Search And Application On The Method Of Multi-modeling Soft Sensor Based On Hybrid Particle Swarm Optimization

Soft-sensor technology is a kind of valid method to solve complex measuring tasks and achieve the online estimation on variables measured difficultly. It can overcome the lack of online analytical instrumentation and estimate real-time the quality of products.Then it provides real-time, effective products quality data for the optimization of production and control operations.So it is focused extensively by the industry and has been successfully applied in many industries. As the single soft sensing model is difficult to describe the global characteristics of a complex system, and thus a kind of multiple model integration is focused highly to improve prediction accuracy and generalization.The existing multi-model integrate of the soft-sensor is lack of complementary property, individual sub-modeling accuracy and robustness of the linear integration method. I start from the model selection and optimization, and adopt different integration method to combining the characteristics of gasoline blending. The main work done is the following:1.Modeling methods of sub-models of PLS,RBF and LS-SVM are chosen according to the rapid modeling, model robustness,the sensitivity of the data sample, prediction accuracy and complementary property. Integration of multi-model structure is demonstrated.According to the property and convenience of industrial applications of integrated multi-model,two kinds of integrating methods based on the nonlinear BP network and linear are determined.2.The nuclear parameters and regularization parameters of RBF and LS-SVM are choosed by the PSO.Aiming at the premature convergence problem with the dimension increasing and diversity of swarm relative decreasing in PSO, the improved harmony search algorithm is introduced to particle swarm optimization.A dynamic harmony search hybrid particle swarm optimization(DHSPSO)algorithm based on a improved harmony search algorithm is proposed. The DHSPSO has better search capability and precision positioning features than the PSO.It lays the foundation to establish high-precision sub-models.3.The heterogeneous multi-models based on hybrid particle swarm optimization algorithm of PLS,RBF and LS-SVM are integrated by the linear method, and it has been applied to the prediction of research octane number (RON) in the system of gasoline blending. It has better prediction accuracy and robustness than single sub-model.4.The final output gained by linear weighted sum of sub-models is lack of robustness.The linear integrated model is sensitive to data sample. The contribution of single sub-model for the total output may contain nonlinear factors.The model must be revised by online. The heterogeneous dynamic static non-linear integrated model based on the DHSPSO is modeled to solve these problems.It is effective to improve the nonlinear effect of the data and online revision of the model and also improve the prediction accuracy and robustness of the model.The heterogeneous multi-model hybrid modeling method proposed in this dissertation can fully utilize the tested system and the measured variables prior knowledge and the existing experimental data. It provides an implementation technology for the online estimates of industry dates.