Data-based Ensemble Methods For Complex Systems Modelling

Modeling is an important method to analyze systems and solve problems. As the complexity of process is on the increase, more and more experiential modeling methods based on data are adopted. Many physical systems exhibit significant nonlinear behavior or have high dimensions. For most modeling methods, it is too hard to achieve the expected performance, or the obtained models are too complicated to analyze or use farther, for example it is difficult to make stability analysis or design the controllers with those complicated models. So new modeling methods are needed, and ensemble method is just one. This dissertation deals with the applications of ensemble methods to complex system modeling (including static system modeling and dynamic system modeling), and the following problems are discussed in detail:1. Ensemble methods are surveyed in three areas, which are methods for constructing ensembles, theoretical research of ensemble methods, and applications of ensemble methods.2. Applications of ensemble methods to static system modeling are studied. First, the AdaBoost.R algorithm is introduced and modified to fit for the small sample better, which is called AdaBoost.Rm method. And the traditional PLS algorithm is modified so that it can be called by the AdaBoost.Rm directly. Then an novel ensemble modeling method, named the LS-Ensem algorithm, is presented by extending the idea of gradient descent search in function spaces. Theoretical analysis shows that the LS-Ensem algorithm converges exponentially, and yields an integrated model with good generalization property. The relationship between the number of iterations and the generalization error of the moedl is also derived. A simulation validates the theoretical results, and shows that the LS-Ensem algorithm can avoid overfitting effectively.3. A PWARX system modeling method exploiting the combined use of search, clustering, pattern recognition, and linear identification techniques is proposed for dynamic system modeling. In this PWARX modeling method, a search strategy using the empirical covariance matrix of the model parameters as the evaluation criterion is applied and the rough clusters are obtained. A refinement procedurecomprising four parts is applied to regulate the initial clusters and the final clusters are decided, where the four parts are the mergence of the similar clusters, assignment of the undecided data, reassignment of misclassified data, and discarding the small clusters. Then the partition of the PWA system and the parameters identification of each submodel can be done easily. This method does not require to fix a priori the number of affine submodels. Two simulations validate the efficiency of the modeling method.4. We apply the ensemble modeling methods proposed in the thesis to two real processes. One is the soft measurement of the gasoline octane number (static system modeling), and the other is the identification of the electronic component placement process in pick-and-place machines. Results of experiments validate the efficiency of the ensemble modeling methods.