The Application Of Support Vector Machine On RF Power Device Modeling

With the development of communication technology, more and more RF power devices work in large signal circumstance, and these devices show severe nonlinearity. The traditional S-parameter theory is suit to analyze linear system, but to nonlinear problems, it is no longer suitable. Therefore, it is necessary to find a new way to substitute the S-parameter theory. Nowadays, many researchers use the nonlinear scattering function theory, an extension of the traditional S-parameter theory, as the substitute. It can solve the problem of modeling nonlinearity in the condition of large signal successfully, so it is used in modeling the characteristic of RF power devices widely. In the meantime, the rapidly developing artificial intelligence technology is widely adopted to model the RF power devices, which is very fast and efficient and has obvious advantages compared to the traditional methods.Firstly, this dissertation introduces the modeling methods of RF power devices and takes the behavior modeling methods as a key point, including Power Series Method, Volterra Series Method, Envelope-domain Model Method, Harmonic Balance Method and Neural Network Model Method. The dissertation also shows some methods used for charactering nonlinearity of devices, such as AM-AM, AM-PM, Intermodulation Distortion and Third-order Intercept Point. Then the nonlinear scattering function theory is analyzed as a key point for the later research.Secondly, the mainstream of RF power device modeling is with the help of CAD and automation design. The way of modeling using artificial intelligence algorithm and nonlinear scattering function is more convenient and efficient than the traditional way. Therefore, this dissertation also introduces the theory of SVM In the later part, preparing to use the regression of SVM to fit the nonlinear scattering function of RF power devices.Then, the SVM theory is applied to modeling RF power devices. Because of the good regression characteristic of SVM, experiments prove that the new method is better than the neural network algorithm in the domain of RF power devices modeling. Next, aiming at the characteristic of train data and nonlinear scattering function, the SVM is improved to form the domain knowledge SVM by leading some knowledge of nonlinear scattering function into its constraint condition. As a result, modeling errors caused by the inaccuracy of the test data are reduced. This dissertation also analyzes the effect of noise on model from the hypothetical point.Finally, this dissertation summarizes the mistakes which are made during the research and finds out the reasons.