【作者】 沈勇；

【导师】 颜建军；

【作者基本信息】 华东理工大学 ， 机械电子工程， 2011， 硕士

【摘要】 中医声诊信号是一种语音信号,语音信号发出是一个复杂的非线性过程,从人的发声机理来看,语音信号是非平稳随机过程,其特性是随时间变化的。这使得基于线性系统理论发展起来的传统语音信号处理和识别技术难以进一步提高。本文基于非线性动力学理论利用替代数据法对声诊信号做了非线性检验,检验出信号具有一定的非线性的结论。将小波包变换和近似熵及样本熵结合起来分析本文的三类声诊信号,讨论了近似熵及样本熵与典型信号时间序列波形的频率、振幅、相位、信噪比等关系,本文对声诊信号进行了小波包4层分解,然后计算了经过小波包分解之后的系数的近似熵和样本熵的值,并进行特征参数的统计分析,选择合适参数利用SVM对特征参数进行分类的研究,得出的分类结果比较理想。利用延迟坐标状态空间重构法对一个典型声诊信号分三段进行相空间重构,选用合适的参数计算方法计算出了合适的延迟时间值,利用关联维数G-P算法分析了声诊信号,计算了所有三类样本信号的关联维数值,经过分析可知,声诊信号的关联维数平均值随着嵌入维数的增加将会达到饱和,在嵌入维数为8的前后,正常组和阴虚类样本信号的关联维数值由前者大变为后者大。最后采用关联维数作为特征参数对三类样本进行分类的研究,结果显示,关联维数对于两类分类结果较好,三类分类时效果不太理想。利用典型信号的RP图形象地说明了RP的不同结构特征,选择平均互信息以及虚假近邻法找出了计算声诊信号RQA特征参数的参数值,本文将小波包分解与RQA参结合起来,这样可以从不同频率段分析信号的动力学行为,更加细致的反应递归图上的细节结构,利用这种方法计算了所有三类样本信号的RQA特征参数值,统计分析了对三类样本具有显著性差异的RQA特征,结果显示大部分数小波包分解系数的DET和L_max两个参数值对于三类样本信号有显著性差异。在提取以上特征之后,利用PCA,FDA和KDA特征优化方法对这些特征进行了优化的分析。融合小波包变换和近似熵、样本熵对心系疾病患者、肝系疾病患者、肺系疾病患者、脾系疾病患者、肾系疾病患者等五脏疾病患者和正常人的五个元音信号/a/,/i/,/u/,/e/,/o/进行了研究,统计分析了信号的不同频率段的非线性特征,发现在五脏疾病患者和正常人的五个元音信号不同频段的非线性特征都有显著差异,为进一步进行五脏疾病患者的分类预测奠定了基础。最后利用VC++6.0平台初步实现了中医声诊系统,可以通过实时采集以及读取已有信号进行特征的提取与分类识别,提取的特征参数包括线性及非线性特征,分类方法采用了SVM分类器,基本实现了信号处理的一般功能。更多还原

【Abstract】 There are some limitations on the speech signal processing based on different traditional segmentation method, conserdering the mechanism of speech, speech signal is a kind of nonstationary and random process, and the characteristics of it is changing as the change of time. And the traditional speech signal processing and identifying technology based on linear system theories are difficult to develop. Firstly non-linear test using alternative method was done to test the auculatation signals, and the conclusion that there are nonlinearity in the signals was made. Based on this conclusion, two parameters Apen and SampEn which measure the complexity of the signals were used to analyzed the signals. Firstly the relationship between the frequency, amplitude, phase, signal-to-noise(SNR) and the value of Apen were analyzed using some typical signals. The combinaiton among the apen, sampen and wavelet packet analysis were proposed to analyze the signals. Then statistics for these feature were made, Finally SVM was chose to classify the samples, and a impressed results were obtained.Correlation dimension and the method of choosing suitable parameters were discussed. Firstly three segmentation of a typical signal were made to conclude that the attactor for the signal is a kind of strange attractor and there are chaos in the signals, then a suitable method was chose to get a appropriate parameter in order to calculate the correlation dimension for all the sample signals using G-P algorithm. Then the conclusion that it is easy classify the samples when the embedding dimension is higher.The theories and parameter choosing on recurrence plots(RP) and recurrence quantification analysis(RQA) were discussed. Then different Structure characteristics of RP were explained using some special RP drawings. Some suitable parameters of calculating the RQA parameters were chose using average mutual inforamtion and false nearest neighbor method. Based all these researches, all RQA parameters were calculated, and in these parameters, DET and L_max were discussed specifically. wavelet packet transform and apen, sampan were combined together to analyze the/a,//I,//u,//e,//o/signals of heart diseases patients, liver disease, lung department is disease patients, spleen and kidney disease patients, then make statistical analysis for the signals which is the basement for the following clasificaiton.Finally TCM ausculatation system was preliminarily realized based on VC++6.0, Users can collect real-time signal and open the existing wav files to do analysis. Users can extract linear and non-linear features in order to recognize what kind of symptoms they belong to. And the system realize the fuction of signal processing basically.更多还原