Weighted Wasserstein Distance-based Improved Serial Principal Component Analysis for Incipient Fault Detection of Complex Industrial Process

Although serial principal component analysis (SPCA) has emerged as one effective process monitoring technology, it doesn’t fully mine the probability distribution information, thus having bad performance on detecting incipient fault. To handle this issue, this paper proposes a weighted Wasserstein distance-based improved serial principal component analysis (WWDSPCA) method to detect the incipient fault of complex industrial process. Firstly, serial principal component analysis is performed to extract both linear and nonlinear principal components. Then, Wasserstein distance (WD), highly sensitive to probability distribution change, is used to construct WD components. Further considering the difference in the fault information carried by the WD components, a weighting strategy based on kernel density estimation (KDE) is adopted to emphasize the WD components with more fault information. Simulation results of a continuous stirred tank reactor (CSTR) system verify the effectiveness of the proposed method for detecting incipient fault.