Process Monitoring, Quality Prediction And Abnormal Variables Detection Based On Improved Opls

It has been a persistent goal for process industry to increase the quality of products and ensure the safety of operation, and how to predict the quality, monitor the process, find and restore fault timely have been the research focus accordingly. With the fast improving of computer technique and so on, abundant or even redundant data can be obtained, which includes information closely related with operation states and final quality. It will be much more meaningful if we can make full use of these data.In this paper, Windows-mean OPLS (Orthogonal partial least squares) is proposed for quality prediction, process monitoring and abnormal variables detection based on multivariate statistical methods and data mining technology. Finally, it is introduced to continuous annealing process. The main contributions of this dissertation can be summarized as follows:The quality prediction, process monitoring and abnormal variables detection based on PLS (partial least squares) and MPLS (multiway partial least squares) are described according to literatures both home and abroad. OPLS is introduced to make the model easier to interprete.The determination of number of orthogonal components is an important issue in OPLS algorithm.An improved OPLS algorithm is proposed to provide an effective method to determine the number of orthogonal components. At the same time, Windows-mean OPLS is put forward to deal with problems of numerous variables for modeling and unequal length of data. It performs OPLS on the average trajectories of variables based on the width of the process segment. It is employed to predict the quality of the steel in continuous annealing process. A method to detect abnormal variables grounded on control limits of contribution and ratio of over-limit is presented. It makes it more reliable to restore fault in process timely.The continuous annealing process has characteristics of many variables and high frequency of sampling. The hardness of steel is difficult to be measured online. It is obtained mainly by offline analysis, and the control is always operated according to experiences, which cause blindness and posteriority. Windows-mean OPLS is applied to the quality prediction, process monitoring and abnormal variables detection. Experiments verify its validity.