Nonlinear process fault detection method under noise environment using KPCA and MVU

Actual chemical process monitoring data have strong nonlinear behavior and are easily disturbed by random noises. A novel kernel principle component analysis (KPCA)-maximum variance unfolding (MVU) based fault detection method for nonlinear process under noise environment is proposed by combining KPCA and MVU feature extraction algorithms. In the dimension reduction process of nonlinear noisy data, local KPCA method is applied to identify and eliminate the noise in the process data in the neighborhood of sample points; and then the nonlinear principal components are extracted in the input data space. Next, under the condition of keeping the Euclidean distances between neighbor points unchanged, MVU is used to map the original high dimension data space to a low dimension embedding space while preserving the data global geometric structure via coordinate rotation and translation transformation. Simulation results of TE process under noise environment and experiment results of acrylonitrile polymerization process show that the improved KPCA-MVU based fault detection model can improve the feature extraction performance of standard KPCA and MVU algorithms for nonlinear noisy data, and effectively enhance the robustness against noise. ©, 2014, Science Press. All right reserved.