Patterns classification of nonlinear multi-dimensional time series based on manifold learning

The multi-sensor signals in industrial process is essentially a nonlinear multi-dimensional time series, so the process fault diagnosis can be implemented by pattern classification of nonlinear multi-dimensional time series. To conquer the effects of nonlinearity, correlation and high dimensionality in the time series, a supervised locally linear embedding (S-LLE) based method is proposed in this paper, in which the locally linear embedding algorithm is improved via the label information and a linear propagation method is applied to deal with the out-of-sample problem. Subsequently, the classifier can be designed by using support vector machines (SVM) and k-nearest neighbor algorithm (knn) in the low dimensional space. The proposed method can greatly preserve the consistency of data local neighborhood structure, effectively extract the low dimensional manifold features embedded in the multi-dimensional time series, and obviously improve the performance of the pattern classification. The experimental results on Tennessee Eastman (TE) process demonstrate the feasibility and effectiveness of the proposed method.