Page 81 - 《应用声学》2022年第1期
P. 81

第 41 卷 第 1 期                                                                       Vol. 41, No. 1
             2022 年 1 月                          Journal of Applied Acoustics                   January, 2022

             ⋄ 研究报告 ⋄



                   水声通信近似L 范数约束的BP网络均衡器
                                                 0




                                           王 凯    1,2  吴立新      1,2†  张青青      1


                                    (1  中国科学院声学研究所      声场声信息国家重点实验室         北京  100190)
                                                (2  中国科学院大学     北京   100049)

                摘要：针对稀疏水声信道的长时延扩展及梯度下降的权值迭代方案导致的神经网络均衡器收敛速度慢的问
                题，提出了近似 L 0 范数约束的 BP 神经网络均衡器。首先在传统 BP 网络均衡器基础上增加判决反馈项，然后
                在代价函数中对均衡器输入层到隐含层的权值增加 L 0 范数约束，构造新的代价函数，利用高斯族函数近似 L 0
                范数约束，并根据不同隐层神经元节点输出权值的 L 2 范数设定近似参数。仿真结果表明，稀疏信道条件下，本
                方法相比传统的 BP 网络均衡器收敛速度更快，误码率更低，可以有效提升神经网络均衡器的性能。
                关键词：水声通信；范数约束；神经网络；信道均衡
                中图法分类号: TN929.3           文献标识码: A          文章编号: 1000-310X(2022)01-0077-07
                DOI: 10.11684/j.issn.1000-310X.2022.01.009






              Approximate L 0 norm constrained BP neural network equalizer for underwater
                                               acoustic communication



                                      WANG Kai  1,2  WU Lixin 1,2  ZHANG Qingqing  1


                  (1  State Key Laboratory of Acoustics, Institute of Acoustics, Chinese Academy of Sciences, Beijing 100190, China)
                                   (2  University of Chinese Academy of Sciences, Beijing 100049, China)

                 Abstract: Aiming at the problem of slow convergence speed of neural network equalizer caused by long delay
                 spread and gradient descent weight iterative scheme of sparse underwater acoustic channel, a BP neural network
                 equalizer with approximate L 0 norm constraint is proposed. Firstly, the decision feedback term is added to the
                 traditional BP network equalizer, and then the L 0 norm constraint is added to the weight from the equalizer
                 input layer to the hidden layer in the cost function to construct a new cost function. The Gauss family
                 function is used to approximate the L 0 norm constraint, and the approximate parameters are set according to
                 the L 2 norm of the hidden layer neurons. The simulation results show that, compared with the traditional BP
                 network equalizer, this method has faster convergence speed and lower bit error rate in sparse channel, which
                 can effectively improve the performance of neural network equalizer.
                 Keywords: Underwater acoustic communication; Norm constraint; Neural network; Channel equalization





             2021-03-08 收稿; 2021-04-09 定稿
             作者简介: 王凯 (1994– ), 男, 陕西咸阳人, 博士研究生, 研究方向: 水声通信。
              通信作者 E-mail: wlx@mail.ioa.ac.cn
             †
   76   77   78   79   80   81   82   83   84   85   86