Development of nonlinea system identification methods for micro PCR (polymerase chain reaction) reactors

Abstract

Various linear system identification methods have been developed. The linear identification method has limitations in describing industrial processes since most of them are actually nonlinear. Recently, many researchers in the field of system identification have been focusing on development of nonlinear system identification methods. Among various nonlinear models, block-oriented nonlinear models and artificial neural networks are generally used. Hammerstein and Hammerstein-Wiener models are useful for representing the chemical processes since most of them include nonlinear actuators. When it is difficult to determine or choose the proper model structure of nonlinear processes, neural networks are preferred due to their excellent ability to fit the nonlinear mapping from the input to the output data. Although many researches have been done for block-oriented nonlinear models as well as neural networks, they have serious problems as follows: (1) Because the previous identification methods cannot separate the two identification problems of the linear system and the nonlinear static function in the Hammerstein models, and an iterative optimization is needed to identify the nonlinear static function, this iterative optimization may cause poor initialization and numerical instability: (2) Very little study has been done on Hammerstein-Wiener models: (3) Most previous artificial neural networks have been developed for discrete-time system identification. Their model performances may be seriously poor when the sampling time is small compared to the process time constant. To overcome the above problems, three separate studies have been carried out. (1) A continuous-time recurrent neural network has been developed for the continuous-time processes, and the training rule for the proposed neural network has been derived. To demonstrate the model performance, the proposed neural network has been applied to a micro PCR reactor. The experiment result shows the model performa...