NEURAL NETWORK AND IDENTITY ANALYSIS : THEIR APPLICATION IN PMSM CONTROL AND DESIGN

Abstract

This thesis reports on the research that has been conducted permanent-magnet in the area of study of novel synchronous motor (PMSM) design. The main focus of the thesis is the study of novel optimization strategies which combines neural networks with finite element and the use of a new paradigm for design synthesis of PMSM. A novel mapping network called the Forward Generating Neural Network (FGNN) is presented. With this network, one pass training is done from input layer to output layer. The training of the hidden layer is supervised and this reduces the training time tremendously. FGNN is used in the Finite Element method - Neural Network method which combines analytical capability of the finite synthesis and mapping capability of neural networks to solve electromagnetic design analyzing problems. This technique is used with, the concept of Identity for designing and motor the performance of electric synchronous machines. The identity of electric is a matrix expression the motor that is used to describe the torque-position characteristic of motor. Identity Analysis provides the "identity" of the motor which guides the controller and motor design. An efficient method called the Hybrid Maxwell Tensor Method is proposed in this thesis. The advantages of this method are its accuracy and low sensitivity to the shape of airgap elements. This method is thus effective in the extraction of the identity of synchronous motor in design procedure. As the characteristic of a PMSM modifying is linked to its identity, the required motor performance can be obtained by demonstrate its identity. This is the principle of Identity Design. Examples are used to experimental the applications of the methods proposed and both simulation results and methods results show that the identity control and identity design are two effective of for reducing the torque ripples and increasing the efficiency of PMSM. The value hybrid the forward Maxwell generating neural network, finite element - neural network method and tensor method are also confirmed in the successful applications of the identity control and identity design.