高压水射流靶物反射声信号的特征提取方法

为了提高利用高压水射流靶物反射声信号识别靶物材质的效率,针对地雷探测过程常见的地雷、石块、砖块和木块4种靶物,采用不同的特征提取方法来识别靶物材质。在分析Mel频率倒谱系数及小波包变换倒谱系数基本原理的基础上,结合靶物反射声信号的特点,提出了一种基于Mel频率倒谱和小波包变换倒谱特征融合的特征提取方法:利用小波包变换将原始靶物反射声信号划分为若干子频段,选取其中一个子频段作为低频和高频的划分层;低频部分提取Mel频率倒谱系数作为特征值,高频部分则提取小波包变换倒谱系数作为特征值,将2组特征值线性合并为一组新的特征向量,用于靶物材质的识别。采用最小二乘支持向量机建立多分类模型,验证基于单一特征和基于特征融合的特征提取方法的识别率。实验结果表明,在取得低频与高频的最佳划分层时,基于特征融合的特征提取方法的平均识别率达到82.812 5%,较单一的利用Mel频率倒谱系数或小波包变换倒谱系数作为特征向量时的平均识别率分别提高了10.312 5%和7.812 5%。

Feature extraction method for reflective sound signal of high pressure water-jet target

In order to improve recognition rate of target materials by using reflective sound signal of high pressure water-jet,in view of four common targets of mine,stone,brick and wood block,different feature extraction methods were used to identify target materials.On basis of analyzing basic principles of Mel frequency cepstral coefficients and wavelet packet transform cepstral coefficients,combining with characteristics of reflective sound signal of target,a feature extraction method based on feature fusion of Mel frequency cepstral coefficients and wavelet packet transform cepstral coefficients was presented.The reflective sound signal of original target was decomposed to several sub-bands by using wavelet packet transform,and one of the optimal sub-band was selected as separate layer of low frequency and high frequency.Mel frequency cepstral coefficients were calculated as eigenvalues in low frequency part,and wavelet packet transform cepstral coefficients were calculated as eigenvalues in high frequency part.The two groups of eigenvalues were merged into a new set of linear feature vector,and the new vector was input into target identification model.LS-SVM classification model was built to evaluate recognition rate of the feature extraction methods based on single characteristic and feature fusion.The experiment results show that when the best division layer between low frequency and high frequency was acquired,the average recognition rate of feature extraction method based on feature fusion reaches 82.812 5%,there was a increase of 10.312 5% and 7.812 5% compared with using Mel frequency cepstral coefficients or wavelet packet transform cepstral coefficients as feature vector.