Research Of Multimode Industrial Process Monitoring Method Based On Locally Linear Embedding

With the expansion of the scale of modern industrial and the complication of the process flow, process safety and product quality are important issues that should be paid great attentions by enterprise; process monitoring can be employed to solve those two aspects. According to the wide use of the distribution control system in industrial processes, large amounts of data were collected, the multivariate statistical process monitoring (MSPM) methods have been widely applied to solve the fault monitoring.The traditional MSPM methods are limited to Gaussian, linear, stationary and single mode processes. However, in the actual industrial process, the process data is so complicated that its distribution is hardly ascertained. And production processes often have not only a stable working condition, most of the production processes are multi-mode process. This dissertation develops the research based on the predecessor’s work, according to the complex distribution of the process data, the main research contents are listed as follows:(1) Dealing with multi-mode production process, the traditional statistical analysis methods neglect the correlation between modes. In this dissertation, from the global analysis of the multi-mode, the common underlying characteristics information among different modes is extracted by manifold learning algorithm:LLE, which is called common subspace. After that, the specific part of each mode is obtained. The common part is the similar variable correlations over modes, and the specific part is the correlations which are not shared by all modes, different KPCA model are established respectively. By the analysis of monitoring the common part and the specific part respectively, this method can give a better understanding of multimode process and a good monitoring effect. When working mode changes, the common part model and the corresponding specific part model are selected for monitoring. The fused magnesium furnace industrial process monitoring simulation results prove the feasibility of this method.(2) Production data of actual process usually have characteristics of nonlinearity and non-gaussian, principal component analysis only involves the second-order feature of signal data, not considering the higher-order statistical characteristics, so the higher-order redundant information are still likely to exist between the transformed data. Moreover, PCA method only removes the correlation between data and doesn’t give the corresponding analysis of the independence of data, which makes PCA and modified PCA method do not work well in above-mentioned case. KLLE-KICA method is proposed to solve this problem. The new method is applied into the cold rolling continuous annealing process. The simulation result of the process monitoring shows this method can reduce the false-alarms and improve the accuracy of the fault monitoring.