Font Size: a A A

Multiple Modeling And Control For Nonlinear Systems Based On Local Learning Strategy

Posted on:2008-12-17Degree:DoctorType:Dissertation
Country:ChinaCandidate:T H PanFull Text:PDF
GTID:1118360215976818Subject:Control theory and control engineering
Abstract/Summary:PDF Full Text Request
With the rapid development of modern industry, large scale, quite complex and wide operating regime are presented in most production processes. Understanding the relationship among all kinds of variables of industry process through the first principle model has been becoming more and more difficult. In this case, it is an effective solution to understand the characteristics of process through the observations by multiple model approach, which based on divide-and-conquer principle. Under this multi-model frame, many mathematical tools such as machine learning, statistical theory and data mining techniques etc. are used. The approach decomposes a complex system into several local models, which are applied to resolve the uncertainty or simplify the complexity of the system, and has gained much more attention in recent years. The major result of this study is to combine the multiple model approach with data cluster and local modeling algorithm to solve the problems in model identification and controller design for strongly nonlinear system. The main works are as follows:Aimed to a class of nonlinear system that can be denoted as PieceWise Affine model (PWA), an off-line model identification is given in this paper through observations which can cover full operating regimes. Those input–output data are collected into several clusters by using an improved G-K clustering algorithm and singularity of the covariance matrix can be overcome in its iterative process of cluster. The number of sub-models can also be estimation based on the several information criteria. In each cluster, the parameter vector of the sub-model is obtained by the weighted least squares method. The two adjacent regions are achieved based on the nearest distance among the cluster centers, and the boundary hyper-plane between two adjacent regions in the regression space can be estimated using soft margin support vector. Simulation researches on a continuous nonlinear function and a real PWARX model are satisfying. The approximate ability of PWA model can be also validated. It is impossible to collect data which cover full operating regime in the real industrial process. This paper presents an online modeling strategy named lazy learning approach which based on statistical local learning theory. Analyzing similarity between the time series, k-Vector Nearest Neighbors (k-VNN) approach has been proposed. The interested observations in a local neighborhood of the query point are selected by k-VNN, and the predicted precision also can be improved. In order to improve calculated efficiency of lazy learning, a kind of recursively identified algorithm and a hierarchical searching process with two levels based on k-Means cluster algorithm have been offered in this paper. A kind of updating strategy for modeling database has been presented, too. Using this strategy, the adaptive ability of lazy learning can be increased and the memory of computer can be saved a lot. Simulation studies on estimation of a nonlinear function are satisfying. Lazy learning approach has been extended to soft sensing domain of industrial process.Based on local learning algorithm, two kinds of controller-designed strategy are presented for a nonlinear system that can be given input/output data. Firstly, the inverse and forward model of nonlinear system according to current operating regime can be identified by lazy learning algorithm from observations. The optimal value of controlled variable can be obtained by iterated optimal algorithm based on optimal performance index. The conclusions are evaluated by a SISO and non minimum-phase nonlinear system which is presented by literature. Secondly, a kind of controller designed approach that combines lazy learning algorithm and model predictive control is advanced. In order to overcome multivariable and many constraints that exist in most industrial processes, a constrained predictive control algorithm is also deduced. Simulation studies on a Fossil-Fuel Power Unit (FFPU) are satisfying. Finally, the performance comparison between lazy learning and other linearization strategies is compared, and some qualitative results are obtained.Combining lazy learning algorithm and process control, PID controller's parameters auto-tuning for nonlinear plants over a wide range of operating regimes are researched in this paper. Firstly, a novel two-level online auto-tuning algorithm is presented. The lower layer consists of a conventional PID controller and a plant process, while the upper level is composed of identification and tuning modules. Locally valid model of system according to current operating condition is identified by data-driven mechanism and lazy learning. Based on this model and performance criteria of General Minimum Variance (GMV), the optimal and suitable parameters of PID controller can be obtained. This scheme of the optimal tuning PID controller can suit demand of different operating regimes. Secondly, a heuristic rule to improve tuning strategy is also presented. Then simulations on Hammersterin model and CSTR process show better performance than other PID strategies.In order to verify the validity of the advanced control algorithm and simulate a practical industrial environment for engineers, an Advanced Algorithm Simulation (AAS) platform has been designed in this paper. Using the opened OPC (OLE for Process Control) protocol and general local area network, the platform has been configured a kind of Server/Client structure. The calculation for model identification and controller design is fulfilled by Matlab and the simulated environment of practical industry is realized by WinCC. This AAS platform can boost the development of advanced control algorithm and offer a friendly simulated condition.
Keywords/Search Tags:Divide-and-Conquer Technique, PieceWise Affine Model, Lazy Learning Approach, Inverse Control, Predictive Control, Simulation Platform
PDF Full Text Request
Related items