非线性多维时间序列模式分类的新方法

Novel method for patterns classification of nonlinear multidimensional time series

多变量非线性时间序列的模式分类是在工业过程领域广泛存在的问题,结合流形学习和支持向量分类机的特点,提出了解决该类问题的一个新方法。该方法应用核化流形学习算法K-Isomap,将高维非线性时间序列映射到低维特征空间实现维数约减,在低维特征空间中采用支持向量机设计分类器实现非线性时间序列的模式分类,该方法充分利用核化流形学习的特点,得到了较好的模型性能。应用该方法对Tennessee Eastman(TE)过程的故障分类进行了实验分析,结果表明该方法的有效性。

Pattern classification from nonlinear multivariate time series is an important problem in process engineering.This paper introduces a generic approach to detect patterns and identify their class incorporating manifold learning and support vector classifier.K-Isomap,a kernelized manifold learning algorithm,is employed to project multidimensional nonlinear time series onto low-dimensional feature space and realize nonlinear dimensionality reduction.Pattern classifier is designed to identify the pattern of nonlinear time series based on support vector machines in low-dimensional feature space.This method takes the advantage of the kernelized manifold learning algorithm and obtains better performance.Experimental results on Tennessee Eastman（TE） process demonstrate the validity and effectiveness of the proposed method.