【作者】 车勋建；

【导师】 胡清华；

【作者基本信息】 哈尔滨工业大学 ， 动力机械及工程， 2011， 硕士

【摘要】 近几年,人工智能,尤其是机器学习和模式识别技术大量应用于设备状态监测和故障诊断领域,采用智能技术检测和分析机械故障成为一种趋势。在故障分析中,用户除了需要知道某设备是否发生故障以及为何种故障外,还需获得故障的严重信息,从而制定适当的维修策略和检修方案。故障程度的智能检测本质上是有序分类问题:将故障程度用一组有序整数n表示(n=1, 2, 3,…),表示‘轻微故障’、‘中等故障’‘严重故障’等。相比于经典分类问题,有序分类的类别号(即故障程度)之间存在大小关系。由于存在这种序的关系,在分类器设计原则上以及分类器性能评价准则上都与经典的模式分类问题有所差别。在故障程度分析中,当表征故障严重性的特征值增加(或减少)时,故障程度随之增加,即故障特征与故障严重性之间存在单调性的约束。称这种能表征故障程度的特征为单调故障特征。这种单调性故障特征为故障程度诊断提供简便实用的信息,可以根据这种特征来衡量故障程度。本文从模式识别中的有序分类问题出发,研究了有序分类问题的本质和从数据中归纳学习有序分类模型的方法,提出2种有序分类规则学习算法,并将其应用于机械故障严重程度建模。具体内容如下:首先,本文系统地介绍了模式识别与机器学习领域中的有序分类问题,指出该问题与经典模式分类问题相比较,发展历史与研究深度相对简单,还有很大的研究空间。其次,介绍了单调分类中特征与决策之间的单调性约束。现有分类器只有在数据集是单调一致时才能训练出单调的分类模型,而现实中非单调噪声广泛存在,单调一致数据集很难得到。本文引入一种衡量特征与决策之间随机单调约束的指标,指出特征与决策是概率上的单调。介绍了有序信息熵模型,该模型继承了经典信息熵的鲁棒性,且能够反映特征与决策之间的随机单调相关性。最后基于该理论构造了基于有序信息熵的决策树学习算法。再次,针对部分特征与决策单调,部分特征不单调的决策问题,构造了随机有序混杂决策树算法。首先用非单调特征分裂数据,再用单调特征继续细化。为测试本文提出方法的性能,分别在人工数据和标准数据上用本文算法与其他经典有序分类算法进行了比较。结果显示,本文提出的两种算法鲁棒性好、泛化误差小。最后,将本文提出的算法应用于实际故障诊断。本文应用在齿轮裂缝故障程度监测实验中。在实验中,用位移传感器监测齿轮箱的振动,由此反映齿轮的故障严重程度。为区分故障等级,人为制造不同深度的齿轮裂缝。在不同负载和转速情况下得到振动数据,然后在频域与时域上提取了一系列特征得到故障数据。结果显示,本文方法分类损失低,展示了该方法的有效性。更多还原

【Abstract】 With the development of artificial intelligence, especially machine learning and pattern recognition, intelligent techniques in detecting of mechanical failure become popular. In practice, we should know not only whether a device is failure or not but also the severity of the failure. Machine systems should reach a certain level before shutdown: It will cause unnecessary waste if shutdown is too early, and will cause a serious accident if too late.In the field of pattern recognition, fault level detection can be seen as ordinal classification problem: the degree of failure can be represented by a group of integer n (n = 1, 2, 3, ...), means ’slight fault’, ’medium failure’, ’serious failure ’and so on. Compared with the classical classification, there exist ordinal relationships in decision class of ordinal classification (i.e. fault level). Due to this ordinal relationship, the design principle of classifier and evaluation criteria has to be reconsidered.Moreover, we consider a special task of fault level detection, in which when the value of a fault value increases(or decreases), fault level increases too. So there is a sort of monotonic constraint between fault features and fault levels. This type of fault features are named monotonic features, which can provide simple but useful information for detecting the level of mechanical failure.In this paper, we design two ordinal learning algorithms to deal with fault level detection, and the details are described as follows:Firstly, we give a systematic introduction to the field of ordinal classification. In literature, there are mainly 5 methods to solve this problem, that is ordinal classification, monotone classification, multiple criteria decision analyze, multi-objective optimize, and ordinal regression. We give a short introduction to each of these methods, and give some comments and cooperation between these methods.Secondly, we introduce monotonicity constraint between features and classes in ordinal classification. In literature, all learning algorithms can derivate monotonic models only if training samples are monotonic, which seldom happen in practice due to the presence of noise. To deal with this problem, we introduce a new constraint to describe the monotonicity between features and classes, namely, stochastic monotonicity constraint. We point out that features and classes are probabilistic monotone. Based on this idea, we introduce rank entropy model, which not only inherits the robustness of classical Shannon information entropy, but also reflects the probabilistic monotonicity between features and classes. Then we apply rank entropy to building decision tree, to deal with monotone classification and also give the theoretical proof of its properties.Thirdly, we recognize that not all features are monotonic. Some features have no such a monotonicity constraint. To solve this problem, Stochastic Hybrid Ordinal Tree (SHOT) is proposed. We conduct numerous experiment to test the performance of the two proposed methods on both artificial data and real-world data. One the one hand, compared with other classical classification algorithms, such as SVM, neural network, the proposed methods can be easy to understand by human beings, which helps us better understand the orderly classification of nature as well as the innate character of fault level detection. One the other hand, our methods are very competent among other ordinal and monotone methods. They are more robust, and produce low generalization error.Finally, we conduct the proposed methods on the practical application of the degree of gear crack fault monitoring experiments. The results show that our methods are very effective and efficient.更多还原