Research Of Support Vector Machine Generalization Ability

On the theoretical basis of Statistical Learning Theory, contrary with other machinelearning methods, Support Vector Machine has its own advantage, so it has wide applicationin soft sensor field already. However, chemistry production process is a complex process withmultiple-conditions, nonlinear, high noise. When Support Vector Machine is used forpredicting unpredictable variables, many problems still will be encountered. Therefore,Support Vector Machine may cover the shortage with the combination of the other techniques.For example, multiple models based on Support Vector Machine and some other intelligentcombination methods can describe production process better and improve the generalizationability of the soft sensor modeling. According to the engineering application background,three methods for soft sensor modeling based on Support Vector Machine (combined withdata mining techniques) are proposed in this thesis. Specific results are as follows:1. A multi-model modeling method based on feature-weighted fuzzy clustering ispresented. This method considers the particular contributions of different features on the basisof traditional fuzzy C-Mean clustering. At the same time, it adjusts feature weights graduallyon the basis of the iterative fuzzy clustering, improves the clustering result.The multi-modelis applied to a soft sensor for components of BPA in a Phenol evaporator outlet, and thesimulation results show that the generalization ability is better.2. A multi-model modeling method based on improved clustering and weighted baggingis proposed in the paper. The method improves clustering result by reducing error dividingprobability with K-neighbors based on traditional fuzzy C-Mean clustering, and the trainingsample set is grouped into several feature sets with correlation analysis. At last, a multi-modelis constructed by support vector machines adaptively according to weighted baggingalgorithm of ensemble learning. Using the proposed algorithm to the soft-sensor model ofBPA component in a Rearrange reactor exports, the result of simulation shows that everyfeature model is assigned with weight reasonably, and the model generalization ability isbetter.3. A combination Support Vector Machine soft-sensor modeling based on SLPP ispresented. The method combines SLPP and combination Support Vector Machine, based onthe idea of multiple knowledge bases mining, makes strengthening dimensional reduction tothe input space. Meanwhile, it considers the weight of sub-model training effect. Finally, itrealizes combination modeling adaptively. The combination model is applied to estimating thecomponents of BPA in a cracking reactor exports, the result of simulation shows that thecombination model is effective.