基于神经网络无铆钉自冲铆接头力学性能预测

Neural network-based mechanical property prediction in the mechanical clinching joints

为研究钢铝一体化结构车身无铆钉自冲铆接接头力学性能,引入反向传播神经网络模型来描述板材厚度、板材硬度和成形接头底部直径等工艺参数与接头剪切力及剥离力强度等力学性能的映射关系。由于标准反向传播网络存在训练精度低、收敛速度慢及泛化能力差等缺陷,采用归一化法与Levenberg-Marquardt算法相结合的算法来优化神经网络预测模型连接权值,提高了神经网络模型的预测精度和泛化能力。对神经网络的预测结果进行检验的结果表明,训练后的神经网络模型能够准确有效地预测无铆钉自冲铆接接头力学性能,证实了神经元网络应用于无铆钉自冲铆接接头力学性能预测的可行性与可靠性,为优质的钢与铝无铆钉自冲铆接接头的设计提供了依据。

In order to investigate the property of mechanical clinching joints in the steel-aluminum hybrid structure car body, back-propagation （BP） neural networks were introduced to describe the mapping relationships among such joining technique parameters as sheet thickness, sheet hardness and joint bottom diameter and joints mechanical properties of shearing and peeling. To overcome the existing disadvantages of the standard propagation algorithm in lower training precision, low convergence speed and weak generalization capability, the algorithms of normalization and Levenberg-Marquardt were combined to optimize the standard BP neural network connection weights and improve the model prediction precision and generalization capability. The training and validating samples were performed by the BTM Tog-L-Loc system with different combinations of technique parameters. Then the parameters of training samples and the corresponding joint mechanical properties were supplied to the neural networks for training, and the experimental data validating samples were used to check up the prediction outputs. The results showed that the neural network prediction models after training could effectively predict the mechanical properties of joints and proved to be feasible and reliable, so as to improve the design of the clinching joints applied in steel-aluminum hybrid structure automotive body manufactures