基于MLP-Bagging的PCB电热耦合建模方法

PCB Electric-Thermal Coupling Modeling Method Based on MLP-Bagging

随着三维集成电路性能的提高和复杂程度的增加,印制电路板(PCB)的散热问题日益突出。研究了PCB在电热多物理场相互作用下各部件的发热情况,提出了基于混合激活函数的多层感知机(MLP)-Bagging多物理参数算法。通过使用ReLU和Sigmoid两个激活函数进行学习和训练,建立了精度更高的MLP模型。之后,结合Bagging算法构建多个并行的MLP模型。所提出的神经网络多物理模型可以快速准确地预测PCB的电热响应。实验结果表明,此方法与有限元法相比,可以节省约97%的计算内存和99%的计算时间,与传统神经网络如随机森林(RF)、长短时记忆(LSTM)网络、MLP相比,该方法表现优良且泛化能力较好,为提高PCB设计效率提供了一种可行方法,为PCB热分析提供了更高效的解决方法。

With the improvement of the performance and complexity of 3D integrated circuits,the heat dissipation problem of printed circuit board(PCB)has become more and more prominent.The heating situation of PCB components under the interaction of electric-thermal multi-physical fields was studied.A multi-layer perceptron(MLP)-Bagging multi-physical parameter algorithm based on mixed activation function was proposed.By using ReLU and Sigmoid activation functions for learning and training,the MLP model with higher precision was built.Then,several parallel MLP models were constructed with Bagging algorithm.The proposed neural network multi-physical model can predict the electric-thermal response of PCB quickly and accurately.The experimental results show that compared with the finite element method,the proposed method can save about 97%of the computing cost and 99%of the computing time.Compared with the traditional neural networks such as random forest(RF),long short-term memory(LSTM)network and MLP,the proposed method has good performance and good generalization ability,and provides a feasible method for improving the efficiency of PCB design,and provides an efficient solution for PCB thermal analysis.