【作者】 李鑫；

【导师】 杨宇；

【作者基本信息】 湖南大学 ， 机械工程， 2022， 博士

【摘要】 旋转机械长期服役于极端恶劣的工况下,其运行状态不可避免地会发生退化,甚至产生故障。一旦旋转机械中某一关键零部件发生故障,如轴承、齿轮、转子等,将会影响整个机械系统的正常运作,造成非计划性停机,降低生产效率,甚至造成重大的生产安全事故。因此,开展旋转机械故障诊断研究对于保障设备服役质量、提高系统可靠性和安全性具有十分重要的工程应用价值。随着监测技术和计算机技术的迅猛发展,基于机器学习的旋转机械智能诊断方法受到了国内外相关学者的广泛关注,并取得了一系列可喜的研究成果。在众多机器学习方法中,以超球体、凸包、超圆盘等几何模型为基础的分类方法由于具有坚实的理论基础和出众的几何可解释性,在故障诊断方面具有一定的优越性。然而,现有几何模型分类(Geometrical Model Classification,GMC)方法大多存在噪声鲁棒性不足、多通道信息融合能力弱、无法直接处理高阶复杂数据等局限性。为此,本文在国家自然科学基金项目(编号:51875183、51975193)的资助下,以旋转机械及其核心零部件作为研究对象,对GMC方法进行了深入地研究和探索,提出了多种改进型凸包和超圆盘分类模型,为促进旋转机械智能故障诊断的进一步发展提供了新的可能。论文的主要研究工作如下:(1)支持矩阵机(Support Matrix Machine,SMM)从几何角度来看是一种矩阵形式凸包分类模型,能够直接处理矩阵形式数据,以挖掘数据中行与行或列与列间的结构信息。然而,SMM在处理复杂矩阵数据时存在一定局限性,如XOR型数据。为此,提出了非平行最小二乘支持矩阵机(Non-Parallel Least Squares Support Matrix Machine,NPLSSMM)模型。与现有矩阵形式分类模型不同,NPLSSMM通过构造一对非平行分类超平面,以最大间隔分类不同类别样本。其中,每一个超平面需尽可能的靠近某一类样本的凸包,同时要尽可能的远离另一类样本的凸包,这种非平行分类策略使分类超平面能根据样本类别分布自适应地调整位置,极大增强了模型的灵活性和泛化性。此外,矩阵形式最小二乘约束的引入明显提高了模型的运行效率。实验结果表明,所提NPLSSMM模型具有良好的滚动轴承故障诊断性能。(2)针对现有凸包分类模型鲁棒性不足、不平衡数据处理能力弱的问题,提出了一种鲁棒不平衡凸包分类(Robustness imbalanced convex hull-based classification,RICHC)模型,用于锥齿轮箱的智能故障诊断。首先,RICHC根据不同样本在类别分布估计中的作用,确定各样本的置信函数,降低异常值和含噪样本的权重,使凸包模型边界更加紧致,以提高模型的鲁棒性。同时,构建自适应模型缩放策略,使RICHC根据多数类和少数类间的动态不平衡因子,调整不同类别间凸包的缩放比例,得到更加准确的分类超平面,从而提升模型的不平衡数据处理能力。采用锥齿轮箱故障数据验证所提方法的有效性及适用性,实验结果表明相较于其他算法,所提方法对噪声和异常值具有更强的抗干扰能力,且该方法具有更优异的类不平衡分类性能。(3)针对传统机器学习方法在处理对称正定(Symmetric Positive Definite,SPD)矩阵数据时的不足,提出了黎曼最大间隔弹性凸包(Riemannian Maximum Margin Flexible Convex Hull,RMMFCH)分类模型。该模型采用黎曼度量评估样本分布,并在黎曼几何空间中构建最优分类超平面,以最大间隔分离不同类型凸包。此外,为了充分挖掘原始数据的全局-局部信息,并保留局部特征间的相互作用,提出了基于统计增强协方差矩阵(Statistical-Enhanced Covariance Matrix,SECM)的故障特征表示方法。将SECM与RMMFCH相融合,构建了基于SECM和RMMFCH的轴承故障诊断方案,并利用轴承故障数据验证了所提方案的有效性和适用性。实验结果表明,SECM能够有效表征轴承的敏感故障信息,且RMMFCH的故障诊断性能要优于传统基于欧式度量的故障诊断模型。(4)针对多通道监测数据下的信息融合问题,提出了基于多通道融合协方差矩阵(Multi-Channel Fusion Covariance Matrix,MFCM)的滚动轴承故障特征表示方法。此外,为了充分利用MFCM的黎曼流形结构信息,将黎曼流形的概念引入超圆盘模型,提出了最大间隔黎曼流形超圆盘(Maximum Margin Riemannian Manifold-Based Hyperdisk,MMRMHD)分类模型,以构建MFCM与轴承故障的非线性映射关系,从而智能识别滚动轴承的不同故障类型。实验结果表明,基于MFCM和MMRMHD的故障诊断方法能够有效融合多通道信息,并准确判别不同轴承故障类型,具有良好的故障诊断效果。(5)将超圆盘几何模型引入单分类领域,提出了单分类超圆盘(One-Class Hyperdisk,OCHD)模型,并将其与辛主成分分析(Symplectic Principal Component Analysis,SPCA)方法相结合,提出了基于SPCA和OCHD的旋转机械故障检测方法。该方法首先采用SPCA将振动信号映射到辛空间,提取最能表征信号主要能量的辛特征值(Symplectic Eigenvalue,SE)作为特征向量。随后,利用SE特征构建OCHD模型,以实现旋转机械的智能故障检测。实验结果表明,OCHD模型能够有效检测出旋转机械早期微弱故障,具有良好的工程应用前景。更多还原

