Manifold Learning Algorithms Based Non-Gaussian Process Monitoring And Its Application To Chemical Process Monitoring

With the rapid development of national economy and the remarkable improvement of human living standard, the society makes an increasing demand to the manufacturing of process industry, so its security and stability should not be underestimated. Ensuring production safety and improving product quality is a major problem faced by the process industry. The process monitoring technology can deal with the problem. Owing to the applications of computer control systems and intelligent instruments, a large number of data are acquired. We can monitor the operating state through the statistical analysis of these data, and it has been a research topic. Multivariate Statistical Process Monitoring (MSPM), as an important data-driven monitoring method, has been widely concerned in academia and industry.However, the traditional MSPM methods have many limitations, such as the linear relationship and Gaussian distribution should be obeyed. The real industry shows the nonlinear relationship and non-Gaussian distribution between the data variables due to various external influences. Through analyzing the practical problems in industry, the main research work on the basis of others are following:In traditional monitoring methods, the nonlinear structure was damaged when modeling, so the monitoring performance was not good. This paper proposes a method based on Negative Selection Algorithm (NSA) to improve monitoring performance. Firstly, the normal data are dealt with dimension reduction method of Maximum Variance Unfolding (MVU). Then the Negative Selection Algorithm is used to get the hyper-sphere group model directly from the low-dimensional manifold. The monitoring is realized and the smooth operation is ensured. At last, the simulation based on TE model is carried out, and the result shows the proposed method has a better monitoring performance compared with other methods.Support Vector Data Description (SVDD) could face the problem of dimension disaster when dealing with large numbers of samples, but it has advantage in handling with small samples. So a process monitoring method based on LTSA-Greedy-SVDD is proposed. Firstly, Local Tangent Space Alignment (LTSA) algorithm is applied to extract the low-dimensional manifold. Then the Greedy method is used to get effective samples to be modeled, so the computational complexity is highly decreased. At last, the simulation based on TE model and the application show the effectiveness of the proposed method. The industrial processes have characteristics of high-dimension and mixture of non-Gaussian and Gaussian distribution. A method based on combined index is proposed to solve these problems. The non-Gaussian&Gaussian strategy is applied to extract process information to get statistical model. Afterwards, the weighted strategy is applied to compute the new statistic. The method based on combined index is confirmed to get good monitoring performance through numerical system. At last, the advantages of the method are proved through the TE model simulation and the application of industrial process.