【作者】 赵德双；

【导师】 王秉中；

【作者基本信息】 电子科技大学 ， 电磁场与微波技术， 2001， 硕士

【摘要】 对于当今高性能快速的电磁工程计算机辅助设计（CAD）来说，建立起精确可靠快速的元器件模型显得十分有必要。近年来，人工神经网络（ANN）作为一种快速准确的建模工具在电磁工程领域中引起了广大工程设计者的极大关注。 本文在人工神经网络建模的基础上引入了可用于复杂的电磁工程建模的知识人工神经网络（KBNN），同时提出了两种新型结构的知识人工神经网络：一种是采用主要元素项分析（PCA）作为网络训练数据前处理器的稳健的知识人工神经网络（RKBNN）；另一种是含有知识人工神经元的人工神经网络（NNKBN），其知识人工神经元的活化函数是由扩展的经验公式构成。 首先，应用RKBNN建立起了高速数字集成电路（HSDIC）共面互连线结构的频变电阻电感网络计算模型。结果表明：通过在网络中加入PCA训练数据前处理器，网络训练的效率和稳定性都得到了很大的提高，同时受训后各RKBNN模型还具有较强的推广能力。 其次，在训练样本数据匮乏的情况下，应用NNKBN建立起了广泛应用于多芯片封装模块（MCM）中带状线间隙的不连续特性网络计算模型。结果表明带状线间隙不连续特性的NNKBN模型在保持较高的建模精度下具有良好的外推特性。 总之，应用本文提出的知识人工神经网络所建立的网络计算模型不仅保持了电磁数值仿真的精度，而且还可以降低它们对CPU和内存等硬件的要求，同时还具有良好的外推特性。因此，通过在网络结构中镶入建模对象先验知识信息的人工神经网络在高性能快速的电磁工程CAD中具有很大的应用潜力。更多还原

【Abstract】 For today’s high performance and fast electromagnetic engineering computer-aided design, it is increasingly necessary to model elements and devices with high accuracy, reliability and efficiency. Artificial neural network (ANN) recently has received extensive attention as a fast and accuate modeling tool in electromagnetic engineering.In this paper, knowledge-based neural network (KBN7N) is introduced to model sophisticated electromagnetic objects by incorporating prior knowledge into artificial neural network structures. Two novel knowledge-based neural network structures are presented. One is the robust knowledge-based neural network (RKBNN) with principle component analysis (PCA) as data pre-processor for network training. The other is the neural network with knowledge-based neurons (NNKBN) where extended prior knowledge analytic formulas work as activation functions of the neurons.Firstly, the RKBNN is used for modeling the frequency-dependent resistance and inductance extraction of coplanar interconnect in high speed digital integrated circuits (HSDIC) . Results show that the network training procedure becomes efficient and stable with PCA as data pre-processor and the developed RKBNN models are robust for generalization.Secondly, the NNKBN trained with insufficient training data is applied to model the discontinuity of nonsymmetrical stripline gap widely used in multi-chip package module (MCM) . Results show that the N7NKBN models are good for extrapolation with high accuracy.All of the proposed models not only preserve the accuracy of the EM simulations, but also simplify their CPU and memory requirements, and at the same time keep good extrapolation capability. Hence, the artificial neural networks incorporated with prior knowledge information of the problems to be modeled have potential power for the high performance and fast CAD in electromagnetic engineering.更多还原