空间分布系统空间混合神经模糊控制器设计与应用

Spatially distributed systems are a class of complex systems with widespread application backgrounds and in possession of profound worthiness of theory research, and have been gotten more and more attentions in recent decades. Since many actual physical systems are usually installed with multiple distributed control sources, and have the feature of two-dimensional space distribution, nonlinearity, and uncertainty, they show extremely complicated system characteristics. However, classical distributed parameter control methods are difficultly applied to the practical physical systems due to their complex theory and their model-based design style. Therefore, this project will investigate a novel spatial hybrid neuro-fuzzy controller, which combines space information, a fuzzy system, and a neural network into a unified framework and integrates the advantages of the expression and processing of spatial information, learning ability, linguistic interpretation ability, and model-free design. The research project consists of five parts. Firstly, the mechanism of expressing and processing two-dimensional space information will be investigated, and a spatial hybrid neuro-fuzzy model will be designed. Secondly, a two-dimensional spatial kernel-based multi-output support vector machine will be proposed to solve the initialization problem of the controller. Thirdly, a self learning method including a phased self-organization of the structure and an online self-learning of the parameters will be concerned. Fourthly, based on the deep analysis of the controller structure, issues about system stability will be discussed. Finally, a snap curing oven as a key equipment in the semiconductor packaging will be taken as an example to validate the effectiveness of the above proposed theory.

空间分布系统作为一类重要复杂系统有着广泛的应用背景，已受到国内外学者的极大关注。由于多数实际系统配置有多个控制源，具有非线性、不确定、二维空间分布等特点，系统特性极为复杂。经典分布参数控制方法由于依赖系统模型、理论复杂难以在工程中得到应用。本项目研究一种新型的空间混合神经模糊控制器，即将空间信息、模糊系统、神经网络有机结合起来，形成集空间信息表征与处理、语言解释、学习、不基于模型等优点于一身的新型智能控制器。本项目从空间分布特性出发，研究二维空间信息表征与处理的方法，在此基础上设计融合空间信息的神经模糊控制器模型；提出基于二维空间核函数的多输出支持向量机学习算法，解决网络初始化问题；研究基于阶段性的结构自组织与在线的参数自学习的控制器自学习方法；在深入分析控制器结构的基础上，探讨系统的稳定性；以半导体封装快速固化炉为例，开展应用研究。

本课题针对实际广泛存在的二维空间分布动态系统，充分挖掘控制系统历史数据和实时数据所蕴含的丰富信息，展开了空间混合神经模糊控制设计与分析等相关问题的研究，主要理论进展和创新成果有：1﹑针对二维空间分布的特点，提出了具有二维空间分布的三域模糊集合与能够处理二维空间信息的三域模糊逻辑推理机，在此基础上，研究并提出了具有七层网络结构的空间混合神经模糊控制器，分析了各层结构的工作机理。2﹑基于已有的隐含有效控制规律的时空耦合输入输出数据，提出了基于支持向量学习﹑基于聚类学习的空间混合神经模糊控制器的结构与参数初始化方法。3﹑依据在线的时空耦合输入输出数据，提出了基于可变参数的在线SVR算法的多输出空间混合神经模糊控制器的结构自组织方法；提出了一种基于模糊变参数的SPSA算法的多输出空间混合神经模糊控制器的参数自组织方法。4﹑从数学的角度，推导出多输出空间混合神经模糊控制器的解析结构；并且以此为基础，分析了二维空间分布多输出空间混合神经模糊控制的稳定性。此外，建立了用于测试和验证上述研究的烘烤系统实验平台与半导体封装快速固化炉实验平台，并在此基础上开发了二维空间3-D模糊控制系统软件。这些平台与软件为空间分布系统智能控制﹑学习与建模奠定了重要基础。