Research On System Identification By Genetic Programming

Modern control objects in automation fields are becoming more and more complex. Accurate system identification and modeling is important for effective implementation/optimization of control strategy, process monitoring, fault diagnosis, simulation, etc. In this thesis, genetic programming (GP) is combined with an advanced sustainable evolution technique named HFC (Herarchical Fair Competition), to effectively, accurately and simultaneously identify structures and parameters of complex systems.The achievements of the thesis can be concluded as follows:1. Existing identification methods and their limitations are analyzed, and then a basic single population GP algorithm is constructed which can be applied to both static and dynamic system identification problems, and realize simultaneous identification of both structures and parameters, thus forming a new unified system identification method based on evolutionary computation, which is superior to other exiting methods.2. On the basis of analyzing premature convergence phenomenon in evolutionary computation, limitations of single population GP are explained-why it can not solve complex problems efficiently. Existing anti-premature convergence methods are explored in detail, and their mechanisms and limitations are examined. On the basis of the basic single population GP identification algorithm, two parallel multi-population GP identification algorithms are constructed especially for stochastic nonlinear system identification:the ISLAND model multi-population GP algorithm and the HFC GP algorithm.3. Object systems of different types are identified:the single population GP identification algorithm is used on two static object systems. All three GP identification algorithms (single population GP, ISLAND GP and HFC GP) are used on five stochastic linear or nonlinear dynamic systems, results are analyzed and compared.4. Performances of all 3 GP algorithms are compared. HFC GP shows the most competitive performance in sustaining evolution, at the same time it also runs with efficiency in the respect of time cost, it's an excellent identification algorithm especially for complex stochastic nonlinear systems.