Design of Direct Digital Frequency Synthesizer Based on Improved RBF Neural Network

Abstract

A high-performance direct digital frequency synthesizer based on an improved radial basis function (RBF) neural network is proposed. Compared with the traditional direct digital frequency synthesizer, it avoids phase truncation error and reduces resource consumption. In order to further improve the training efficiency and stability of the RBF neural network, an improved RBF neural network training algorithm is proposed. In the coarse adjustment stage, the K-means++ algorithm is used to quickly determine the initial activation function center, making the activation function center distribution more reasonable; in the fine adjustment stage, the L-BFGS-B algorithm is used to fine-tune the optimal center obtained in the coarse adjustment stage to further reduce the output error. The experimental results on the general FPGA platform show that when the output clock frequency of the direct digital frequency synthesizer based on the improved RBF neural network is 1.53 MHz, the spurious free dynamic range is 85.26 dB, the phase noise is -90.50 dBc/Hz@100 kHz, and no additional ROM resources are required.