中文摘要：

针对一类非线性、快时变或系统参数跳变的离散时间系统，需要解决的问题是：设计多模型自适应控制器使得系统的输出能够更快、更准确的跟踪参考输入。为此，本文提出了一种新的基于聚类优化多模型集的方法。该方法首先根据系统的输入输出数据构造数据样本集，采用模糊核聚类自适应算法来对数据样本集进行自适应聚类，再对各类数据采用最小二乘法来建立局部模型。在多模型自适应控制过程中，当系统产生一个新的输入输出数据时，依据该数据与各聚类中心的距离来判断数据归属于哪个聚类，并利用递推最小二乘法来对相应的局部模型进行更新。如果不属于任何一个聚类的新数据累积到一定的数量时，则要重新对已有的数据进行聚类建模。另一方面，随着数据量的增加，多模型集中的模型数量也有可能逐渐增加，为了防止模型数量的无限增长，设定了一个模型数量的阈值，当模型数达到这一个闽值则根据新数据与各聚类中心的距离来删除最不符合当前系统状态的局部模型。最后，给出的MATLAB仿真结果表明：与传统方法相比，本文提出的方法能够使系统的真实输出更快更好的跟踪参考输入。可以得出，利用本文提出的一种基于聚类优化的多模型自适应方法可以更有效的解决一类具有非线性、快时变、参数有跳变的系统的控制问题，并经过了大量的仿真实验证明了该方法的有效性。

英文摘要：

For a class of nonlinear discrete time system with fast time-varying or jumping parameters, a Multiple Models Adaptive Controller （MMAC） based on cluster-optimization is proposed in this paper. The main control problem is to design the multiple models adaptive controller which can guarantee that the real output of system can track the reference output more quickly and accurately. For this purpose, a MMAC utilizing cluster-optimization is designed. Firstly, a sample data set would be established by random input and the corresponding output The fuzzy kernel clustering adaptive algorithm has been presented to classify the data in the sample set into several clusters. Then the least square method is adopted to construct the local models of the multiple models set respectively corresponding to each clusters. When a new data point is got, it could fall into a cluster which has the shorter distance between it and the cluster centre than the mean of all points in this cluster. The related local model can be updated by the recursive least square method. Otherwise, the new data should be called unclustered data. If the unclustered data increase to a predefined threshold, all the clusters and the models are needed to be reconstructed. On the other hand, along with the number of data added up, the number of the local models may also increase. When the number reached a given value which is designed to prevent excessive number of models, a local model would be deleted depended on the similarity of models and the real system. Moreover, the result of matlab simulation can prove that the approach proposed is more effective than the traditional one. It can make the real output track the reference output more quickly and accurately. Finally, a conclusion can be drawn that the control problem of the nonlinear system can be solved by the proposed method effectively.