中文摘要：

针对建模数据中包含噪声和离群点会降低相应软测量模型准确性的问题，提出一种结合2层变量空间分析的预处理方法．用多变量修剪法在原始变量空间预处理；并提出支持向量聚类（SVC）的预处理方法，将建模数据映射到高维特征空间，构造一超球体来排除离群点．SVC无需像传统预处理方法假设数据服从正态或近似正态分布，更符合实际的高炉过程．预处理后的数据再用支持向量回归建立软测量模型．在一工业高炉铁水硅含量的建模和预报实验结果表明，所提出方法能够更有效排除离群点。且提高了支持向量回归模型的鲁棒性和预报性能．

英文摘要：

A novel preprocessing method integrated two-level spaces of process variables was proposed to overcome the effect of noises and outliers in modeling data, which can degrade the performance of the related soft sensor model. The multivariate trimming was first utilized for preprocessing in the primary space. Then, a support vector clustering （SVC） strategy was proposed for outlier detection. The main idea of SVC is to map the data into the feature space and then to find a hyper-sphere with the minimal radius that contains most of the mapped data. Different from most of traditional preproeessing methods, SVC dose not assume that data are distributed （approximately） normally, and thus is more suitable for industrial ironmaking processes. After SVC-based preprocessing, a support vector regression soft sensor model was built. An experiment study for an industrial blast furnace is investigated and the results show its superiority, including efficient outlier detection, a more robust support vector regression model and better prediction performance, compared to traditional approaches.