Algorithm Of Neural Network Training Data Generation Used In Automatic Modeling Of Microwave Applications

In recent years, with the rapid development of wireless communication industry,microwave device and circuit design has become the focus of attention. Microwavedevice and circuit design is moving to the direction of high efficiency, low cost andhigh quality. Artificial neural network provides efficient tool for nonlinear mappingand real-time computing. Aneural network model for microwave device or circuit canbe developed by learning and abstracting from measured or simulated microwave data,through a process called training. Once be trained, neural network model can be usedduring microwave design to provide instant answers to any pre-assigned task.Nowadays, artificial neural network techniques has been widely used to solve variousproblems of microwave/RF devices and circuits, such as transistor modeling, coplanarwaveguide element modeling, amplifier modeling, filter modeling, nonlinearmicrowave circuit optimization and so on.Typically, traditional microwave device and circuit modeling technique istime-consuming. Therefore, it is indispensable to design an automatic neural networkmodeling system and the automatic training data generation is the key aspect.Microwave modeling problems are often highly nonlinear and multidimensional. Thenumber of samples needed for developing a neural model with desired accuracy andtheir distribution in the input space are not obvious. While too many samples areexpensive, too few samples lead to over-learning of the neural network.Firstly, this paper introduces the key technologies of microwave modeling andartificial neural network and analyzes the relevant data distribution techniques. Whatis more, this dissertation provides a comprehensive discussion on existing datageneration algorithms and discovers that the existing data generation algorithms existsa series of shortcomings, such as excessive number of training data needed,time-consuming and so on. Finally, this paper proposes a novel neural networktraining data generation algorithm and designs an intelligent neural network modelingsystem. Examples of power amplifier and microstrip filter modeling are used toillustrate the effectiveness and efficiency of the novel data generation algorithm.