基于Huffman最优二叉树支持向量机的舱音记录器背景信号识别  

Signal Recognition Based on Huffman Optimal Binary Tree SVM for CVR

作　　者：杨琳 王从庆[2] 缪鹏[2] 姜龙生[2] 

机构地区：[1]中国民用航空科学技术研究院,北京100028 [2]南京航空航天大学自动化学院,南京210016

出　　处：《宇航学报》2011年第6期1428-1434,共7页Journal of Astronautics

基　　金：国家自然科学基金(60776819)

摘　　要：飞行器舱音记录器(CVR)记录的舱音信号,通常是语音声、警告声、开关按钮声和背景噪声等混合而成。目前国内对该类信号的分析和辨别主要是计算机译码后利用人耳进行辨听,存在不易准确分辨各种独立声音信号的缺点。针对舱音信号是一种非平稳性的时频信号,提出了基于多尺度最优小波包基的CVR背景信号特征提取算法,将10种典型信号进行小波包分解,以分解得到的子带能量作为信号初始特征,再根据类间最大距离准则选取最优小波包基,从而确定待识别信号最具有代表性的特征向量,最后基于Huffman最优二叉树支持向量机进行CVR背景信号分类。仿真实验结果表明,该方法的平均识别率为94.62%,可以应用于CVR背景声音信号的自动识别。Signals recorded by CVR（Cockpit Voice Recorder） on aircraft are mixed signal composed by voices,switching knob,alarm sound and noises.So far in domestic,the analysis and identification of these acoustic signals are mainly depended on human audition,difficult to separate independent sounds.Because these signals are non-stationary time and frequency signals,an optimal wavelet packet basis-based algorithm is given by using the arbitrary time-frequency decomposition of wavelet packet transform.Initial characteristics for ten kinds of signals are constituted by the filial generation energy of the wavelet packet decomposition.Optimal wavelet packet basis is determined to obtain the best representative characteristic vectors in terms of the largest distance rule.Automatic recognition experiment based on Huffman Optimal Tree SVM（Support Vector Machine） network is carried out,and results show that it reaches 94.62% recognition rate,so it can be applied to the recognition of non-voice acoustic signals of CVR.

关 键 词：舱音记录器 非话语背景声音信号 Huffman最优二叉树支持向量机 信号自动识别 

分 类 号：TP391.42[自动化与计算机技术—计算机应用技术] TN912.3[自动化与计算机技术—计算机科学与技术]