【作者】 徐小平；

【导师】 钱富才；

【作者基本信息】 西安理工大学 ， 控制理论与控制工程， 2010， 博士

【摘要】 作为系统和控制科学的核心问题之一,系统辨识在控制系统的设计与分析方面得到了广泛地应用,并且已经深入到了生物学、生理学、医学、社会学等众多的科学技术领域。从而,系统辨识成为目前相当活跃的学科之一,吸引了大量的科技人员对其理论进行研究,探讨在不同实际问题中应用的可能性。在系统辨识方法研究中,由于工程实际对象的复杂性,使得非线性系统辨识方法在工程对象的分析中表现出明显的优势。然而,因为非线性系统固有的复杂多样性,目前关于非线性的研究还远没有达到成熟的程度,蕴涵在复杂现象背后的“本质问题”还没有完全揭示出来。因此,传统的辨识方法始终未能很好地解决复杂的非线性对象的辨识,非线性系统的辨识一直是当今国际辨识界所关心的问题。辨识非线性系统的困难之一就是缺乏描述各种非线性系统特性的统一的数学模型。为此,这就需要人们分门别类地去探讨研究,解决所遇到的各种具体问题。因此,本文对模块化非线性系统的辨识算法进行了探讨和研究,期望在实际问题的应用中取得良好效果。论文的具体研究工作如下：1.利用粒子群优化算法研究了非线性系统的辨识问题。基本思想是：将非线性系统模块化后,给出基于粒子群优化算法的新型辨识方法。即就是说,首先,由典型数学模型相互组合来构成系统模型,将系统结构辨识问题转化为组合优化问题；然后,采用粒子群优化算法对系统的结构与参数进行辨识。利用仿真结果说明了所给的辨识方法是合理的和有效的。2.针对非线性环节为分段函数的模块化非线性系统,利用关键变量分离的原理,分离出线性环节的关键变量,将非线性环节表达式代入到其中去,使整个系统的输出表示为所有待估参数的回归方程形式。然后,将回归方程信息向量中未知真实输出、未知中间变量和不可测噪声项分别用辅助模型的输出、中间变量估计值和估计残差来代替,将非线性系统的辨识问题转化为参数空间上的函数优化问题。接着,采用粒子群优化算法获得该优化问题的解。最后,进行了仿真研究,其结果验证了所给辨识方法的有效性。3.研究了由一线性动态子系统和一无记忆的非线性增益串联构成的模块化非线性模型的盲辨识算法问题。当系统输入不采用高斯随机信号时,利用循环平稳输入信号的一阶统计特性和模型非线性部分的逆映射,将有输入信号的系统辨识过程转变为无输入信号的辨识过程,并且利用输出信号恢复了所有的中间变量。然后,通过支持向量机线性回归算法和高阶累积量方法来求取模型的参数。最后,仿真结果表明了所给的盲辨识方法是有效的和稳定的。4.针对一类多输入单输出的模块化非线性模型,其辨识问题可等价成以待估计参数为优化变量的非线性极小值优化问题。该方法的基本思想是将非线性系统的辨识问题转化为参数空间上的优化问题,然后,采用粒子群优化算法获得该优化问题的解。为了进一步增强粒子群优化算法的辨识性能,提出了利用一种混合粒子群优化算法。最后,通过仿真结果说明所给辨识方法的有效性和实用性。5.研究了对单输入单输出的采样模块化非线性Hammerstein-Wi ener模型的盲辨识算法问题。当系统输入不采用白噪声信号时,通过盲信号处理方法,仅仅利用输出信号恢复了所有中间变量。然后,通过支持向量机线性回归算法求取了模型的线性部分和非线性部分的参数。最后,在仿真实验中,与使用最小二乘法进行了比较,结果表明了所给的盲辨识方法是切实可行的。6.为了提高小波神经网络对非线性系统的辨识性能,利用一种改进粒子群优化算法对BP小波神经网络参数进行训练,求得最优值,达到对非线性系统辨识目的。在数值仿真中,与采用标准粒子群优化算法相比,结果显示了所给的方法在收敛性和稳定性等方面均得到了明显地改善。更多还原

