基于BP神经网络和遗传算法的刚挠背板设计

Anti-vibration design of rigid-flex backplane based on BP neural network and genetic algorithm

选取加强筋宽度等5个关键因素,采用正交设计法采集刚挠背板基频形成训练样本。由试验确定BP神经网络拓扑结构。选用LM算法训练的BP神经网络(BPNN)作为遗传算法目标函数求解器,用于优化抗振结构。结果表明,网络拓扑结构为4-6-1时网络泛化能力强,测试误差小于1.6%;获得最优结构参数组合x1～x5分别为0.0039,0.004,0.0268,0.0242和0.0018m;优化后,基频提高92.7%,振幅降低82.77%,计算误差为0.636%。

In this study, stiffening width and other four parameters were selected as key factors. Training samples were generated based on the fundamental frequency of rigid-flex backplane collected using orthogonal design method. The topological structure of neural network was determined via testing. To optimize the anti-vibration structure, the back propagation neural networks （BPNN）, trained by Levenberg-Marquardt （LM） algorithm, was used as the objective function solver for genetic algorithm （GA）. The results indicate that, when the topological structure of neural network is in the state of 4-6-1, the network generalization capability is large and the testing error is less than 1.6%; The key factors x1 to x5 for the optimal anti-vibration structure are 0.003 9, 0.004 0.026 8, 0.024 2 and 0.001 8 m respectively; After optimization, the fundamental frequency increases by 92.7% and the resonant amplitude decreases by 82.77%, with a calculation error of 0.636%.