Clustering-Based Multiple Modeling Approach And Its Application In Soft Sensor

In view of multiple models can significantly improve model's estimation accuracy and generalization performance, combining with actual industry application background, it is used to construct multiple models of crystal tower and realize online monitoring of process variables.In sorts of multiple modeling method, clustering-based ones get the most widespread concern. However, in the traditional clustering algorithm, many problems, such as how to decide cluster numbers and centers, are still unsolved. And it indirectly constrains the development of multiple models. At the same time, most clustering algorithms are sensitive to abnormal sample points what greatly reduces the effectiveness of clustering results. Moreover, traditional clustering algorithm has inherent shortcomings which only uses the input sets, while ignores the enormous impact of output set in the process of clustering sample set results in determining the quality of clustering. At last, as the most important part of the multi-model, sub-model will have a direct bearing on the accuracy of multiple modes.To solve the problems mentioned above, the paper improves clustering algorithm and modeling approach from the following four aspects so as to establish effective multiple models.In view of the traditional cluster algorithm's shortcomings of heavily relying on the priori knowledge and initial parameters, a single-parameter adjustment expanding search clustering method which is suitable for arbitrary shape sample distribution is presented. The new clustering method is based on the nearest neighbor algorithm. By definingε-neighborhood of samples and applying expanding search method, the algorithm classifies all associatedε-neighborhood samples into one cluster, and therefore, the work of clustering sample set is achieved. The proposed algorithm is used for clustering sample set, and obtains multiple modeling techniques based on expanding search clustering.The existing of outliers will severely affect clustering results. A multiple modeling method based on manifold clustering with local reconstruction and merging is proposed. In order to restraining the impacts of outliers to clustering results, data set is split into several small disjoint sub-clusters. By reconstructing linear manifold level based on every sub-cluster respectively, it completes the work of clustering through merging sub-clusters who are not only closer but also belonging to the same manifold level. Meanwhile, Support Vector Machine is used to construct regression model in each sub-class and multiple models is obtained finally.The traditional clustering algorithm can't deal with incomplete information very well. A multiple modeling approach based on secondary data partition is presented. The proposed method carries out the secondary classification on the sub-class by improved rough set classifier which obtains from clustering sample set, so as to eliminate affect of contraction samples on model's accuracy to some extent. Support vector machine is used for building regression sub-model on each subclass, and finally obtain the soft-sensing multiple models. At the same time, in view of possible appearance of unbalanced classification problem, the improved weighted rough set classifier is adopt to improve above multiple modeling method further more so that significantly boosts the classification accuracy of classifier and ensures the reliability of multiple models.The accuracy of final multiple models directly depends on effect of sub-models. A novel local penalized weighted kernel partial least squares algorithm is presented. The proposed method map original inputs into a high dimensional feature space so as to realize the linear treatment of nonlinear problems, and meanwhile, partial least squares algorithm is used to extract the principal components. According to local learning theory, a local penalized weighted least squares regression model is constructed based on the new data set, which is formed by the principal component, in order to differentially treat the contribution of each sample value, reduce the model sensitivity of abnormal data and optimize the model parameters. In view of multiple models can improve the estimated accuracy and generalization of model, the expanding search clustering algorithm and local penalized weighted kernel partial least squares are used to cluster sample set and establish the regression sub-models on corresponding sub-cluster respectively. Finally, a soft sensor system based on multiple models is obtained.The proposed algorithms are used in a soft sensor model for the Bisphenol-A productive process, and the result of simulation shows the effectiveness of the algorithm.