Research On Soft Sensor Modelling Method Using ESN In Chemical Processes

Dynamic soft sensor modelling is a hot research in the field of chemical processes, it can effectively resolve some problems of online analytical instruments, which include lag measure, expensive and complex maintenance for complex strongly nonlinear systems, it can attain requirement of real time for actual production process. At present, main dynamic soft sensor modelling methodes include neural networks, support vector machines(SVM) etc. As a new dynamic recurrent neural network(RNN), ESN has been widely concerned by Scientists. The structure of hidden layer of ESN is a State Reservoir(SR) which has echo state property(ESP), and learning algorithms of ESN have the advantages of simplicity, effectiveness and fast convergence. Thus ESN has strong dynamic approximation ability.Based on the theory of ESN, in this thesis, a class of dynamic soft sensor modelling method using leaky integrator echo state network(LiESN) are proposed. The corresponding learning algorithms are also given. Dynamic soft sensor modelling methods are then applied to chemical production proce ss for refinery, experimental results show that this method has a good modeling effection and great potential in industrial application. The main contents are as follows:(1) A dynamic soft sensor modelling method with different time series modeles are studied. A 4-plot model residuals analysis of statistical analysis method to assess the effection of soft sensor modelling method is given.(2) Based on the theory of ESN, LiESN and their learning algorithms are studied. LiESN ridge regression offline learning algorithm and recursive least squares(RLS) online learning algorithms are proposed. By adding a regularization coefficient, ridge regression algorithm could control large sizes of output weight matrix and improve the properties of ESN solution. O n- line learning algorithm could allow on- line processing of large data sets and attain requirement of real time for process modeling. In addition, the global parameters optimization based on gradient descent with Least Mean Square(LMS) algorithm is further studied.(3) Dynamic soft sensor modelling methods combining LiESN with nonlinear moving average(NMA) or nonlinear autoregressive(N ARX) dynamic time series model are studied. By using dynamic soft sensor modelling method, the butane(C4) concentration in the bottom flow of a debutanizer column is estimated and the tail gas composition in the sulfur recovery unit(SRU) are computed. Compared with existing dynamic soft sensor modelling methods such as ESN, SVM etc, under the same condition, expe rimental results confirm the employed LiESN method can achieve better performance, the accuracy of the model and stability can meet the practical need. Therefore, research results of this thesis have great significance for improving the quality of products effectively and achieving the minimization of the emission pollutants in refinery.