Modeling Technology Of Soft Sensor Based On Data Driven And Its Industrial Application

In industrial process, there are variables which are very important for guaranteeing the quality of products or keeping the process running properly, but they are very difficult to be measured directly for reasons of technology or economy. To solve this problem, a new conception named soft-sensor is created and has been one of the most important research directions in the area of process control. Its basic principle is to select a set of secondary variables that are easy to be detected and possess close relationship with the primary variables according to certain "optimal" criteria. The selected secondary variables are then used to obtain the on-line estimation of the primary variables by constructing some mathematic relationship between these variables. The main research works conducted in this dissertation are described as follows:1. The development, actuality and traits of the soft-sensor technique are summarized, and the concept, basic model and some common modeling technologies of soft-sensor are introduced. Thereinto, some typical soft-sensor modeling technologies based on data driven are stressed.2. In order to solve the problem of deciding the number of nodes of the hidden layer of neural network, a soft-sensor modeling algorithm which can prune away the redundant nodes of the RBF neural network at one time by analyzing the output information of its hidden nodes with PLS-Pruning algorithm is presented and the corresponding optimized mathematic analytical model can be obtained. And simulation research proves the validity of this algorithm.3. In order to improve the generalization of the soft-sensor model, the algorithm of least squares support vector machine (LS-SVM), which is based on statistical learning theory(STL) and structural risk minimization(SRM) principle, is used to build the soft-sensor model and simulation in laboratory demonstrates that the modeling algorithm needless samples and is accurate in forecasting precision.4. The RBFNN soft-sensor modeling algorithm based on fast-pls-pruningand the LS-SVM algorithm are applied to build soft-senor models severally for the average diameter of an industrial Purified Terephthalic Acid (PTA) crystal in one chemical industry and for the nutrient concentration of the multiplex fertilizer device in one vitriol factory. Simulation results with factual industrial data demonstrate that the two methods are both fast in computation and accurate in prediction, so they are suitable to practical application.Finally, the dissertation is concluded with a summary and discussions of the problems for further research and exploration on soft-sensor modeling technique.