【Abstract】 As one of the key issues of system and control science, system identification has been widely applied to the design and analysis of control system, and has been gone deep into many fields of science and technology, such as biology, physiology, medicine, sociology, and so on. Consequently, system identification becomes one of the current very active subjects, attracting a large number of scientific and technical personnel for their theoretical study to examine the practical problems in different application possibilities.In research on the method of system identification, because many engineering objects are very complex, the identification methods of nonlinear systems show remarkable advantage in the analysis of the engineering object. However, because of the inherent complexity and diversity of nonlinear systems, the current research on the nonlinear far not reached the degree of maturity; it has not been fully revealed that the "essence problem" of nonlinear was contained behind the complex phenomenon. Thereby, the complex nonlinear object recognitions are still not well solved via the traditional identification methods; the identification of nonlinear system was the main topics in the current international identification fields.For the identification of nonlinear system, one of the difficulties is short of unified mathematical model to description of various nonlinear system features. To this end, one should study on different specific problems. Consequently, in this paper, the identification algorithms of blocking nonlinear systems are discussed and studied, and expected good results in practical applications. The main research work is as follows:1. The identification problem of nonlinear system was studied based on a particle swarm optimization algorithm. The basic idea is that identification method was given by a particle swarm optimization algorithm, after the nonlinear system was blocked. That is to say, first of all, the idea of the method employed a system model composed with classical models so as to transform the system structure identification problem into a combinational problem. And then, a particle swarm optimization algorithm was adopted to implement the identification on the system structure and parameters. Simulation results showed the rationality and effectiveness of the presented method.2. The parameter identification method of blocking nonlinear model with two-segment piecewise nonlinearities was presented. Its basic idea is that:Firstly, expressing the output of the blocking nonlinear models as a regressive equation in all parameters based on the key term separation principle and separating key term from linear block and nonlinear block. Secondly, the unknown true outputs of the information vector are replaced with the outputs of an auxiliary model; the unknown internal variables and the unmeasured noise terms are replaced with the estimated internal variables and the estimated residuals, respectively. Accordingly, the problem of nonlinear system identification was cast as function optimization over parameter space, and then, a particle swarm optimization algorithm was adopted to solve the optimization problem. Finally, simulation results showed the effectiveness of the proposed method.3. A blind identification algorithm to a blocking nonlinear model, where the nonlinear block is a linear subsystem is followed by a memoryless nonlinear gain, was investigated. When the input signal of system didn’t adopt a Gaussian random signal, the identification process with the input signal was changed into the one without input signal using the first-order moment of the cyclostationary input signal and the inverse nonlinear mapping of the nonlinear model. And all internal variables were recovered based on the output signal. Then, parameters of the model were obtained through support vector machine linear regression algorithm and higher order cumulants method. Finally, the efficiency and stability of the presented algorithm were demonstrated by simulation examples.4. For a class of multi-input single-output blocking nonlinear model, the model identification problem is equivalent to the nonlinear minimization problem with the estimated parameters as the optimized variables. The basic idea of the method is that the problem of the nonlinear system identification was changed into an optimization problem in parameter space and a particle swarm optimization algorithm was then adopted to solve the optimization problem. In order to enhance the identification performance of the particle swarm optimization algorithm, a hybrid particle swarm optimization algorithm was also adopted. Finally, simulation results showed the effectiveness and practicality of the proposed identification method.5. A blind identification algorithm for the single-input single-output sampled blocking nonlinear Hammerstein-Wiener model was proposed. When the input signal of the system didn’t adopt a white noise signal, by a blind signal disposal approach, all internal variables were recovered solely based on the output signal. Then, parameters of linear and nonlinear parts of the model were obtained through support vector machine linear regression algorithm. Finally, in simulation experiments, compared with using least square method, the effectiveness of the presented algorithm was illustrated.6. In order to improve the identification performance of the wavelet neural network for the nonlinear system, the parameters of a BP wavelet neural network were trained via an improved particle swarm optimization algorithm to obtain optimal values to achieve the purpose of identification for the nonlinear system. In numerical simulation, compared with using standard particle swarm optimization algorithm, the results showed that the presented algorithm was obviously improved in the convergence, stability, and so on.更多还原