【作者】 任世锦；

【导师】 吴铁军；

【作者基本信息】 浙江大学 ， 控制科学与工程， 2006， 博士

【摘要】 尽管数据挖掘技术已经成功应用到很多领域，但是由于测量手段的局限性以及客观对象本身不确定性等原因，在自然和社会科学领域中存在大量待挖掘的数据是不准确、不确定的。如果不充分考虑这些数据的不确定性，不采用合适的数据挖掘方法，那么必然会导致挖掘模型性能下降，甚至是不可接受的。如何从大量不确定性数据中找出隐藏的知识，仍然是人们研究的主要课题。根据待挖掘数据所依据的不确定理论，不确定性数据挖掘方法可分为4种类型：随机数据挖掘、灰色数据挖掘、模糊数据挖掘和区间数数据挖掘方法。本文以国家863项目钢铁生产过程的质量控制为研究背景，为解决生产过程数据不完整、不确定、大数据量的数据挖掘困难问题，提出基于SVM以及核方法的区间数数据挖掘建模方法，并用于钢铁生产过程的质量预测和质量分析。本文主要工作包括以下几个方面：1．提出了两种区间数回归建模方法。针对钢铁工业过程中的实际需要，本文提出两种区间数回归建模方法：(1)基于SVM的区间数输入、区间数输出回归建模方法。该方法把SVM从实数回归分析推广到区间数回归分析，同时继承了SVM的优点；(2)基于SVM的精确数输入、区间数输出的回归分析建模方法。该方法对区间数上下界逼近回归建模的同时还考虑了上下界之间的相互信息。本算法能通过解凸二次最优化问题得到全局最优解，有效地避免了回归模型的上界小于下界的问题。2．提出了基于多尺度径向基小波SVM的精确数输入、区间数输出的鲁棒回归分析建模方法。首先提出并使用基于多尺度径向基小波SVM方法对带有离群点的多尺度区间数样本数据集进行训练，分别得到区间数上下界初始回归模型，初始回归模型不仅能够有效地逼近多尺度样本信号、具有较好的鲁棒性和推广性能，而且在正常样本点处残差较小，在离群点处残差较大；然后以加权M—估计函数作为代价函数，并使用梯度下降法同时调整区间数上下界回归模型参数，且根据间数上下界模型输出关系改变加权系数，消除了离群点对区间数回归模型的影响，解决了区间数模型输出上界小于下界的问题。3．提出了基于SVM的区间数分类建模方法。根据区间数的比较定义，该方法把区间数线性分类器转化为精确数输入的线性分类器，从而能够建立基于SVM的区间数分类器模型。本文通过设计合适的核函数，把区间数数据样本映射到高维特征空间，然后在高维空间中对样本数据进行线性分类，顺利地解决了区间数样本线性不可分问题，克服了现有的区间数分类算法对输入维数敏感以及不适合大量数据样本等问题。4．提出一种基于核方法的区间数模糊聚类算法。本文通过设计合适的核函数，有效地扩大样本模式之间的差异性，提高聚类算法对非对称数据以及混合数据模式的适应能力，且避免了直接在高维特征空间求解聚类问题。采用区间数遗传算法来求解非凸聚类优化问题得到聚类问题的全局最优解，大大提高了聚类质量。该方法解决了现有区间数聚类算法对多种聚类模式以及不对称数据聚类性能不理想的问题。5．以钢铁生产过程为应用背景，阐述了基于区间数的不确定性数据挖掘在实际生产中的应用，并与其它方法相对比。结果表明，本文提出的数据挖掘方法能够克服现有算法的缺点，并在钢铁过程的数据挖掘中取得较好的效果。更多还原

【Abstract】 Data mining approach has been successfully applied into various fields, but in fact many data are uncertain and imprecise in natural and social sciences because of limits of measurement technology and uncertainties of objects et al. The inappropriate data mining methods based on these uncertain data will result in poor, even unaccepted quality of mining models. How to extract hidden knowledge in uncertain and imprecise data is worth researching. According to the uncertain theory on which data mining methods are based, four major approaches have proposed for data mining under uncertainty: data mining for stochastic data, data mining for grey data, data mining for fuzzy number and data mining for interval number.To overcome the shortcomings of data mining approaches for uncertain, imprecise and numerous production data, the thesis takes national 863 projects, i.e. steel& iron production quality control for instance, and proposes SVM and kernel method-based data mining models for interval number, and applies the proposed algorithms to quality analysis and prediction.The main research work is depicted as the follows:1. Two regression analysis models for interval number are proposed. According to practical problems in steel& iron production, two regression analysis models for interval number are presented. (1) A SVM based regression model with interval number input and interval number output. The SVM regression method is generalized from real number domain to interval number domain, and simultaneously keeps the merits of SVM; (2) A SVM based regression model with accurate number input and interval number output. The method takes advantage of the relationship between upper and low limitation regression models simultaneously. The proposed algorithm can prevent lower limitation from exceeding upper limitation by solving convex quadratic optimization problem.2. A multi-scale radial wavelet SVM robust regression model with accurate number input and interval number output is proposed. A multi-scale radial wavelet SVM is presented and applied to train the upper and lower limitation regression models respectively according to multi-scale interval number sample set contained by outliers. The two regression models have good robustness and generalization, and can approximate multi-scale signals effectively. Also the regression residuals of normal samples are small, while those of outliers are large. Then a weighted M-estimator is considered as cost function, and gradient algorithm is adopted to adjust the parameters of upper and lower limitation of regression models simultaneously. The weights and change with the relationships between the upper and lower limitation regression models. So, the effects of outliers can be reduced gradually and the problem that the lower limitation exceeding upper limitation is prevented.3. A SVM-basedclassification method for interval number is presented in this thesis. According to the comparison of interval numbers, the linear classification modele with interval number input is transformed into that with precise number input,and then SVM-based classifier for interval number is constructed. By designing an appropriate kernel function, the interval number samples are mapped to high dimension feature space in which a linear classification model is built. Thus the problem that interval number sample set can’t be separated linearly is solved easily, and the proposed algorithm overcomes the deficiencies of the existing classification algorithms for interval number, such as sensitive to input dimension, suitable for the small number of samples etc.4. A kernel-based fuzzy clustering algorithm for interval number is presented. By designing an appropriate kernel function, the algorithm can effectively cluster asymmetric/mixed data structure of multi-pattern prototypes, and meanwhile avoiding doing this directly in feature spaces. Also, interval number genetic algorithm is designed in this thesis to solve a highly non-convex optimization problem to get the global solution to the clustering problem, greatly improving quality of cluster. The proposed algorithm can overcome a common shortcoming of existent interval number clustering approaches. i.e. the cluster performance is unsatisfying for multi-pattern clusters and unbalanced data structure.5. Several interval-based uncertain data mining methods are used to steel & iron process, and the results of these approaches are compared with those of the existence methods. The results show that the proposed approaches can overcome the defiecencies of the existence methods, and have better results in steel& iron process data mining.更多还原