【作者】 谢珺；

【导师】 谢克明；

【作者基本信息】 太原理工大学 ， 电路与系统， 2009， 博士

【摘要】 随着信息技术的发展,人们采集到的数据日益膨胀,这些数据具有维数高、类别多的特征,如何从海量数据中挖掘出有用信息,成为人工智能面临的重要课题。神经网络擅长从数据中找到特定模式,但不能确定哪些特征是冗余的,哪些特征是有用的,难以评估特征的重要性。而新兴提出的粒计算能摆脱繁琐的、不重要的细节,抓住问题的本质,从适当的粒度层来求解问题,降低算法复杂度,快速获得问题的解。将粒计算与神经网络融合研究具有前沿性、科学性及优越性。故障诊断可以看做是一个从海量的故障数据中挖掘关键特征,进而对故障样本进行分类识别的过程。本文将粒计算和神经网络结合,并应用于机械的故障诊断,其研究成果为粒计算和神经网络理论增添了新的内涵,拓展了粒计算的应用领域,为工业领域的故障诊断提供了一种新的思路和方法。本文的研究内容属于信息科学、自动化科学、计算机科学等学科的交叉领域。本文的主要创新性工作包括:(1)提出一种基于粒计算的二进制可辨矩阵属性约简算法。引入粒度层的观点,把二进制可辨矩阵的不同列组合看作不同的粒度层,通过不同列之间的访问实现了不同粒度层的跳转。针对过热器模型参数因工况参数变化而变化的问题,且模型参数与工况参数间的关系复杂,将所提出的约简算法用来挖掘工况参数与模型参数间的关系和影响规律,实现对过热器的建模。实验结果表明该算法计算量少,能够快速挖掘到规则,且易于编程实现。(2)提出一种基于粒度混合系统的控制方法。定义了粒度混合自动机的概念,针对倒立摆的强耦合性和自然不稳性的特点,将倒立摆的起摆和稳定过程划分成3个粒度世界,并给出不同粒度世界之间跳转的阈值条件。通过在3个粒度世界间的跳转,实现对倒立摆系统的稳定控制。仿真结果表明基于粒度混合系统的控制方法能够在较短时间内摆起倒立摆,并稳定在目标位置。(3)提出一种基于二进制粒矩阵的属性约简算法(Binary Granular Matrix-based Attribute Reduction, BGMAR)。该算法由2个子算法构成:针对决策表存在重复和矛盾对象的情况,提出一种基于二进制粒矩阵的论域约简算法,经过论域约简后的决策表对象数目减少,缩小了进行最优属性约简的样本空间;针对现有的属性约简算法大多无法找到最优约简的现状,提出一种基于二进制粒矩阵最优属性约简算法,在论域约简的基础上,通过计算各个条件属性的重要度,求出属性核;在属性核的基础上,采用决策属性集对条件属性集的依赖度作为启发信息进行搜索,直到得到最优属性集。设计了2个算例对BGMAR算法的有效性进行验证,结果表明该算法不仅对一致决策表的最优属性约简有效,而且还适用于不一致决策表约简,并且能得到其最优属性集。(4)提出一个二进制粒神经网络模型框架,并在该框架下提出二进制粒神经网络分类器(Binary Granular Neural Network Classifier, BGNNC)及其算法。将粒计算的降维与约简能力与神经网络固有的并行处理与学习能力相结合,构建粒计算作为神经网络前端处理器的二进制粒神经网络模型,提出一种二进制粒神经网络分类器的7元组模型及其分类算法。利用BGMAR算法约简特征空间,找到最优特征后,采用动量—自适应学习率BP算法训练BGNNC实现模式分类。选取UCI机器学习数据库中的Breast Cancer Wisconsin (Diagnostic)、Iris、Wine、Zoo 4个标准数据集进行测试,将BGNNC算法与标准BP算法进行了分类结果的比较。实验结果表明BGNNC是一种有效的分类算法,具有较好的泛化能力,不仅可以解决2分类问题,在多分类任务上也是可行且正确率较高的一种方法。(5)将本文提出的BGNNC算法应用于机械故障诊断。针对往复式机械(内燃机)和旋转式机械部件(滚动轴承)各自的故障机理,利用BGNNC算法对这两类机械设备进行故障诊断。设计了验证BGNNC算法有效性的对比实验,分别采用BP算法、免疫群体网络(Immune Population Network, IPN)算法对上述工程实例进行故障诊断,并比较上述3个算法的诊断结果。仿真实验表明BGNNC算法训练速度快,诊断准确率高,不失为一种可行的、高效的智能诊断方法。更多还原

