【作者】 杨先勇；

【导师】 周晓军；

【作者基本信息】 浙江大学 ， 机械制造及其自动化， 2009， 博士

【摘要】 人们对现代化生产系统运行可靠性和安全性越来越高的要求,是机械故障诊断技术产生并迅速发展的原因。随着机械设备向高速化、重载化、大型化和和复杂化方向发展,传统故障诊断技术越来越难以满足诊断需求。信号分析技术和人工智能方法的发展为提高故障诊断水平提供了有力工具。本文结合“车辆变速箱检测试验台”、“车辆分动器检测试验台”、“车辆变速箱齿轮检测试验台”等项目,针对机械设备发生故障时振动信号往往表现出时变性、非平稳性特点,以齿轮和轴承为对象,围绕故障特征信息提取和智能诊断方法这两个核心问题开展研究,重点在于基于信号局部特征的分析方法和基于AIS(Artificial Immune System)的智能诊断模型及其应用。主要工作和研究成果如下:(1)介绍了研究对象的主要故障形式、振动产生机理和信号特点,以及实际中振动信号的测量和传播路径的影响,为后续章节奠定基础。(2)针对测得的信号中伴有背景信号和噪声,研究了形态小波降噪和特征提取方法。基于形态小波的一般框架构造了极值提升形态小波和复合结构元素形态非抽样小波两种形态小波,将其用于冲击信号特征的提取,具有比传统小波和解调分析更强的特征提取能力;将极值提升形态小波应用于齿轮和轴承的小波灰度矩分析,结果表明极大改善了其对信号局部时频能量分布特征的刻画能力。(3)针对一般方法的基对信号局部特征不具有自适应性,研究了基于局域波的特征提取方法。提出了一种改进局域波分解特征提取性能的方法,根据局域波互信息剔除噪声分量和虚假分量,并引入形态小波分析,以抑制模态混叠和虚假分量,提高了信号分解的准确性和瞬时参数提取的时效性。在此基础上,提出局域波域的信息熵特征分析方法,用于定量刻画信号在基本模式空间中分布的复杂度。用实例验证了其有效性。(4)针对特征的重要性不同及其相关性和冗余性,结合局域波分析,从特征提取和特征选择两个方面研究了获得对分类有利的特征子集、改善分类性能的方法:核主元分析-最小二乘支持向量机和贝叶斯证据框架-序列后向选择-最小二乘支持向量机。前者通过核主元分析将数据映射到高维特征空间,在抑制冗余度和噪声的基础上进行特征提取;后者运用启发式搜索策略,在分类器学习过程中实现自适应的多特征子集选择优化。用实例验证了以上方法在提高分类器的学习和泛化性能上的有效性。(5)AIS本质上是模拟生物体自身的故障诊断机制,具有很好的可解释性。基于免疫识别和免疫网络隐喻机制,提出了基于V-detector算法的故障诊断模型和RS-ABNet故障诊断模型,前者适于处理低维空间问题,后者在处理高维空间问题时更具优势。通过对齿轮和轴承的故障诊断,证实了以上方法的实用性和有效性,为机械故障诊断提供了新的智能方法。更多还原

【Abstract】 The higher and higher demands for operational reliability and safety of the modern production system is the reason for generation and rapid development of the mechanical fault diagnosis technology. With the development of the mechanical equipment towards high speed, heavy load, large scale and complication, traditional fault diagnosis technology can’t meet the requirement for equipment diagnosis. Development of signal analysis and artificial intelligence technology supplies powerful tools for improving the fault diagnosis level. Focusing on time-varying and non-stationary feature of vibration signal when faults occur, based on projects of vehicle gear-box testing system, vehicle power transfer testing system, vehicle gear-box gear testing system, and on the research subjects of gear and bearing, research was carried out around the two core problems of fault diagnosis, feature extraction and intelligent diagnosis method. Research keys were the analysis method based on signal local characteristics and intelligent diagnosis model and its application based on artificial immune system (AIS). The main research works are as follows:(1) The main fault types, vibration generating principle and signal feature of the research subjects were introduced, and influences of practical measure and transmitting path on vibration signal were analyzed, which is the basis for latter chapters.(2) Focusing on practical data interfered by noises and background signal, the denoising and feature extraction methods based on morphological wavelet were researched. Two morphological wavelets, minimax-lifting morphological wavelet (MLMW) and multi-element morphological undecimated wavelet decomposition (MMUWD), were constructed based on the general frame of morphological wavelet, then they were used to extract shocking signal’s feature, show stronger feature extraction capability than that of traditional wavelet and demodulation analysis. Using MLMW to the CWT grey moment analysis of gear and bearing, results show that MLMW greatly improves the describing capability of CWT grey moment for signal’s local time-frequency energy distribution characteristics.(3) Aiming at general analysis methods’ basis being not adaptive for signal’s local feature, the feature extraction method based on local-wave was researched. A method improving the feature extraction capability of local-wave was proposed, in which noise components and pseudo-components were removed by calculating local-wave mutual information, and morphological wavelet analysis was combined to constrain mode mixing and pseudo-components. With which the decomposition quality and feature extraction capability are improved. Then the information entropy feature analysis method based on local-wave was proposed, which was used to describe signal’s complexity in the intrinsic mode space. Simulation and practical data prove the proposed methods’ effectiveness.(4) Focusing on features’ different importance, correlation and redundancy, and combining local-wave analysis, methods of obtaining feature subset preferable for classification and improving classification ability were introduced from feature extraction and feature selection, and the models of KPCA(kernel principal component analysis) - LSSVM (least squares support vector machines) and BEF(Bayesian evidence framework)- SBS (sequential backward selection)- LSSVM were proposed correspondingly. The former mapped data to high dimension feature space through KPCA, and extracted features on the basis of constraining redundancy and noises. While the latter adopted heuristic searching strategy, and realized adaptive multi-feature subsets selection and optimization. The practical data proves the proposed methods’ effectiveness in improving learning and generalization performance of LSSVM.(5) In nature, AIS stimulates biological fault diagnosis mechanism for itself, and has good interpretation. Based on immune identification and immune net metaphor mechanism, the fault diagnosis models based on V-detector algorithm and RS-ABNet fault diagnosis model were proposed. The former is suitable for processing low dimension space problem, while the latter has advantage in processing high dimension space problem. The fault diagnosis of gear and bearing prove the practicability and effectiveness of the above methods, and it will supply a new intelligent method for mechanical fault diagnosis.更多还原