Research On Soft Sensor Modelling Of Chemical Process Using Dynamic Neural Network

Soft sensor technology is a vital component of advanced process control technology, and it's also an important means of process detection and process control in chemical process area. It can effectively overcome the defects of online analytical instruments, which is expensive, big lagging and complex maintenance, and improve real-time monitoring of process variables in chemical companies, further improve the quality of products, to meet the needs of industrial production control system, and effectively enhance the competitiveness of the enterprises. Currently, the neural network has become an important tool for soft sensor modelling. Recurrent neural network(RNN) as a kind of dynamic neural network, has been successfully applied in soft sensor modelling based on data-driven.This dissertation uses kalman filtering as a method of state space model, which uses connection weights of neural network in each layer. This would make a state of filter to update and replace the traditional algorithm of RNN to train the network, it also can effectively improve the prediction precision of the network, and has been successfully applied in time series prediction. On the basis of the kalman filter, this dissertation gives the extended kalman filtering(EKF), cubature kalman filtering(CKF), and square-root cubature kalman filter(SCKF) algorithm for RNN training. Satisfactory results have been achieved by using this soft sensor modeling method to chemical process instances. The dissertation's main research contents are as follows:(1) The topology of simple recurrent network(SRN) and fully connected recurrent neural network(FCRNN) are studied, and the SRN as the key research object gives the basic training algorithm.(2) As the shortage of traditional training algorithm, based on time suboptimal state estimation technology and kalman filtering, a FCRNN method based on EKF training algorithm is proposed; in order to further improve the accuracy of network and the stability of algorithms, this dissertation researches on CKF algorithm and SCKF algorithm, and a SRN method based on SCKF algorithm is proposed.(3) The SRN based on SCKF algorithm which combined with Nonlinear Moving Average(NMA) or Nonlinear Autoregressive(NARX) dynamic time series model are used to construct dynamic soft sensor modelling. By using dynamic soft sensor modelling method, C4 concentration in the bottom flow of a debutanizer column is estimated and the SO2 and H2 S in the tail gas composition in the sulfur recovery unit(SRU) are computed. In order to validate the effectiveness of the method, the method is compared with the existing soft sensor modelling methods such as the method of SRN based on traditional training algorithm, the method of FCRNN based on traditional training algorithm, the method of multilayer perceptron(MLP) based on EKF algorithm, the method of FCRNN based on EKF algorithm, the method of SRN based on EKF algorithm, the method of MLP based on SCKF algorithm and the method of FCRNN based on SCKF algorithm under the same condition. The results shows that the thesis' s method can obtain a high modeling accuracy, which is a kind of effective soft sensor modelling method.