【作者】 吴彭江；

【导师】 孙明轩；

【作者基本信息】 浙江工业大学 ， 控制理论与控制工程， 2008， 硕士

【摘要】 系统辨识是高性能自动化技术(如适应控制与学习控制)中的重要内容。随着人们生活水平和工业生产水平的提高,人们对工业产品质量的要求越来越高,生产工艺对自动化技术的依赖程度也随之提高,对系统辨识理论与方法提出了更高的要求。许多实际系统动力学特性中的参数随时间变化,这要求辨识方法具有跟踪能力,能够估计时变参数。在实际中也存在着受突发异常干扰的辨识环境,因此,有必要研究鲁棒辨识算法。针对这样两个方面,本文的研究工作主要包括以下几部分:1.讨论时变遗忘因子和加权配合的递推辨识算法的算法性质和确定性收敛性。2.针对一类在有限时间区间上重复运行的离散时变系统,讨论用于系统时变参数估计的迭代学习辨识算法。文中进一步讨论迭代学习投影算法、迭代学习最小二乘算法的理论性质,并给出数值仿真结果;推导得到两种迭代学习辨识算法:迭代学习贝叶斯法和迭代学习随机牛顿法。不同于常规跟踪算法的有界收敛性,迭代学习辨识算法可在有限时间区间上实现对时变参数的完全估计,使得沿整个区间参数估计误差为零。3.将近似最小l_l模递推算法应用于回波消除,分析了这种算法的收敛性能。分析表明,这种算法一旦收敛,对于有界干扰具有强鲁棒性。这一性能有利于提高回波消除质量。双讲时在不增加系统硬件(对讲检测器)情形下,算法仍具有良好的收敛性能,避免了在双讲情况下设置双方对讲监测器。4.依据非线性PCA准则,以迭代加权最小二乘算法极小化目标函数,近似一类l_p模指标下的盲分离递推算法,文中给出了算法性质。和RLS算法相比,所提算法未显著增加计算复杂度,对幅度较大的非正态分布或统计特性未知的噪声有较强的抑止能力。5.设计实现冰球式蓄冷空调系统。用时变系统来近似实际系统,采用遗忘和加权配合的辨识算法来预测能量。在系统实现上,利用组态王实现数据采集、友好人机界面设计,利用MATLAB完成复杂的预测算法,充分发挥组态软件在工业控制上组态方便、开发周期短和MATLAB强大数值计算功能的优势,两者之间通过DDE技术实现数据实时通信。更多还原

【Abstract】 System identification plays an important role in high performance automation technologies such as adaptive control and learning control. More high quality of products is required, in nowadays plants, to meet the needs of people’s production and living demands. It should be challenged for system identification when high control performance is pursued. The parameters in dynamics of many real systems are time-dependent, which requires that identification techniques are of the ability of tracking, and can estimate time-varying parameters. We also need robust identification algorithms owing to the presence of abrupt disturbances. Take into account the issues above mentioned, this thesis focuses on the following aspects:1. The properties and deterministic convergence of recursive identification algorithm are considered, in which both a forgetting factor and weights have been incorporated.2. Iterative learning identification algorithms are presented for estimating time-varying parameters of a class of discrete time-varying systems over a finite time interval. The theoretic properties of iterative learning projection and least squares algorithms are further discussed, and corresponding numerical simulations are presented. Two prototype algorithms of iterative learning identification, iterative learning Bayes and stochastic Newton algorithms, are proposed with detail. Different from the bounded convergence performance obtained by conventional tracking algorithms, perfect estimation for the time-varying unknowns is achieved through iterative learning, and the parameter estimation error is guaranteed to converge to zero over the entire time interval.3. The approximate least l₁-norm recursive algorithm is presented for adaptive acoustic echo cancellation. The convergence performance of the algorithm is analyzed, by which the algorithm exhibits robustness with respect to bounded disturbances at steady-state. It is in turn ? useful for performance improvement of the acoustic echo cancellation. During duplex talk, the proposed algorithm works well without requiring Double-Talk Detector and without requiring more hardware in the system. 4. Based on the nonlinear PCA criterion, an iterative weighted least squares algorithm is presented for blind source separation. The iterative algorithm is for the purpose to approximate the l_p-norm index, and the theoretical properties of the algorithm are derived. In comparison with the conventional RLS algorithm, the proposed algorithm has comparable computational complexity, and is robust with respect to non-Gaussian disturbances or those with unknown statistical distribution.5. An ice storage air condition system is designed and implemented. The real system is modeled as a time-varying system, and the energy is predicted by the identification algorithm with both forgetting and weighted factors. For the implementation, KingView is adopted for data acquisition, with friendly human machine interface design, and MATLAB is used to carry out the on-line prediction computing. In the ice storage system, DDE technique is applied for the real time data exchange between the configuration software and MATLAB.更多还原