【作者】 周志红；

【导师】 周新聪； 严新平；

【作者基本信息】 武汉理工大学 ， 载运工具运用工程， 2007， 硕士

【摘要】 摩擦学系统的状态监测与辨识日益引起学术界与工业界的重视和关注，以润滑油中的磨损颗粒为研究对象的磨粒分析技术是开展摩擦学系统状态描述和辨识研究的一种有效方法。由于磨粒的多样性和复杂性，磨粒的识别长期依赖专家的经验，导致磨粒分析技术的推广应用受到较大限制。随着显微镜技术、计算机技术、传感器技术、信息技术及现代数学理论的长足进步，磨粒分析及识别技术虽然已经向智能化、自动化方向发展，但在磨粒特征参数的选择优化以及识别方法等方面的研究还存在着不足。如何建立磨粒特征参数描述体系是磨粒分析领域内的一大难点，本文从磨粒特征集内在的关联性出发，在对磨粒特征的相关性和冗余性分析的基础上，针对磨粒特征参数评价优化问题，设计了两种不同的特征选择算法——Recorre算法和遗传算法(GA)。研究表明：这2种算法对磨粒特征空间都能有效的约简。Recorre算法根据磨粒训练样本的统计特性进行特征选择，效率高，所得到的特征子集通用性强，但精度一般，在数据的处理量很大、系统的实时性要求高时可考虑采用Recorre算法；遗传算法根据磨粒识别系统的准确率来选择特征子集，所得到的特征子集是使系统性能最好的特征子集，因而能获得较高的系统识别率，但该方法效率低，适用于对系统识别率要求高的场合。集成学习通过将多个分类器以某种方式集成，可以显著提高磨粒分类的准确性。本文将集成学习理论引入磨粒的自动识别研究，研究了基于Bagging集成技术的磨粒识别方法，结果表明该方法可以进一步提高现有磨粒识别方法的准确率。在上述研究的基础上，结合GA特征选择和Bagging算法的优点，建立了一种新的基于GA特征选择和Bagging集成技术的磨粒识别模型——GA-Bagging，该方法在构建个体分类器的同时选择对磨粒识别最有效的特征，通过该算法可以有效去除磨粒特征集中的冗余特征和不相关特征，并在新的更少的数据集上进行建模。结果表明该方法可以显著提高现有磨粒识别系统的泛化能力和准确性，对完善和发展磨粒自动识别系统的理论、技术和方法提供了新的思路和解决方案。更多还原

【Abstract】 Tribosystem condition monitoring and recognition has been paid more and more attention in academia and industry. Analysis of wear particles in lubricating oil has been recognized as one of the most effective methods in the study of tribosystem condition description and recognition. However, classification of types of wear particles was usually examined by experts in the field due to the diversity and complexity of wear particles, which limited its industrial applications. With the advancement of microscope, computer, sensor technology, information technology and modern mathematical theory, although wear particle analysis tends to head for an automated, intellectualized stage, there are still insufficient for practical uses in existing automatic wear-particle classification systems with respect to wear-particle image processing, feature optimization and pattern recognition.One challenging issue in wear particle analysis is to establish the effective characteristic parameter system to describe wear particles properly using a few numerical features. In this paper, feature relevance and feature redundancy of wear particle are defined according to their intrinsic relationship between wear-particle features. To solve the optimization and selection of wear particle parameters, thepaper proposed two different algorithms------Recorre and genetic algorithm (GA). Italso analyses the advantage and application area of each method respectively. It is believed that the developed methods can reduce dimensionality of input features significantly. Recorre algorithm relies on general characteristics of training data to select some features without involving any learning algorithm, while GA requires one predetermined wear-particle classifier in feature selection and uses its performance to evaluate and determine which features are selected. GA can perform better than Recorre but it trends to be computationally more expensive. Recorre algorithm has the advantage of high speed and ability to scale to large datasets but its performance is moderate. GA is applicable to the cases when the accuracy is preferentially considered while Recorre is more practical if the number of instances becomes very large. Ensemble learning is an active field in intelligent learning field, which combines multiple component learners that have been trained in the same task to improve the accuracy of classification systems. The paper applies ensemble learning to classify wear particles and presents a new ensemble method, i.e. GA-Bagging for automated wear particle analysis which combines the feature selection technique, GA, with ensemble learning technique, Bagging algorithm. The proposed method can improve the accuracy and generalization performance of automated wear-particle classification systems significantly and provides a new approach for automatization and intelligentization of wear particle analysis.更多还原