基于GSC的改进生成对抗网络噪声干扰抵消算法  

An Improved Generated Adversarial Network Noise Cancellation Algorithm Based on GSC

作　　者：向亚丽 颜冰[1] XIANG Ya-li;YAN Bing(School of Weapon Engineering,Naval University of Engineering,Wuhan 430033,China)

机构地区：[1]海军工程大学兵器工程学院,武汉430033

出　　处：《武汉理工大学学报》2020年第7期83-91,共9页Journal of Wuhan University of Technology

基　　金：国防预研项目(41419010208).

摘　　要：广义旁瓣抵消方法(GSC)对加性噪声环境下的含噪语音信号具有很好的噪声抑制作用,然而由于阻塞矩阵设计过于简单、后置滤波采用的LMS自适应滤波算法存在收敛速度较慢等问题,导致GSC辅助通道阻塞矩阵的输出残留部分期望信号,残留的期望信号导致后置滤波算法的滤波性能恶化。提出将深度学习中广泛使用的生成式对抗网络引入到广义旁瓣噪声抵消算法中,通过改进生成器模型,利用生成模型强大的学习能力生成与主通道中噪声干扰分量相关性更强的参考分量,提高了主通道中噪声干扰分量估计的准确性。为了验证所提算法的滤波性能,在实验室环境下搭建天线阵列,设置不同干扰类别、不同噪声干扰方向进行多组实验。结果表明,提出的改进生成对抗网络具有更好的学习能力、更好的鲁棒性,降低了算法的运算复杂度,提高了接收信号的信噪比。The generalized sidelobe offset method(GSC)of additive noise environment noise speech signal has very good noise suppression effect,however,due to the design of block matrix is too simple,the rear filter USES LMS adaptive filtering algorithm problem such as slow convergence speed,and result in GSC auxiliary channel blocking matrix output remain part of the desired signal,the residual of the desired signal leads to the deterioration of the rear filtering algorithm of filtering performance.Proposed in this paper will study the depth of the widely used to generate the type against network is introduced into the generalized sidelobe noise cancellation algorithms,by improving the generator model,using the generation model of learning ability to generate a stronger correlation with the noise components in the main reference component,improve the precision of the main noise components.In order to verify the filtering performance of the proposed algorithm,an antenna array is set up in the laboratory environment,and multiple experiments are carried out with different interference categories and different noise interference directions.The experimental results show that the proposed improved generative adversarial network has better learning ability and robustness,reduces the computational complexity of the algorithm,and improves the SNR of the received signal.

关 键 词：改进的生成对抗网络 生成器 广义旁瓣抵消 阻塞矩阵 后置滤波算法 

分 类 号：TN911.7[电子电信—通信与信息系统]