Neural network application to high performance electric drives systems

With the introduction of improved training algorithms and new network topologies, neural networks (NN) have demonstrated their feasibility and practicality in several applications including electric drives. An artificial neural network using parallel and distributed processing units can achieve the functions of system modeling and control. NN have several key features that make them highly suitable for high power drive (HPD) applications. There are several issues that must be addressed before the NN is used for practical applications. Among them are the selection of the NN structure, the proper training algorithm, the memorization and saturation of the NN, the correlation of the training data, the size and range of the training data, the distribution of the training data in the operation space, the inversion of the NN for contour tracking and the implementation of query techniques. The application and implementation of the neural network in HPD are becoming more complex and sophisticated, the open-loop black-box approach to the design, training and verification can no longer be effective or reliable. This paper shows that any NN design can be greatly enhanced when ancillary techniques are used before, during and after training. An application to a brushless DC motor is discussed