【Abstract】 Rotating machinery chronically serves under extremely harsh working conditions,and thus its operating state will inevitably degrade or even fail.Once the key components in rotating machinery cause faults,such as bearings,gears,rotors,etc.,it will affect the normal operation of the entire mechanical system,resulting in unplanned shutdowns,reduced production efficiency,and even serious production safety accidents.Therefore,the research on rotating machinery fault diagnosis has very important engineering application value for ensuring the service quality of equipment and improving the reliability and safety of the system.With the rapid development of monitoring technology and computer technology,the intelligent diagnosis method of rotating machinery based on machine learning has received extensive attention from relevant scholars at home and abroad,and a series of gratifying research results have been achieved.Among numerous machine learning methods,the classification methods based on geometric models,such as hypersphere,convex hull,and hyperdisk,have obvious advantages in fault diagnosis due to the solid theoretical foundation and outstanding geometric interpretability.However,most of the existing geometric model classification(GMC)methods have the shortcomings of insufficient noise robustness,weak multi-channel information fusion ability,and inability to directly deal with high-order complex data.Therefore,under the support of the National Natural Science Foundation of China(No.51875183,51975193),this dissertation takes rotating machinery and its core components as the research object,conducts in-depth research and exploration on GMC methods,and proposes a variety of improved classification models based on the geometric models of convex hull and hyperdisk,providing new possibilities for further development of intelligent fault diagnosis of rotating machinery.The main research contents of this dissertation are concluded as follows:(1)Geometrically,support matrix machine(SMM)is a matrix-form convex hull classification model,which can directly process matrix data to mine their structural information between rows or columns.However,SMM has poor classification performance when dealing with complex matrix data,such as XOR data.To this end,a non-parallel least squares support matrix machine(NPLSSMM)model is proposed in this dissertation.Different from the existing matrix-form classification models,NPLSSMM constructs a pair of non-parallel classification hyperplanes to classify samples of different classes with the maximum margin,where every hyperplane is required to be as close as possible to the samples of one class while being as far as possible from other samples.This nonparallel classification strategy enables the classification hyperplanes to adaptively adjust their position according to the category distribution of samples,largely enhancing the flexibility and generalization of the proposed model.In addition,the designed matrix-form least-squares constraints significantly improve the computation efficiency of NPLSSMM.The experimental results show that the proposed NPLSSMM model has good fault diagnosis performance of rolling bearings.(2)Aiming at the problems that the existing convex hull classification models have insufficient robustness and weak ability to process unbalanced data,a robustness imbalanced convex hull-based classification(RICHC)model is proposed for intelligent fault diagnosis of bevel gearboxes.First,according to the role of different samples in class distribution estimation,a confidence function is designed for RICHC to reduce the weights of outliers and noisy samples,which will make the boundary of the convex hulls more compact and improve the robustness.At the same time,an adaptive scaling strategy is constructed for RICHC to control the scaling of the convex hulls between different classes,and the scaling ratio is determined by the dynamic imbalance factor between the majority class and the minority class.Based on this strategy,a more accurate classification hyperplane will be obtained for RICHC to improve the unbalanced data processing ability.The effectiveness and applicability of the proposed method is verified on bevel gearbox fault data,and the experimental results show that compared with other models,the proposed method has stronger anti-interference ability against noise and outliers,and the method has more excellent class imbalance classification performance.(3)Aiming at the shortcomings of traditional machine learning methods in classifying symmetric positive definite(SPD)matrices,a Riemannian maximum margin flexible convex hull(RMMFCH)model is proposed.RMMFCH uses the Riemannian metric to evaluate the sample distribution,and constructs the optimal classification hyperplane in the Riemannian geometric space to separate different types of convex hulls under the principle of maximum margin classification.In addition,a fault feature representation method based on statistical-enhanced covariance matrix(SECM)is proposed to fully mine the global-local information of raw data and preserve the interaction between local features.By integrating SECM and RMMFCH,a bearing fault diagnosis scheme based on SECM and RMMFCH is constructed in this paper.The effectiveness and applicability of the proposed scheme are verified by two bearing fault datasets,and the experimental results show that SECM can effectively characterize the sensitive fault information of bearings,and the fault diagnosis performance of RMMFCH is better than traditional fault diagnosis models based on Euclidean metric.(4)Aiming at the problem of information fusion under multi-channel monitoring data,a fault feature representation method of rolling bearing called multi-channel fusion covariance matrix(MFCM)is proposed.In addition,to make full use of the manifold structure information among MFCM,the concept of Riemannian manifold is introduced into the geometric model of hyperdisk,and a maximum margin Riemannian manifold-based hyperdisk(MMRMHD)model is proposed to construct the nonlinear mapping relationship of MFCM and bearing faults,so as to intelligently identify different faults of rolling bearings.The experimental results show that the fault diagnosis method based on MFCM and MMRMHD can effectively achieve multi-channel information fusion,and accurately distinguish different bearing faults with outstanding fault diagnosis performance.(5)A one-class hyperdisk(OCHD)model is proposed by introducing the hyperdisk geometric model into the one-class classification field.By integrating symplectic principal component analysis(SPCA)and OCHD,a fault detection method of rotating machinery is designed in this dissertation.First,SPCA is used to map raw vibration signals into a symplectic space,and then the symplectic eigenvalues(SE)that can best characterize the main energy of the original data are extracted as insensitive eigenvectors.Subsequently,OCHD is constructed with the extracted SE features to achieve the intelligent fault detection of rotating machinery.The experimental results show that the proposed OCHD model can effectively detect the early weak faults of rotating machinery,and has good engineering application prospects.更多还原