New modeling method of soft sensor based on Support Vector set-RBF neural network

The soft-sensing technique provides a new approach to detect and control primary variables which are very difficult or impossible to be detected, and is one of the most important research directions in the area of process control. Generalization performance is the key of soft-sensing technique. Therefore, it is significant that conducting research aiming at improving the generalization performance of soft sensor. This paper analyses the principle and disadvantage of empirical risk minimization, indicates the reasons why artificial neural network based on empirical risk minimization has the poor generalization performance. From the perspective of relationship between complexity and generalization capacity on Machine learning, principle of structural risk minimization is expounded and the advantages comparing to empirical risk minimization are indicated. On the basis of contrast of mathematical expressions, topology and main training algorithm on RBF neural network and Support Vector Machine, in order to improve generalization performance of soft sensor, a modeling method of soft sensor based on Support Vectors-RBF neural network suitable for regression is proposed and expounded in theory: Taking classical cholesterin data set in MTLAB 6.5 as an example, simulation presents that Support Vector set-RBF neural network modeling method is superior to RBF neural network and evenly can compare with SVM in generalization performance.