摘要
提出了一种基于改进型径向基函数(Radial basis function,RBF)神经网络的高性能直接数字频率合成器,相比于传统的直接数字频率合成器避免了相位截断误差并降低了资源消耗。为了进一步提高RBF神经网络的训练效率及稳定性,提出一种改进型的RBF神经网络训练算法。该算法在粗调阶段,利用K-means++算法快速确定初始激活函数中心,使激活函数中心分布更加合理;在细调阶段则采用L-BFGS-B算法,对粗调阶段得到的最佳中心进行精细调整,进一步降低输出误差。通用FPGA平台的实验结果表明,基于改进型RBF神经网络的直接数字频率合成器当输出时钟频率为1.53 MHz时,无杂散动态范围为85.26 dB,相位噪声为-90.50 dBc/Hz@100 kHz,且无需占用额外ROM资源。
A high-performance direct digital frequency synthesizer based on an improved radial ba-sis function neural network is presented which avoids phase truncation errors and reduces resource con-sumption compared to traditional direct digital frequency synthesizers.To enhance the training efficien-cy and stability of the RBF neural network,an improved RBF neural network training algorithm is pro-posed.In the coarse-tuning stage,the K-means++algorithm is used to quickly determine the initial centers of activation functions,ensuring a more reasonable distribution.The fine-tuning stage employs the L-BFGS-B algorithm to adjust the optimal centers obtained in the coarse-tuning stage,further re-ducing the output error.The experimental results on a general FPGA platform show that the designed synthesizer achieves a spurious free dynamic range of 85.26 dB and a phase noise of-90.50 dBc/Hz@100 kHz at an output clock frequency of 1.53 MHz,without needing additional ROM resources.
作者
倪崧顺
张长春
王静
张翼
NI Songshun;ZHANG Changchun;WANG Jing;ZHANG Yi(School of Integrated Circuit Science and Engineering,Nanjing University of Posts and Telecommunications,Nanjing,210003,CHN;State Key Laboratory of Millimeter Waves,Southeast University,Nanjing,210096,CHN)
出处
《固体电子学研究与进展》
CAS
2024年第2期149-156,共8页
Research & Progress of SSE
基金
国家自然科学基金资助项目(62174090)
毫米波国家重点实验室开放课题(K202325)。
