中文摘要：

针对间歇过程数据的批次不等长和强非线性的特点,结合核偏最小二乘和核熵分析,提出了多向核熵偏最小二乘（multi-way kernel entropy partial least squares,MKEPLS）的过程监测及质量预测方法.该方法将三维历史数据沿新的展开方式展开,克服了批次不等长和数据缺失的问题,通过核映射将过程数据从低维输入空间映射到高维特征空间,实现变量之间非线性相关关系的线性转换,解决了数据的非线性特性;根据核熵的大小将特征值和特征向量进行排序并对数据进行降维,弥补了MKPLS方法只按照数据特征值的最大化进行降维的不足.同时,引入核特征提取算法降低核空间的计算量,使其能够在线应用.数值实例和实际工业过程数据的验证效果表明：MKEPLS方法不仅能对故障进行有效监控,提高故障的报警率,同时还能对最终产品质量进行预测.

英文摘要：

Aiming at the nonlinear, unequal length of the data in the batch processes, a multi-way entropy partial least squares （MKEPLS） process monitoring and quality prediction method was proposed combining with kernel partial least squares and kernel entropy analysis. The method solved the unequal length and missing data problem by expanding the original three-dimensional data in a new unfolding way, and the nonlinear character of the data could also be solved through the kernel mapping process, which mapped the data from low dimensional input space into a high dimensional feature space to achieve the nonlinear relationship between variables linear transformation. Then, the data dimensionality reduction was conducted according to the kernel entropy eigenvalues and eigenveetors, which made up the shortcomings in MKPLS method. Moreover, the kernel feature extraction algorithm was introduced to reduce the computational kernel space to enable the online applications of MKEPLS. The numerical examples and practical industrial process data performance show that MKEPLS method can monitor the fault effectively, improve the fault alarm rate, and predict the quality of the final product at the same time.