Design o f an FPGA Based Neural Controller

This paper describes an efficient implementation of neural multi-layer networks on FPGA fabric (field programmable gate array). A back-propagation algorithm was used for the training task while implementation and synthesis tools are centred on the ISE 6.3 of Xilinx with the targeted components being Virtex II and Virtex II Pro. A fixed point and a floating point number representation were used for encoding real numbers and for data processing, respectively. The realization of the activation function was carried out according to three methods, for which the results of simulation and synthesis are also presented. The implementation performances were tested using an approximation of some linear and non-linear functions. Of particular importance, an experimental evaluation involving the speed control of a DC motor is given to demonstrate the features of the adopted methodology.