广义KKT约束的增量支持向量机质量建模研究

Research on Quality Model of the Generalized KKT Constraint Incremental Support Vector Machine

传统支持向量机在求解大规模样本训练时,解二次规划问题占用大量内存,收敛速度慢;并且当有新样本加入时,所有样本需要一同重新训练,浪费大量时间,进而限制了其使用范围。针对上述缺陷,提出了广义约束的增量学习支持向量机回归模型的动态质量建模方法,利用KKT条件及时淘汰对后续训练影响不大的样本,同时保留了含有重要信息的样本。以带钢热镀锌生产中锌层质量模型为研究对象,建立生产过程参数与质量结果之间的回归模型。用某钢厂带钢热镀锌的实际生产数据进行验证。结果表明,该算法在保证预测精度的同时,有效的提高了学习速度及对大样本学习的能力并降低了内存占用。

When training large-scale sample, there were some disadvantages of traditional support vector machine （SVM）. The solution of quadratic occupies a lot of memory; the convergence speed is slow; all samples should be training together when new samples joining which waste a lot of time. In this paper, a generalized constraint SVM of dynamic quality modeling method was proposed. This method delete the samples timely which non. effect to the training and retain the samples which contains important information through KKT conditiotu As research object, the production of galvanized steel with zinc quality model was used to build the regression model between the production process parameters and quality. An actual data of strip steel galvanized was used to verify the method. The result shows the prediction accuracy, the learning speed and learning ability of large sample was effectively improved and the occupied less memory.