Innovation Neural Component Analysis for Monitoring Nonlinear and Dynamic Processes

Non-linearity and dynamics are very common in contemporary industrial processes, but the performance of traditional PCA methods is poor in these fields. In recent years, many scholars have proposed lots of methods to solve these problems. For dynamic processes monitoring, scholars expand the dimension by adding historical data to the original data, so that the established model contains time series information, but the actual monitoring performance is not satisfactory. In the field of nonlinear process monitoring, the kernel method is one of the effective strategies, but there are also many defects such as inconvenience of nuclear parameter setting. In this work, we propose a new monitoring algorithm-innovation neural component analysis (INCA). First, the dynamic and static structures are separated through the Autoregressive model. Then considering the effectiveness of the artificial neural network for nonlinear feature extraction, it is adopted to extract the features, which contain the main information, to construct statistics and finally realize the effective monitoring of dynamic nonlinear process. Simultaneously, we designed a set of mathematical experiments to verify the effectiveness of the method.