基于深度学习的复合材料螺栓连接失效预测

FAILURE PREDICTION OF BOLTED CONNECTION OF COMPOSITE MATERIALS BASED ON DEEP LEARNING

针对复合材料螺栓连接失效强度分析与预测问题,利用深度学习神经网络强大的非线性映射能力,将不同参数对复合材料螺栓连接失效载荷的影响进行非线性拟合,并分配各参数的影响权重;通过有限的训练样本构建预测模型,来预测复合材料螺栓连接的失效峰值载荷。使用有限元软件计算得到层合板螺栓连接失效峰值载荷的数据组,以此来构建深度学习神经网络。通过测试确定当隐藏层数量为两层时深度学习模型开发效果最佳,以预测值与有限元仿真值之间的均方误差作为损失函数、学习速率取0.01,当均方误差最小时停止训练,此时得到最佳深度学习预测模型;利用该模型预测得到所有失效峰值载荷预测结果中的最大值以及对应的参数组合,并与同样参数的仿真结果进行对比,两者相差1.4%;相比有限元仿真和拟合经验公式的预测方法,深度学习预测方法具有明显的时间效率优势。

Aiming at the problem of failure strength analysis and prediction of bolted composite connection,the strong nonlinear mapping ability of deep learning neural network was used to non-linear fit the influence of different parameters on the failure load of composite bolting,and the influence weight of each parameter was allocated.A prediction model was constructed based on limited training samples to predict the peak failure load of bolted composite joints.Using finite element software,the data set of peak failure load of bolted laminates was calculated to construct the deep learning neural network.Through the test,it is determined that the development effect of deep learning model is the best when the number of hidden layers is two.The mean square error between the predicted value and the finite element simulation value is taken as the loss function,and the learning rate is set at 0.01.When the mean square error is the minimum,the training is stopped,and the best deep learning prediction model is obtained.The model is used to predict the maximum value of all the prediction results of peak load failure and the corresponding parameter combination,and compared with the simulation results of the same parameters,the difference between them is 1.4%.Compared with the prediction methods of finite element simulation and empirical formula fitting,the deep learning prediction method has obvious advantages in time efficiency.