基于群体智能和聚类的模块化非线性系统智能辨识

Nonlinear systems widely exist in industrial processes, but the related theories of system identification fail to meet the needs of accurate modeling. Here, there exist two problems: 1. unreasonable descriptions of nonlinear system structures; 2. inaccurate parameter estimations under known and complex structures. In response to the first problem, this project intends to adopt the block-oriented nonlinear models that separate the system statics and dynamics, in order to describe the system characteristics sufficiently. For the nonlinear parts of models, different methods will be involved according to various sources of nonlinear characteristics. Specifically, using artificial neural networks is to approximate the strong nonlinear process mechanisms, and using unified nonlinear characteristics is to describe the nonlinearities of actuators or sensors. Regarding to the second problem, this project intends to apply the novel intelligent identification techniques to achieve robust estimation in the presence of stochastic noises. The whole scheme is about “swarm intelligent optimization – cluster analysis – local optimization”. Firstly, by improving the key techniques of swarm intelligent optimization, the adaptability of identification algorithm can be improved in noisy environment. Secondly, in order to avoid the global search being trapped in local optima, cluster analysis will be implemented for elite swarm obtained by the first step. In this process, multiple underlying promising search areas can be determined by cluster analysis results. Finally, the local optimization algorithm will be applied for optimizing the individuals under these promising areas, then the global optimal parameter estimation will be harvested according to individual fitness values.

在工业过程中，非线性系统广泛存在，而相关的辨识理论不能满足精确建模的要求。这里存在两个问题：1. 非线性系统的结构描述不合理；2. 已知复杂结构下的参数估计不准确。对于第一个问题，本项目拟采用静态与动态分离的模块化非线性模型以充分表征系统特性。对于该模型下的非线性模块，项目将根据非线性特性的不同来源采用不同的描述：利用人工神经网络来逼近具有强非线性关联的过程机理，并使用具有泛化能力的统一非线性特性来描述执行器或传感器部位的非线性。对于第二个问题，本项目拟运用“群智能优化——聚类分析——局部寻优”这一新型智能辨识技术来实现随机干扰下的鲁棒估计。首先，通过改进群智能优化的关键技术环节来增强辨识算法在噪声环境中的适应性。其次，为了避免全局搜索陷入局部最优，对第一步所得精英种群进行聚类分析以确定多个潜在较优区域。最后，对这些区域中的个体进行局部寻优，根据其适应度大小确定出全局最优的参数估计。

项目负责人围绕“基于群体智能和聚类的模块化非线性系统智能辨识”开展了深入细致的研究。首先，负责人对模块化非线性系统的适用范围进行了深入分析，发现Wiener结构具有较为广泛的适用性。它具有线性动态与非线性特性相串联的结构，该结构对动态过程内部不同时间尺度的动态特性、双线性过程、甚至对于带有复杂非线性的动态过程都具有较好的描述能力。其次，负责人针对常见的模块化非线性系统，如Hamerstein系统、Wiener系统、Wiener-Hammerstein系统、Hammerstein-Wiener系统，开展了一系列辨识方法的研究。对重点关注的系统切换特性、慢时变特性等进行了辨识方法上的详细设计，且对包含大量参数的系统模型采用了改进群智能的搜索算法进行有效辨识。其中群智能搜索算法的主要改进之处有两点：其一是针对性能指标的改进以应对脉冲噪声的影响；其二是针对精英种群进行K-means聚类分析以缩小搜索范围。最后，项目负责人进行了多项数字仿真实验，针对不同的非线性动态过程进行建模和辨识，验证了模块化非线性系统模型的有效性以及智能辨识的准确性。