Research Of Fault Isolation For Nonlinear Processes

In modern industrial process, the accurate way of process monitoring and fault diagnosis is the key to ensure process safety, increase production efficiency and improve product quality. As the continuous development of detection and computation technology, and their widely application in the industrial processes, the scale of collected data becomes larger and larger. How to use the latent information of historical data to guide the production and practice and to improve the performance of process monitoring has become a very important research topic in the field of process control.After summarizing the development of multivariate statistical analysis based process monitoring and fault isolation methods, this thesis studies the monitoring and fault isolation of nonlinear process. The main work of this thesis including:(1). Aimed at the defect of traditional PCA based reconstruction method, that is, it does not take the normal operating data into consideration when builds the reconstruction models, so that it cannot distinguish the faulty information from the normal one in the faulty samples with its fault directions. On the basis of traditional reconstruction method, the relationship of faulty samples and normal operating data is analysed deeply, in order to extract the faulty directions that cause the out-of-control statistic for each class of fault. Reconstruct the faulty samples using the reconstruction model of each fault in turn. Only the current fault model can remove the fault information from faulty samples accurately and eliminate the alarming of statistic, the fault isolation is therefore performed. The simulation results of fused magnesium furnace industrial process prove the feasibility of this method.(2). The PCA based reconstruction method is too complex, it need to build the reconstruction models for all kinds of faults firstly, and need to perform the fault reconstruction with these models in turn. To solve the above problems, the KLSR based fault isolation method is proposed, in order to solve the fault isolation problem with classification algorithm. This method projects different faulty samples to its corresponding targets by multivariate linear regression analysis in the feature space. In the training process, the correction factors are used to optimize the weight vectors. Since too many training samples would lead to an increase in the computation and storage of kernel matrix, the coefficient matrix with sparsity is acquired by introducing the L2,1 norm in the objective function, in order to extract the samples more important in building KLSR model. The validity of this method can be proved by the simulation results of iris data set and fused magnesium furnace industrial process.