Research On Application Of Identification And Control Of Chemical Processes Based On ELM

Chemical processes are objects with high complexity and uncertain nonlinearity, so that it is difficult to obtain the precise mechanism models of chemical processes. Therefore, the computational intelligence methods can achieve dynamical identification models of nonlinear objects and intelligent adaptive controllers of online control, which play important roles in the identification and the control of chemical processes, such as neural networks(NNs) and fuzzy neural networks(FNNs) and so on. Extreme learning machine(ELM) as a single-hidden layer feedforward neural networks(SLFNs), which only needs to analytically determine the output weights of SLFNs, tends to provide good generalized performance at extremely fast learning speed and avoids falling into local minimum risk using gradient descent method of neural network. In view of these, the ELM network should be a good candidate of computational intelligence methods. The model identification and neural control method of chemical processes based on ELM are studied in this dissertation. The main contributions of this dissertation are as follows:(1) The basic structure characteristic of ELM is analyzed and its offline and online training algorithm are further studied. The dissertation has combined the ELM with the nonlinear auto regressive with exogenous inputs(NARX) and aimed at two typical chemical processes whose models are unknown. The two typical chemical processes are continuous stirred tank reactor(CSTR) process with strong nonlinearity and pH neutralization process with serious nonlinearity and large time delay characteristic. The ELM network is applied to off-line identification and online-identification. First of all, the off-line identification is proposed based on NARX-ELM network. In order to improve the generalization performance of the off-line identification method, NARX-regularization ELM is proposed. Finally, the on-line identification method based on NARX-OSELM is proposed. Under the same conditions, the off-line and on-line identification methods of ELM are compared and analyzed.(2) On the basis of regular ELM, the basic structure of KELM and its training algorithm are studied when the feature map function of ELM is unknown. Aiming at CSTR process and pH process, the NARX-KELM network identification method based on two different types of kernel functions are proposed, and the effection of two different kernel functions on the system performance are discussed. Under the same conditions, the results show that NARX-K ELM can effectively reflect the system's dynamical performance with smaller identification error in comparison to NARX-ELM, NARX-regularization ELM and NARX-OSELM and some existing references.(3) Aiming at the CSTR nonlinear chemical process, which is transformed into a class ofnon-affine nonlinear pure feedback dynamical systems with uncertainty, a backstepping adaptive neural control method based on ELM neural network is proposed. In the proposed method of every step of the backstepping design, ELM network is used to approximate subsystem of the unknown nonlinearities online. The weight parameters of the adaptive regulating law is designed based on Lyapunov stability analysis, so that all signals of closed-loop nonlinear system are guaranteed semiglobal uniformly ultimately bounded and system output convergences in a small neighborhood of the desired trajectory. Simulation results show the effectiveness of the control method.