【Abstract】 With the development of information technology, the acquired data are sharply increasing, which own the features as high dimension and multiple target classes. It must face for us how to mine the useful information from the mass data. Artificial neural networks are good at find the specific patterns in data, however it cannot know which features are redundant and which are key. On the other hand, GrC latest proposed by T. Y. Lin can get rid of tedious and unimportant details and grasp the essence in data. It solves problem from proper granule levels, which reduces computational complexity and obtains satisfactory solutions to the problem rapidly. It belongs to the preceding research which combines GrC and ANN.Fault diagnosis can be regarded as such procedure that mines the key features from the fault data and then finds out the fault patterns hidden in them. This dissertation took GrC as the front-end processor of ANN and set up a framework, namely a binary granular neural network. Under the framework, a binary granular neural network classifier, BGNNC in short, has been modeled and used to solve the mechanical fault diagnosis problem. The achievements in the dissertation add new connotation both to GrC theory and ANN theory, expand application regions of GrC and provide a new idea for fault diagnosis in industry field. The research contents of the dissertation belong to the crossing field of information science, automation science and computer science et al.The main innovative works in the dissertation include:(1) Designed a granular computing-based binary discernibility matrix attribute reduction algorithm. A granular layer viewpoint was introduced into the binary discernibility matrix. It took the different collums combinations as the different granular layers. The jump between different granular layers was realized by combining different collums. Aiming at the problems which the model parameters of superheater vary with the change of working parameters, the proposed algorithm was used to mine the relation and rules between working parameters and model parameters so as to build the model of the superheater. The simulation results show that the proposed algorithm has less calculation, extracts the ruls fast and is easy to implementation with programme.(2) Proposed a granular hybrid system-based control method. The core of it was the definition of granular hybrid automata. Inverted pendulum system is a nonlinear unstable system. According to its strong coupling, the swing and stabilizing procedure can be divied into three granular worlds. The threshold conditions of jumping between different granular worlds were given. It can fulfil the stability control of the inverted pendulum through the jump among the three granular worlds. The control simulation results indicated that the proposed control strategy can swing the inverted pendulum and stabilize it at the goal position in a short time.(3) Presented a binary granular matrix-based attribute reduction algorithm (BGMAR in short). BGMAR algorithm consists of two sub-algorithms which are binary granular matrix-based universe reduction (BGM-UR) and binary granular matrix-based minimum reduction (BGM-MR). Due to the repeated or contradictory objects in the universe, the BGM-UR algorithm can delete such objects and shrank the sample space where the BGM-MR will work on. Aiming that the most existing reduction algorithms cannot find the minimum attributes set, the BGM-MR find the core by calculating the wightness of each condition attribute and exploits the dependency degree as heuristic information to search the minimum reduction. The BGMAR algorithm is evaluated effective by two examples. It concludes that BGMAR algorithm is suitable not only to consistent decision table but also to inconsistent decision table, which can get the minimum reduction for both.(4) Built a framework of binary granular neural network. It takes GrC as the front-end processor of ANN. A binary granular neural network classifier (BGNNC) was proposed under the framework. It was defined as a septuple. BGNNC algorithm reduces the feature space and can find the key features, and then uses the additional momentum adaptive learning rate adjustment to train BGNNC, finally realizes the pattern classification. Breast Cancer Wisconsin (Diagnostic), Iris, Wine and Zoo datasets from UCI database are chosen to test the effectiveness of BGNNC algorithm. The results were compared with standard BP classification algorithm and shew that BGNNC has higher classification precision and better generalization than standard BP algorithm.(5) Applied the BGNNC algorithm to solve the mechanical fault diagnosis problem. In light of respective fault mechanism of the internal combustion engine (as the example of reciprocating machine) and the rolling bearing (as the example part of rotating machine), BGNNC was applied to diagnose the faults in them. The contrast experiment was designed to compare BGNNC algorithm with BP algorithm and Immune Population Network algorithm. The experiment results indicated that BGNNC has fast training speed and high diagnosis accuracy. It is proved that BGNNC is a feasible and efficient intelligent diagnosis method.更多还原