【作者】 蒋刚；

【导师】 肖建；

【作者基本信息】 西南交通大学 ， 交通信息工程及控制， 2006， 博士

【摘要】 论文主要对机器学习问题、核机器方法、小波核机器、模糊小波核机器等内容进行探讨和分析,构建了几种核机器模型,理论分析和实验结果验证了它们的可行性和有效性。主要研究内容如下: 处理大规模样本时,支持向量机难以满足实时性要求,针对这一问题,提出了一种支持向量预提取方法:先用核感知机模块提取准支持向量,然后将所得结果输入到支持向量机中进行二次处理。核感知机采用的是简单的迭代结构,即使在样本规模较大时,花费的时间也很少;此外,准支持向量的数目可以通过设定阈值进行控制。在一定精度要求下,能从很大程度上提高数据的处理效率。由于核函数和误分界的引入,在综合使用支持向量机的基础上,能处理线性可分、非线性可分、非线性不可分带噪声数据以及回归等问题,理论分析和实验结果较好地验证了这一结论。 对非平稳信号进行处理时,信号细微特征的提取非常关键。论文尝试将小波技术、主分量分析及核方法相结合,用于处理这类信号。对采用小波基构建核函数的可行性进行了探讨,证明了它满足Mercy条件及其在Hilbert空间具有再生性的命题,以此为基础,结合主分量分析,探讨了小波核机器的构建方法,构造出一种核机器模型,并作了实例仿真。实验结果表明,复Gaussian小波核和复Morlet小波核的性能大致相当,它们都优于常规的高斯核和多项式核,初步展示出该方法的可行性和优越性。 对模糊逻辑和小波技术的相关理论进行探讨和分析,构建了一种模糊小波容许核函数,并与支持向量机结合,构造出一种核机器模型,对该模型的一致逼近性作了证明。在此基础上,提出了一种模糊小波支持向量核机器方法FW-SVKM,对参数的选择与预测结果的内在关系作了较为详细的分析,与三层神经网络ANN进行短期峰值负荷预测的对比实验,结果表明FW-SVKM优于ANN,具有较大的实用价值和较好的应用前景。 针对学习机器在参数较多时,优化时间过长、效率过低,不利于工程应用的问题,提出了一种多参数同步优化策略。实验结果表明,该方法在实际应用中是行之有效的,能大幅减少多参数模型的优化时间,增强核机器方法的实用性和有效性。更多还原

【Abstract】 This paper dealed with machine learning, kernel machine method, wavelet kernel machine and fuzzy wavelet kernel machine technology detailedly. Several kinds of kernel machine models were constructed, and have been applied to non-stationary signals processing. Theoretical analysis and implementation results show their validity and feasibility. Some important issues have been discussed in this paper as follows.It is well known that general SVM (Support Vector Machine) costs too much time on large scale data sets. As a valid solution, support vectors pre-extraction method has been discussed in this paper. Kernel perceptron firstly has been used to extract quasi-support vectors. And then, quasi-support vectors were input to standard SVM to process accurately. This method takes advantages of the high speed of perceptron for its simple iterative structure. Perceptron costs fewer time than general SVM, especially on large scale data sets, and can control number of quasi-support vectors easily by a threshold variable, much time will be economized in latter process. Some special technologies, such as kernel function and error boundary, etc, have been adopted to conduct linear separable, nonlinear separable, nonlinear unseparated model recognition and regression questions effectively.Generally, imperceptible features are very important in non-stationary signal processing. Some kinds of complex methods were discussed in this paper, which combined wavelet, Principal Component Analysis (PCA) and kernel function teconology. Wavelet kernel function was constructed after proofs of propositions, that it can meet Mercy condition needs and has reproduction feature in Hilbert space. A kind of kernel machine model was presented and some numerical simulation experiments were applied to validate its correctness. Experiment results show that complex Gaussian wavelet kernel almost has the approximate performance as complex Morlet wavelet kernel, however, excel than general Gaussian kernel and polynomial kernel.Fuzzy and wavelet technology were adopted to construct a kind of fuzzy wavelet kernel function. After that, a kind of kernel machine model based on SVM was built and proofs of consistant approximation were shown immediately. Based upon these theories foundation, Fuzzy Wavelet Support Vector Kernel Machine (FW-SVKR) was formed. The close relationship between parameters and forecastresults was expatiated later. Contrast experiments between FW-SVKR and Artificial Neurial Network (ANN) show that the former was superior to the latter in electric power system load forcasting, and seems to have more applied value in this domain.Theory analysis suggested that much more time should be cost when training a multi-parameters model, which formed the mainly obstacle in application. Aimed at this question, a kind of new technology named multi-parametes synchronous optimization method was proposed. It can save a lot of time when training parameters, and enhance applied value remarkably. Experiment results show its advantages in application.Along with the boost of country economy, city traffic block should be solved urgently. Based on analysis of features of city traffic flow, some kind of kernels such as general kernels, compound kernels and fuzzy wavelet kernels, have been adopted to do realtime traffic flow forecast. Contrast experiment results show the different performance of those kernels, which can help to improve the city traffic control power effectively. Another application of kernel machine method was discussed in this paper, complex kernel function method and support vectors preextraction technology have been adopted to retrieve the Oceanic Chlorophyll-a Concentration in SeaWIFS data sets. Further more, ANN altorithm and twelve kinds of empirical algorithms have been adopted, too. Contrast experiment results show that, retrieve precision with complex kernel function method is higher than that of other algorithms;it seems to be more suitable in this domain.更多还原