太原理工大学 硕士学位论文 粒计算神经网络算法及其在旋转机械故障诊断中的应用 姓名：李凤 申请学位级别：硕士 专业：控制理论与控制工程 指导教师：谢克明 20090501 太原理工大学硕士研究生学位论文 I 粒计算神经网络算法及其在旋转机械故障诊断中的应用 摘 要 本论文在进行粒计算和人工神经网络的理论研究的基础上，进一步将 两者结合起来，提出了粒计算神经网络算法，并将其应用到旋转机械 故障诊断之中。 粒计算理论（Granular computing theory, GrC）是美国学者 T.Y.Lin 教授 提出的，他曾指出“粒计算理论是数据约简的一个好工具” 。粒计算从不同 粒层次上研究问题，主要用于处理不确定的、模糊的、不完整的和海量的 信息，是信息处理的一种新的概念和计算范式，覆盖了所有有关粒的理论、 方法、技术和工具的研究。作为一种新的智能信息处理技术，粒计算属于 “软计算”的一种，在近几年受到国内外学者的广泛关注。人工神经网络 (Artificial Neural Networks，ANN)因其具有自学习、自组织、容错性好和并 行处理信息等能力，广泛地应用于各种领域。 本文在此基础上提出的粒计算神经网络松耦合算法将粒计算与人 工神经网络相结合，各取其优点，优势互补，其核心思想是将粒计算作为 神经网络的前端处理器，即利用粒计算理论强大的约简能力对原始信息进 行化简，从而得到与原始信息等价的最小属性集，最后构建基于最小属性 集的神经网络，从而进行网络训练。实验证明，基于粒计算的约简算法简 单明了，计算量小，简化了网络结构，提高了训练效率。另外，本文还提 出了二进制粒神经网络（Binary Granular Computing Neural Networks, 太原理工大学硕士研究生学位论文 II BGrCNN）模型，该模型将二进制粒与神经网络紧密地融合在一块，构建了 一种新的神经网络。该模型的输入输出及其运算过程均为二进制数，能大 大提高网络的训练时间。 本文的主要创新成果有： 1) 首次将基于 Rough 集的粒计算与人工神经网络相结合，各取其优点， 建立了粒计算人工神经网络的模型。粒计算是对数据进行约简的一 种方法，人工神经网络是对数据进行分类的一种手段，二者的结合可 以彼此加强处理数据的能力。 2) 提出粒计算人工神经网络故障诊断算法，以改进人工神经网络在故 障诊断应用中的内在缺点。 3) 将粒计算人工神经网络松耦合故障诊断算法在仿真平台上进行测 试，并应用到大型旋转机械系统的故障诊断中。 4) 构建了二进制粒神经网络模型，将粒计算与神经网络紧密融合，是一 种新型模型，也是理论研究的一个新课题。 关键字：粒计算,人工神经网络,故障诊断,约简 太原理工大学硕士研究生学位论文 III GRANULAR COMPUTINGNEURAL NETWORK ALGORITHM AND ITS APPLICATION IN THE ROTATING MACHINE FAULT DIAGNOSIS ABSTRACT The paper proposes the granular computing - - neural network algorithm, and applies it into the revolving mechanical fault diagnosis, after studying granular computing and artificial neural networks. Granular computing theory is proposed by American scholar Professor T.Y.Lin, and he once pointed out “Granular computing theory is a good tool of data reduction”. Granular computing theory is one new concept and the computation model about information processing, which has covered all related research about granules theories, methods, technical and tools. Granular Computing studies problems which are imprecise, partial true, fuzzy and unnumbered, based on different levels of granules. Granular computing as one kind of new intelligence information processing and management technology, belongs to one kind of “the soft computation”, and is receiving the domestic and foreign scholars widespread attention in recent years. Artificial Neural Networks is a kind of parallel information processing algorithms mathematical model which simulates animal neural network. It is the hot spot of artificial 太原理工大学硕士研究生学位论文 IV intelligence research. Artificial Neural Networks widely applies in each kind of domain because of artificial neural networks self-learning, self- organizing, good tolerance, and abilities of parallel processing information and so on. The granular computing - - neural network loose coupling algorithm is proposed in this article. The core thought is putting granular computing as neural network front processor to reduce the original sample set to obtain smallest attribute sets, namely minimum reduction. And then neural network was added into the fault diagnosis system using the reduction samples to train. The simulation result proves that reduction algorithm based on granular computing understands simply, decreases the computation, simplifies the networks architecture, and raises training efficiency. Moreover, one kind of new neural network- Binary Granular Computing Neural Networks (BGrCNN) is also proposed in this article, which is similar with the fuzzy neural network. Because this models inputs, outputs and operate process is using the binary number, it can enhance the training time greatly, therefore this model has displayed granular computing and neural networks advantage respectively. It is a research good topic. The main innovations of this paper are listed as follows: 1) Unify Granular computing based on a Rough set and the artificial neural networks for the first time to establish the granular computing - artificial neural networks model. Granular computing is one method of reducing the data, artificial neural networks is one method of classifying the data. Unifying two 太原理工大学硕士研究生学位论文 V theories can strengthen the processing data ability. 2) Propose the granular computing - artificial neural networks fault diagnosis algorithm to improve the intrinsic shortcoming when artificial neural networks apply in the fault diagnosis. 3) Test the granular computing - artificial neural networks fault diagnosis algorithm in the simulation platform, and apply it into the large-scale revolving mechanical systems fault diagnosis. 4) Construct the binary granule neural network model, fusing granular computing closely with neural network. It is one kind of new model, and is also a new research topic. KEY WORDS: Granular Computing, Artificial Neural Networks, Fault Diagnosis, Reduction 太原理工大学硕士研究生学位论文 XI 图索引 图 1-1 故障诊断系统模型6 Fig.1-1 The Model of Fault Diagnosis System6 图 1-2 故障诊断系统的实现框图7 Fig.1-2 The Schema of Fault Diagnosis System7 图 1-3 论文章节安排及其关系10 Fig.1-3 Frame Work of Dissertation.10 图 2-1 人工神经元模型示意图12 Fig.2-1 The Model of Neural Nerve F