多方法融合特征提取在不同时序分类中的应用研究

发布时间：2018-04-03 11:44

本文选题：时序分类　切入点：特征提取　出处：《哈尔滨商业大学》2017年硕士论文

【摘要】：时序数据普遍存在于许多重要应用领域,其中大多数问题的解决可归属为时序数据挖掘问题,时序分类是时序数据挖掘的重要研究内容之一。衡量分类性能好坏的一个重要指标是分类准确率,而提高分类准确率的方法主要有两种:一是改进分类器;二是改进时序分类特征提取方法。针对时序分类问题,通过改进分类特征提取方法来提高分类准确率,提出一种多方法融合的时序分类特征提取方法,应用多个仿真实验验证了该方法的性能,并将其应用于不同领域的时序分类问题中,提高了时序分类问题的分类准确率。在对时序数据特点深入分析的基础上,将小波、分形和统计三种方法相结合,提出一种多方法融合的时序分类特征提取方法。首先应用小波对原始时序数据去噪,并对去噪重构后的时序数据进行分解,计算各尺度空间上的平均高频系数构成特征向量作为第一部分时序分类特征;然后在多尺度下对去噪重构后的时序数据进行多重分形分析,依据具体时序数据的特点和分类的需要,提取相关多重分形谱参数作为第二部分时序分类特征;再次根据不同领域时间序列的特点,提取相关时序统计特征作为第三部分时序分类特征;最后对上述小波、分形、统计方法提取的特征进行综合分析,结合时序的自身特点和实验结果,选择确定最终的时序分类特征。其次,通过不同领域时序分类数据的仿真实验验证了该方法的性能。使用UCI数据集中的Japanese Vowels数据和Synthetic Control数据分别进行仿真实验,通过与其他特征提取方法的多角度比较分析,验证了本文方法的有效性和优越性。最后,对多方法融合特征提取在不同时序分类中的应用进行研究,将该方法应用于不同时序问题,验证了该方法在不同时序分类问题中的实用性和优越性。将该方法应用于EEG信号分类,对来自科拉罗多州发布的EEG脑电信号分类问题展开研究,通过与以往使用相同数据进行特征提取的分类结果比较,说明了该方法在所有分类下的分类准确率均明显高于以往方法的分类准确率;同时与其他两种方法结合进行特征提取的分类结果对比,从结果可见,使用该方法进行EEG分类特征提取在所有分类下的分类结果均高于其他两种分类特征提取方法结合对应的分类结果。然后将该方法应用于手部动作表面肌电信号分类,针对两通路6个类别对应的六种基本手部动作表面肌电信号的时间序列数据,应用该方法实现了分类,从多角度、多层次对分类结果进行了比较分析,结果说明该方法的分类准确率不仅从整体上可以达到较高的分类精度,从各个动作分类上也能得到很好的分类结果。对多方法融合特征提取方法在不同时序分类中的应用作了初步探索。
[Abstract]:Time series data generally exist in many important application fields, and most of the problems can be attributed to time series data mining. Time series classification is one of the important research contents of time series data mining.The classification accuracy is an important index to measure the classification performance, and there are two main methods to improve the classification accuracy: one is to improve the classifier; the other is to improve the temporal classification feature extraction method.In order to solve the problem of temporal classification, an improved feature extraction method is proposed to improve the classification accuracy. A multi-method fusion method is proposed to extract the feature of temporal classification. The performance of the method is verified by several simulation experiments.It is applied to the time series classification problem in different fields, which improves the classification accuracy of the time series classification problem.On the basis of deeply analyzing the characteristics of time series data, this paper proposes a multi-method fusion time series classification feature extraction method by combining wavelet, fractal and statistical methods.Firstly, the original time series data is de-noised by wavelet transform, and the reconstructed time series data is decomposed, and the average high-frequency coefficients in each scale space are calculated as the feature vector of the first part of time series classification.Then multifractal analysis is carried out on the time series data after de-noising and reconstruction in multi-scale. According to the characteristics of specific time series data and the need of classification, the related multifractal spectrum parameters are extracted as the second part of the time series classification features.Thirdly, according to the characteristics of time series in different fields, the statistical features of correlation time series are extracted as the third part of time series classification features. Finally, the features extracted by the above wavelet, fractal and statistical methods are analyzed synthetically.Combined with the characteristics of time series and experimental results, the final time series classification features are selected.Secondly, the performance of this method is verified by simulation experiments of time series classification data in different fields.The Japanese Vowels data and Synthetic Control data in UCI dataset are used to carry out simulation experiments, and the effectiveness and superiority of this method are verified by comparing and analyzing with other feature extraction methods.The results show that the classification accuracy of this method is obviously higher than that of the previous methods, and the classification results are compared with the other two methods, and the results can be seen from the results.The classification results of EEG classification feature extraction under all classification conditions are higher than those of other two classification methods combined with corresponding classification results.Then the method is applied to the classification of EMG signals in hand action. According to the time series data of six kinds of EMG signals corresponding to six kinds of basic hand actions, the method is applied to classify the EMG signals from different angles.The results show that the classification accuracy of this method can not only reach a higher classification accuracy, but also obtain a good classification result from each action classification.The application of multi-method fusion feature extraction in different time series classification is preliminarily explored.
【学位授予单位】：哈尔滨商业大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP311.13

【参考文献】