Application Of Bayesian Kernel Learnhig Based On Soft Sensor To Thickening Process

Our country is rich in non-ferrous metal resources, but with the rapid social and economy development and the steady progress of industrialization, accelerating the consumption of non-ferrous metal resources and the non-ferrous metal resources increasingly depletes, which severely restricts the economic and social sustainable development. Hydrometallurgy can deal with miscellaneous and low grade mineral, benefits for environment and uses the mineral resources efficiently, which is easy to achieve continuous and automatic production process. Thickening process is an important unit operation in hydrometallurgy. The underflow concentration of thickener is the key quality index. To achieve the control and optimization of thickening process, it needs to establish an accurate model of the thickening process.Aiming at the difficulty of online measuring key variables and based on deeply analyzing the characteristics of thickening process, this thesis uses hybrid modeling method which is composed of mechanism and data-driven modeling methods to predict the key variables of thickening process online. Then a model correction strategy based on model evaluation is developed to realize the online application of thickening process model. The main researches are summarized as follows:(1) Through analyzing the basic principle of sludge sedimentation and based on solid flux theory and mass conservation theory, the sludge concentration distribution model is established. Through the simulation of the proposed model, the characteristics of the thickening process are revealed to find the main factors of the thickening process, to determine the secondary variables of the soft sensor model, and to establish the foundation of soft sensor model.(2) Aiming at the difficulties of applying the mechanism model to the industrial field directly, soft sensor model of thickening process is established adopting parallel hybrid model. The model is composed of the simplified mechanism model and data compensated model, therefore the advantages of different modeling methods can be exerted. Considering the nonlinear characteristics of thickening process, least squares support vector machine (LSSVM) approach is used to realize the nonlinear fitting of prediction error. And Bayesian method is used to tuning the parameters of least squares support vector machine. Simulation results verify that hybrid method is more effective in comparison with the mechanism model.(3) Aiming at the difficulty of tracking slowly time-varying characteristics, a model correction strategy based on model evaluation is proposed. Gaussian Mixture Model (GMM) is established to describe model prediction error distribution characteristics. Then the statistic is constructed from the GMM to evaluate the hybrid model performance. Based on the assessment results, a hybrid correction strategy is proposed, which combines output offset compensation and online model parameters updating. When the prediction error is within the normal range, the hybrid model is corrected by model output offset compensation; When the prediction error is out of the normal range and the samples reaches a certain amount, the hybrid model is corrected by online model parameters updating and model output offset compensation. Through combining the model output offset compensation and online model parameters updating, the model prediction accuracy is further enhanced, and achieves the online measurement of key variables for thickening process.(4) Based on the research of thickening process, the simulation platform of thickening process is designed. A detailed description of hardware and software structure is given. The measurement of key variables for thickening process is achieved, and the operating guide for thickening process is provided.