非线性混沌理论在脑卒中患者声音时间序列中的分析和应用

发布时间：2018-08-17 14:46

【摘要】：脑卒中是一类高发病率,高致死率的疾病。在预测脑卒中发生以及在脑卒中患者康复观察的过程中,并没有很好的客观评价方法,只能通过医生的临床经验。所以本论文通过结合人体声音产生的生理学特点,利用非线性动力学方法分析声音时间序列,提取特征量分析脑卒中患者的大脑的损伤状态,尝试可以找到度量脑卒中患者大脑状态的特征量。为脑卒中患者康复及预防提供客观评价。本文对脑卒中患者的诊断判别方法进行了分析和研究,并最终通过声音诊断技术实现了脑卒中患者和健康人的分类。对声音诊断技术的四个方面(即脑卒中病人声音采集、脑卒中病人声音信号分析处理、脑卒中患者诊断特征量的构造和脑卒中患者分类)进行研究探索,并取得了以下研究结果： 1)提出了通过声音来研究大脑状态的方法,并且从语音产生的神经机制和脑成像机制的角度对利用声音分析大脑状态提供了理论上的支撑。并且结合脑卒中患者实际情况提出了最适合采集的音节。 2)提出了基于声音时间序列的混沌特性用非线性动力学的方法来分析脑卒中患者的大脑状态的方法。对声音非线性时间序列进行相空间重构,分别用互信息法得到的时间延迟和用改进伪最近邻法得到的CAO方法得到的嵌入维数重构相空间和吸引子。最后用小数据法计算声音时间序列的最大Lyapunov指数,提取得到的这些混沌特征量都证明了声音时间序列具有混沌特性。 3)首次提出了利用改进的替代数据法得到一种新的非线性特征量来反映声音时间序列的非线性特征,进而用于反映脑卒中患者的大脑状态。该方法将替代数据法和关联维数相结合得到了新的非线性特征量即归一化方差检测量。这一新的特征量反映了非线性声音时间序列和声音序列的替代数据(不具有混沌特性)在关联维数上间的差异,比单纯的非线性声音时间序列的关联维数更好得反映了脑卒中患者因为脑损伤导致的变异声音的非线性性质。 4)对声音特征量进行模式分类。本文首先对所有声音样本提取声音特征量,包括互信息图的第一个最小值,关联维数,最大Lyapunov指数以及归一化方差检测量。对这些特征量按照健康人和脑卒中病人进行统计对比分析。通过图表的形式生动直观得看出两类声音信号的差异。然后再利用K近邻分类方法对组合特征量进行分类,分类结果表明了新的归一化方差检测量能够提高分类准确度,而提取出的非线性特征量可以用于对脑卒中患者和健康人进行分类。这也为以后用声音时间序列分析度量大脑状态提供了研究方向和基础。
[Abstract]:Stroke is a kind of disease with high incidence and high mortality. In predicting the occurrence of stroke and in the course of rehabilitation observation of stroke patients, there is no good objective evaluation method, only through the doctor's clinical experience. Therefore, combining the physiological characteristics of human voice production, this paper uses nonlinear dynamic method to analyze the sound time series, and extracts the characteristic quantity to analyze the brain damage state of stroke patients. Try to find characteristic quantities that measure the state of the brain in stroke patients. To provide objective evaluation for the rehabilitation and prevention of stroke patients. In this paper, the diagnosis and discrimination methods of stroke patients were analyzed and studied. Finally, the classification of stroke patients and healthy people was realized by sound diagnosis. In this paper, four aspects of sound diagnosis technology (namely, the sound acquisition of stroke patients, the analysis and processing of sound signals of stroke patients, the construction of diagnostic characteristics of stroke patients and the classification of stroke patients) were studied and explored. The following results are obtained: 1) A method to study the brain state through sound is proposed, and it provides theoretical support for the analysis of brain state by sound from the perspective of the neural mechanism of speech production and the brain imaging mechanism. According to the actual situation of stroke patients, the most suitable syllable is put forward. 2) A method of analyzing the brain state of stroke patients by nonlinear dynamics is proposed based on the chaotic characteristics of sound time series. The phase space reconstruction of sound nonlinear time series is carried out. The time delay obtained by mutual information method and the embedding dimension obtained by CAO method with improved pseudo-nearest neighbor method are used to reconstruct the phase space and attractor respectively. Finally, the maximum Lyapunov exponent of sound time series is calculated by small data method. The extracted chaotic features prove that the acoustic time series have chaotic properties. 3) A new nonlinear feature of sound time series is first proposed by using the improved alternative data method to reflect the nonlinear characteristics of the sound time series. In turn, it is used to reflect the brain state of stroke patients. This method combines the substitution data method and the correlation dimension to obtain a new nonlinear characteristic measure called normalized variance detection. This new feature reflects the difference between the correlation dimension of the nonlinear sound time series and the alternative data of the sound sequence (which does not have chaotic characteristics). The correlation dimension of the time series is better than the correlation dimension of the nonlinear acoustic time series, which reflects the nonlinear characteristics of the variant sound caused by brain injury in stroke patients. 4) the pattern classification of the sound characteristic quantity is carried out. In this paper, sound features are extracted from all sound samples, including the first minimum of mutual information graph, correlation dimension, maximum Lyapunov exponent and normalized variance detection. These characteristics were statistically compared with those of healthy people and stroke patients. The difference between the two types of sound signals can be seen vividly and intuitively through the form of a chart. Then the K-nearest neighbor classification method is used to classify the combined feature quantity. The classification results show that the new normalized variance detection method can improve the classification accuracy. The extracted nonlinear features can be used to classify stroke patients and healthy people. This also provides the research direction and foundation for the measurement of brain state by sound time series analysis.
【学位授予单位】：浙江大学
【学位级别】：硕士
【学位授予年份】：2014
【分类号】：R743.3

【参考文献】