Neural networks for RF and microwave design: Toward automatic model generation.

摘要

Over the last decade, rapid progress in telecommunications, VLSI and wireless industries resulted in ever-increasing circuit complexities and signal frequencies. This, coupled with tightened design tolerances and shortened time-to-market cycles, necessitated powerful computer aided design (CAD) methodologies. Artificial neural networks (ANN) have recently gained recognition as accurate and fast vehicle for RF and microwave CAD including modeling, simulation and optimization. Key objectives of the thesis are, (a) To develop robust neural modeling techniques that can address complicated high-frequency modeling challenges, and (b) To formulate efficient neural network modeling algorithms that can facilitate automatic generation of accurate RF and microwave neural models.; Major contributions of the thesis include, the iterative multi stage (IMS) algorithm, the automatic model generation (AMG) algorithm, and the knowledge-based automatic model generation (KAMG) algorithm. Through iterative decomposition and stage-wise training, the IMS allows neural modeling of otherwise complicated highly nonlinear and multi-dimensional RF/microwave behaviors even in the presence of gross errors or sharp variations in training data. Starting with no training data and proceeding with adaptive training and validation data augmentation during neural network training, the AMG automatically generates accurate neural models with reduced use of data and shortened model development time. Motivated by the space-mapping concept, the KAMG utilizes extensive coarse data but fewest fine data as compared to any other existing neural modeling techniques to generate neural network models that accurately match fine data, with minimal human intervention. By virtues of automation, efficiency and robustness, the work presented in the thesis enhances today's high-frequency electronics design and strengthens the ground for computerization of neural network based end-to-end CAD methodologies in future.