Research On Soft Sensor Method In Chemical Process Based On Bayesian Network

In the chemical process,there are often a large number of important variables that cannot be measured online in real time by hardware sensors.Therefore,soft sensor technology was invented and received widespread attention.Data-driven soft sensor technology uses easily available auxiliary variables to estimate quality variables that are difficult to measure.Among many data-driven soft senser method,Bayesian network,as a combination of probability model and graph model,can effectively deal with uncertain relationships and causal reasoning.Because the data in the non-stationary process has complex characteristics such as nonlinearity and dynamics,it increases the difficulty of soft sensor technology to predict quality variables.Therefore,this thesis focuses on the complex chemical process and studies the soft-sensing modeling method based on Bayesian networks.The main research work of this thesis is as follows:(1)The modeling process of soft sensor method is studied.The concepts,inference and learning methods of Bayesian static and dynamic networks are summarized,which lays the foundation for the methods proposed in this thesis;(2)Aiming at the influence of uncertain factors such as non-stationary data nonlinearity and random noise interference in the chemical process,a Bayesian network method for soft-sensing modeling based on hidden space is designed.At first,in order to extract latent information in complex process variables,a supervised Bayesian network is proposed to extract quality-related latent variables that are applied to local weights.Next,a double-layer similarity locally weighted method based on latent variables is proposed,which allows the weight to take into account the relationship between samples and the global information.Finally,the effectiveness of the method is proved by two case studies.(3)The actual chemical process often changes dynamically,so the soft-sensing modeling method based on dynamic Bayesian network is designed.In order to deal with nonlinearity appropriately,two similarities for dynamic Bayesian networks are calculated.By weighting the two parts of the likelihood function.This improvement makes dynamic Bayesian network more adaptable to dynamic changes.Two case studies have proved the predictive performance and dynamic tracking ability of the method.