Fault Detect Based On Quality-relevant Kernel Independent Component Analysis And Manifold Regularization With Feature Extraction

The feature of complex in the modern industrial control system make the information of monitoring data become more rich,so the demand with safety,efficiency and sustainability for the process control is also higher.In order to detect the fault information quickly and accurately and predict the quality information for production preferably,the method of fault detect is getting more and more attention.This paper firstly introduced the background of fault detect and its significance,then presented the development of process monitoring and multivariate statistic process and current situation at home and abroad,Secondly introduced the nuclear independent principal component method,non-gaussian maximization method,kernel partial least squares,manifold regularization method and learning method based on knowledge information.We proposed the quality-relevant method of QKICR based on the combination nuclear independent principal component with kernel partial least squares.The QKICR improve the objective function of the KICR to make independent component selected consider the problem of independence and the correlation of independent component and quality variable.This greatly improves the performance of the quality prediction.We also come up with a study of fault detect based on manifold regularization with feature extraction which consider the manifold regularization,the knowledge learning method and the multi-mode characteristics,it put a few data with expertise information and mass of data with no-expertise information into the learning of manifold regularization to make the better effect in the process of fault detect.These methods are applied to the continuous annealing process and Tennessee Eastman process for fault diagnosis and quality prediction respectively.Some methods are used for comparing.The results indicate that the proposed methods are effective.