Study Of Multivariate Statistical Method Based On Probabilistic PCA On The Process Monitoring

After long time running, there are some changes in any producing system, which will unavoidably influence the quality of products and even result in great accidents. Therefore, the traditional methods entirely based on manpower are outdated and can not satisfy the complicated desire of quality control. Having no use of complex mechanism model, multivariable statistical process monitoring method can monitor process through extracting important information from raw data using statistical method and then transforming them into several significative indices. The method not only takes sufficient use of the existing information and is well realizable, but also greatly reduces the procedure of process monitoring system, which has been developed more than thirty years, in which lots of research results have been acquired and applied widely.Probabilistic principal component analysis(PPCA) firstly assume the distribution of latent variables and error vector, secondly evaluate the generative model by the expectation and maximization (EM) algorithm, so it can detect fault effectively and performs on-line fault identification, which make it attractive both in industry and academia. However PPCA is a linear way, the preconditions of its application are that process are normally distributed and no auto-correlation among them. But most of industrial process are complicated and always violate the preconditions, so the PPCA-based method behaves unsatisfactorily.Aiming at the disadvantages of the PPCA-based method, the main contributions are as follows:1. The dissertation proposes an improved monitoring way which monitors the norm of whiten measurement variables and perform on-line fault identification by monitoring every whiten variable, so the load of monitoring is reduced. At last PCA and PPCA are compared in monitoring process of chemical separation.2. Aiming at the strong dynamic characteristic of the industrial process, using EM algorithm, the dynamic PPCA model is built to cope with the data matrix extend by time series. According to the technique, static PPCA can be extended to monitor dynamic multivariate process, and auto-correlation among process variables is effectively eliminated.3. On the monitoring of nonlinear process. The dissertation propose a dynamic kernel PPCA, it maps the compressed data matrix extended by time series into high-dimensional space by kernel function, then PPCA can be used to monitor the linear mapped value or process variables. The method is used in the continuous and reforming heating stove system process, and the results verify its effectiveness.