【作者】 王广斌；

【导师】 刘义伦；

【作者基本信息】 中南大学 ， 机械电子工程， 2010， 博士

【摘要】 故障诊断需解决的基本问题是根据传感器采集到的机械设备运行状态信号,提取特征参量,设计决策函数,最终求出其故障状态,核心是特征提取和模式识别问题。由于机械设备运行状态复杂、工作环境恶劣、运行时间较长,因此其状态信号具有数据量大、非线性程度高、噪声干扰强等特性,人们对这些庞大而复杂的数据信息地驾驭和处理越来越难,表现在一方面我们可以获取的数据量越来越大,而另一方面却难以得到更多有助于决策的信息。从2000年开始发展起来的流形学习方法,将数据分析与状态决策从欧氏空间扩展到流形,从而能够从分布在高维流形上的数据集中高效快速地挖掘出数据的本质特征,找到数据产生的内在规律,达到准确故障诊断的目的。论文主要进行了以下研究工作：在非线性降噪方面,提出三种流形学习降噪方法。基于本征维数的局部切空间降噪方法,直接将高维相空间数据直接约简到信号主流形所在的本征维数空间上,再反求一维信号实现对信号的降噪,避免了主流形提取时约简目标维数的盲目性。局部切空间均值重构降噪方法,将局部切空间降噪后的低维数据在全局范围内通过求取各点均值的方法重构到原高维空间,相当于对各点的局部坐标进行均值处理,实质是二次降噪,避免了相空间数据在全局排列过程中出现相点畸变的问题。基于高阶累积量的局部切空间降噪方法,利用高阶累积量理论上可完全抑制高斯有色噪声干扰的特性,用四阶累积量函数代替二阶矩函数构造协方差矩阵,提高了对含有色噪声信号的降噪效果。为解决了局部Fisher判别分析求解不对称特征方程时得出的投影基向量不正交使得数据重建困难的问题,提出了基于迭代正交和Schur正交的局部Fisher判别方法。通过迭代正交或Schur正交分解的方法构建正交基函数,可有效保留故障信号流形空间中的与近邻距离有关的结构信息,并在主特征求取的过程中,保留类别信息,使提取的主特征量能在尽量保持甚至降低类内散度的同时,使得类间信号特征量之间的距离尽可能远离,进而更好地实现故障诊断。将核方法引入正交局部Fisher判别中,提出了基于核的迭代正交和Schur正交局部Fisher判别方法,通过非线性核函数将信号投影到高维特征空间,在此空间进行正交局部Fisher判别分析,进行故障特征提取,实现了线性流形学习方法到非线性方法的转变,取得了比线性正交方法更好的故障诊断效果。以局部边界邻域点来构建Fisher判别函数进行故障特征提取和诊断,提出局部边界Fisher判别方法,直接利用邻域空间边界点对来计算局部类内散度和类间散度,大大提高了方法效率。为避免伪边界点干扰,还设计了用模糊聚类来寻找真实局部边界的方法。用核方法实现了局部模糊聚类边界Fisher判别由线形向非线形方法的转变,基于核的方法具有更强的故障诊断能力。在监督流形学习方面,对增量局部切空间排列和线性局部切空间排列方法进行改进,并引入了非线性支持向量机分类器,提出了监督增量局部切空间排列-支持向量机和监督线性局部切空间排列-支持向量机故障诊断方法,既解决了非线性流形学习的泛化能力不足的问题,又增强了流形学习方法的故障诊断能力。更多还原

【Abstract】 The basic problem of fault diagnosis is to obtain fault status by extracting feature, design decision functions based on information of equipment running,feature extraction and pattern recognition is the core of the problem.Machinery generally runs in a complex and poor working conditions,so the state signal has large amount of data, high nonlinearity characteristic, strong noise and interference.We are more and more difficult to manage and process these large and complex data, specific performance is that on the one hand we can get the amount of data, on the other hand it is hard to get more help in decision. Since 2000, manifold learning methods begin to develop and become research focus of the machine learning and pattern recognition. This method extends from euclidean space to manifold space in data analysis and state decision, efficiently and quickly to dig out the essential characteristics of the data from the high-dimensional data sets, to find the internal laws of data, and to achieve an accurate diagnosis.The main work include the following:In the perspective of nonlinear noise reducation, three manifold learning methods are proposed.In local tangent space alignment algorithm based on the intrinsic dimension, at first the intrinsic dimension of signal is obtained, and then the data in high dimensional phase space are reducated to the intrinsic dimension space, at last one-dimensional signal is obtained by reverse process. Algorithm avoid from blindness of dimension reduction targets’ selection, improve the efficiency of noise reduction. In local tangent space mean reconstruction algorithm, low dimension data after noise reducation in local tangent space are reconstructed to the high dimension data by obtaining the mean of each point in global space. Algorithm’s nature is the second noise reduction, not only enhancing effect of noise reduction,but also avoiding from the distortion of phase space data in the course of the global arrangement. Making use of restraining characteristic to colored noise of high-order cumulan,covariance matrix is constructed with a fourth-order cumulant function instead of construct second-order moment function covariance matrix,local tangent space alignment algorithm based on fourth-order cumulan is also proposed. This algorithm improves effect of noise reduction to signal with colored noise.In local fisher discriminant analysis, projection basis vectors obtained by calculating asymmetric the characteristic equatio are no-orthogonal,this leads to be difficult to data’s reconstruction. To solve this question,iteration orthogonal and schur orthogonal local fisher discriminant methods are proposed. Orthogonal local fisher discriminant algorithm may effectively preserve the structure information of nearest neighbors in manifold space, and in the prosess of main features’seeking, class information are retained,and then main features obtained can maintain or even reduce the with-class divergence of the same category sample, at the same time make between-class distance becaome far, better achieve fault classification.In this paper,kernel method is introduced orthogonal local fisher discriminant analysis, iterative orthogonal and schur orthogonal local fisher fault diagnosis algorithm based on kernel method are proposed. Feature signals are projected into the high dimensional kernel space by nonlinear kernel function, and make orthogonal local fisher discriminant analysis in this space. Algorithm has achived transformation from linear to nonlinear method, and obtain better effect than linear orthogonal fault diagnosis.In the perspective of fault diagnosis based on the concept of local margin, local fuzzy clustering margin fisher discriminance is proposed. The fisher discriminant function is built by directly computing local within-divergence and between-class divergence using local margin points in neighborhood, instead of using all points, greatly increased the efficiency of the algorithm. In order to avoid from using possible pseudo margin points, a method is proposed ny means of fuzzy clustering algorithm to find the real local boundary.Meanwhile,by kener method local fuzzy clustering margin fisher discriminance becomes non-linear algorithm,and has better fault diagnosis ability.In the perspective of supervised manifold learning, increment local tangent space alignment(ILTSA) and linear local tangent space alignment(LLTSA) algorithm are improved, and nonlinear support vector machine(SVM) classifier is introduced, supervised ILTSA-SVM and supervised LLTSA-SVM are proposed. Two algorithms increaze generalization ability and fault diagnosis ability of the non-linear manifold learning.更多还原