Research And Implementation Of Neural Network Controller Based On FPGA

Nowadays the application requirement of complex data processing makes the neural network become increasingly important. Neural network has been more and more widely used in intelligent control, pattern recognition and other aspects with its characteristics of self adaptability, fault tolerance and parallel processing. In some applications such as real time control, the implementation of the neural network based on the microprocessor can not reflect the parallelism nature of the neural network. But the field programmable gates array (FPGA) has been widely used in the realization of neural network, which is characterized by its rich on-chip resources and high-speed parallel data processing.This dissertation mainly focuses on the application of different digital design methods in the FPGA implementation of neural networks. Firstly, the parameter learning and structure optimization of neural network are introduced briefly, then how to realize the nonlinear activation function of neural network efficiently is studied, and the characteristics of several FPGA realization methods of nonlinear activation function are compared in this dissertation. The principle, the error of implementation and resource usage of STAM method are analyzed in detail, and the FPGA implementation of this method is simulated and verified. After analyzing the computing process in the whole training period, the structure of the basic neuron is improved according to the principle of time division multiplexing, and the neural network is divided into different states in the whole training period. Based on these work, the architecture of neural network is designed based on the finite state machine. In view of the problem of multiplication of constant coefficient, this dissertation introduces the two methods of distributed computing and sub-expression redundancy. And then the two methods are applied to the design of off-line neural network to reduce the compution complexity. Finally, the FPGA implementation of these two kinds of neural network structure are simulated and analyzed.