanical properties of composites.Applied Physics Letters, 2019, 115(16): 161901doi:10.1063/1.5124529 [15] 赵地, 赵莉芝, 甘永进等. 基于支撑先验与深度图像先验的无预训练磁共振图像重建方法. 物理学报, 2022, 71(5): 350-362 (Zhao Di, Zhao Lizhi, Gan Yongjin, et al. Undersampled magnetic resonance image reconstruction based on support prior and deep image prior without pre-training.Acta Physica Sinica, 2022, 71(5): 350-362 (in Chinese) [16] Hu ZL, Zhao Q, Wang J. The prediction model of cotton yarn intensity based on the CNN-BP neural network.Wireless Personal Communications, 2018, 102(2): 1905-1916doi:10.1007/s11277-018-5245-0 [17] Bahdanau D, Cho K, Bengio Y. Neural Machine translation by jointly learning to align and translate//IEICE Transactions on Fundamentals of Electronics, Communications and Computer Sciences, 2014: 1-15 [18] Hu J, Shen L, Samuel A, et al. Squeeze-and-excitation networks.IEEE Transactions on Pattern Analysis and Machine Intelligence, 2019, 42(8): 1-13 [19] Wang Q, Wu B, Zhu P, et al. ECA-Net: Efficient channel attention for deep convolutional neural networks//IEEE Computer Society Conference on Computer Vision and Pattern Recognition, 2020: 1-12 [20] Woo S, Park J, Lee JY, et al. CBAM: Convolutional block attention module//European Conference on Computer Vision, 2018: 1-17 [21] 王溢琴, 董云云, 刘慧玲. 基于GoogLeNet和空间谱变换的高光谱图像超分辨率方法. 光学技术, 2022, 48(1): 93-101 (Wang Yingqin, Dong Yunyun, Liu Huiling. Super-resolution method of hyperspectral image based on GoogLeNet and spatial spectrum transformation.Optical Technique, 2022, 48(1): 93-101 (in Chinese) [22] 周杰, 赵婷婷, 陈青青等. 基于 GoogLeNet 的混凝土细观模型应力-应变曲线预测. 应用数学和力学, 2022, 43(3): 290-299 (Zhou Jie, Zhao Tingting, Chen Qingqing, et al. Prediction of concrete meso-model stress-strain curves based on GoogLeNet.Applied Mathematics and Mechanics, 2022, 43(3): 290-299 (in Chinese) [23] 狄少丞, 冯云田, 瞿同明等. 基于深度学习和细观力学的颗粒材料本构关系研究. 力学学报, 2021, 53(10): 2712-2723 (Di Shaocheng, Feng Yuntian, Qu Tongming, et al. Data-driven stress-strain modeling for granular materials through deep reinforcement learning.Chinese Journal of Theoretical and Applied Mechanics, 2021, 53(10): 2712-2723 (in Chinese) [24] Efe GD, Zohar R, Sebastian HD, et al. Symmetries and phase diagrams with real-space mutual information neural estimation.Physical Review E, 2021, 104(6-1): 064106 [25] 胡丹丹, 张忠婷, 牛国臣. 融合 CBAM 注意力机制与可变形卷积的车道线检测. 北京航空航天大学学报, 2022, doi:10.13700/j.bh.1001-5965.2022.0601Hu Dandan, Zhang Zhongting, Niu Guochen. Lane line detection incorporating CBAM attention mechanism and deformable convolution.Journal of Beijing University of Aeronautics and Astronautics,2022, doi:10.13700/j.bh.1001-5965.2022.0601(in Chinese) [26] Zhang YH, Chen QQ, Wang ZY, et al. 3D mesoscale fracture analysis of concrete under complex loading.Engineering Fracture Mechanics, 2019, 220: 106646 [27] Li BB, Jiang JF, Xiong HB, et al. Improved concrete plastic-damage model for FRP-confined concrete based on true tri-axial experiment.Composite Structures, 2021, 269: 114051doi:10.1016/j.compstruct.2021.114051 [28] 陈青青, 张煜航, 张杰等. 含孔隙混凝土二维细观模型建模方法研究. 应用数学和力学, 2020, 41(2): 182-194 (Chen Qingqing, Zhang Yuhang, Zhang Jie, et al. Study on a 2 D mesoscopic modeling method for concrete with voids.Applied Mathematics and Mechanics, 2020, 41(2): 182-194 (in Chinese) [29] Krizhevsky A, Sutskever L, Hinton GE. ImageNet classification with deep convolutional neural networks.Communications of the ACM, 2017, 60(6): 84-90doi:10.1145/3065386 [30] 姜克杰, 胡松, 韩强. 基于长短期记忆网络的FRP约束混凝土圆柱循环轴压应力应变预测模型. 工程力学, 2022, 39: 1-15 (Jiang Kejie, Hu Song, Han Qiang. Cyclic axial compressive stress-strain prediction model for FRP-constrained concrete cylinder based on long short-term memory networks.Engineering Mechanics, 2022, 39: 1-15 (in Chinese) [31] Loffe S, Szegedy C. Batch Normalization: Accelerating deep network training by reducing internal covariate shift//IEICE Transactions on Fundamentals of Electronics, Communications and Computer Sciences, 2015: 1-8 [32] 马闯, 杨晓龙, 陈含爽等. 基于平均场近似的BP算法求解随机块模型. 物理学报, 2021, 70(22): 345-356 (Ma Chuang, Yang Xiaolong, Chen Hanshuang, et al. A mean-field approximation based BP algorithm for solving the stochastic block model.Acta Physica Sinica, 2021, 70(22): 345-356 (in Chinese) [33] 汪恩良, 田雨, 刘兴超等. 基于WOA-BP神经网络的超低温冻土抗压强度预测模型研究. 力学学报, 2022, 54(4): 1145-1153 (Wang Enliang, Tian Yu, Liu Xingchao, et al. Prediction model of compressive strength of ultra low temperature frozen soil based on WOA-BP neural network.Chinese Journal of Theoretical and Applied Mechanics, 2022, 54(4): 1145-1153 (in Chinese) [34] 陈苏, 丁毅, 孙浩等. 物理驱动深度学习波动数值模拟方法及应用. 力学学报, 2023, 55(1): 272-282 (Chen Su, Ding Yi, Sun Hao, et al. Methods and applications of physical information deep learning in wave numerical simulation.Chinese Journal of Theoretical and Applied Mechanics, 2023, 55(1): 272-282 (in Chinese) [35] Aleboyeh A, Kasiri MB, Olya ME, et al. Prediction of azo dye decolorization by UV/H2O2using artificial neural networks.Dyes and Pigments, 2008, 77(2): 288-294doi:10.1016/j.dyepig.2007.05.014