基于核独立成分分析的声信号去噪方法

A denoising method for acoustic signal based on kernel independent component analysis

野外环境无线传感侦查网络中的声识别技术面临着复杂的自然环境噪声的挑战,尤其是由强风噪声造成的影响。独立成分分析(ICA)方法是一种能够较好地解决这种复杂环境去噪的方法。引入一种基于核方法的非线性ICA方法——核独立成分分析(KICA)。基于该算法,针对强风噪声的特性,设计一种应用于单声传感器降噪的方案。通过降噪仿真实验,对KICA与ICA的典型算法快速ICA算法进行比较。实验结果表明:在以均方误差作为指标的降噪性能和实际分类识别率两方面,KICA算法相较于快速ICA算法对于该种强噪声具有更为优秀的降噪效果。

The acoustic recognition technology in wireless sensor surveillance network in wild environment is facing the challenge of the complicated and strong acoustic noise, especially the effect of the wind noise. Independent component analysis（ICA） method has a good performance in denoising in complicated environment. On the basis of kernel independent component analysis（ KICA）, a denoising scheme for single acoustic sensor is designed. It is a typical nonlinear independent component analysis method. Through denoising simulation experiment, the KICA algorithm is compared with the typical algorithm of ICA method, fast ICA algorithm. Results show that in a denoising application aimed at the strong wind noise in wild environment,with the mean square error criterion and the correct classification ratio criterion, KICA algorithm is superior to the fast ICA algorithm.