基于HHT和决策树的滚动轴承故障诊断研究

发布时间：2018-06-01 13:57

本文选题：滚动轴承 + 特征提取　；参考：《华中农业大学》2011年硕士论文

【摘要】：由于轴承的寿命具有很大的离散性和不可预知性,对其进行故障诊断与实时监测来保证机械设备安全运行尤为重要。轴承故障诊断技术包括信号采集、特征提取、故障模式识别等环节,是一门包括机械、信号处理、人工智能、传感器等技术的交叉学科。本文中以滚动轴承6205为例,对其进行故障诊断分析,其主要故障形式有内圈故障、外圈故障、滚动体故障三类。主要研究轴承振动信号的故障特征提取方法、特征降维方法及多故障诊断算法在诊断技术中的应用,并将些算法在MATLAB GUI工程中实现,开发出故障诊断软件。特征提取是故障诊断技术的一个重要环节,对其进行选择的特征应该能准确而敏感地反映轴承的工作状态,本文结合相关文献利用以下几种方法综合提取特征：计算并提取信号的时域统计量作为特征,即常用的各阶中心矩与原点矩；在频域中应用参数法和非参数法估计信号的功率谱分析其频率特征；利用时频分析技术希尔伯特黄变换(HHT)来提取特征。主要开展的工作如下：将希尔伯特黄变换用于轴承的振动信号分解,计算各信号成分的希尔伯特谱和希尔伯特边际谱,提取相应的故障特征；为减轻经验模态分解出现的混叠现象,提取真实的频率成分,本文分析了信号及其导数的关系,提出一种用信号二阶导数的极值点取代原经验模态分解算法中的信号极值点的方法进行三次样条插值,并将其与改进前进行对比,发现改进的算法能准确分解出信号中幅值分量表现不明显的高频信号,具有很好的实用价值；针对多故障诊断的人工智能算法进行研究,本文中主要用到的算法是多分类支持向量机和决策树。在多分类支持向量机算法中,其方法是：一是从时域和频域中综合提取7个特征,在经过主成分分析降维,得到4个主成分,包含原有特征98.8%的信息。在MATLAB中对多分类支持向量机进行编程,对样本信号的主成分进行训练和预测,试验结果表明：该方法能有效诊断轴承的几种典型故障,总的正确识别率达97.3%。第二种诊断方法是将希尔伯特黄变换和决策树相结合。其方法是：首先用EMD对轴承振动信号进行分解处理,并对得到的固有模态函数进行自相关降噪,提取主要频率作为特征,再根据样本特征生成C4.5故障决策树,并将决策树的节点信息反馈给信号处理算法,以减少在预测工作过程中不必要的属性计算和特征提取,降低算法复杂度,形成一个闭环的诊断系统。实验结果表明该算法正确诊断率达96%。
[Abstract]:Because the life of bearing is discrete and unpredictable, it is very important to make fault diagnosis and real-time monitoring to ensure the safe operation of mechanical equipment. Bearing fault diagnosis includes signal collection, feature extraction, fault pattern recognition and so on. It is an interdisciplinary subject including machinery, signal processing, artificial intelligence, sensor and so on. In this paper, the rolling bearing 6205 is taken as an example to diagnose and analyze its faults. The main fault forms are inner ring fault, outer ring fault and rolling body fault. This paper mainly studies the application of fault feature extraction method, feature dimension reduction method and multi-fault diagnosis algorithm in the diagnosis technology of bearing vibration signal. The software of fault diagnosis is developed by using these algorithms in MATLAB GUI engineering. Feature extraction is an important part of fault diagnosis technology. The selected features should accurately and sensitively reflect the working state of bearings. In this paper, the following methods are used to extract the features: the time-domain statistics of the signals are calculated and extracted, that is, the central moments and the origin moments of each order; In the frequency domain, the power spectrum of the signal is estimated by the parametric method and the non-parametric method, and the feature is extracted by using the time-frequency analysis technique, Hilbert Huang transform (HHT). The main tasks are as follows: The Hilbert-Huang transform is used to decompose the vibration signal of bearing, the Hilbert spectrum and Hilbert marginal spectrum of each signal component are calculated, and the corresponding fault characteristics are extracted. In this paper, the relationship between the signal and its derivative is analyzed, and a cubic spline interpolation method is proposed to replace the signal extremum of the original empirical mode decomposition algorithm with the extremum of the second derivative of the signal. Compared with the improved algorithm, it is found that the improved algorithm can accurately decompose the high frequency signal which the amplitude component in the signal is not obvious, and has good practical value. The main algorithms used in this paper are multi-classification support vector machine and decision tree. In the multi-classification support vector machine algorithm, the first method is to extract seven features from time domain and frequency domain synthetically. Through principal component analysis, four principal components are obtained, which contain 98.8% of the original feature information. The multi-classification support vector machine (SVM) is programmed in MATLAB to train and predict the principal components of the sample signal. The experimental results show that this method can effectively diagnose several typical faults of bearing, and the total correct recognition rate is 97.3%. The second method is to combine Hilbert Huang transform with decision tree. The method is as follows: firstly, the bearing vibration signal is decomposed by EMD, and the natural mode function is automatically correlated to reduce the noise, the main frequency is extracted as the feature, and then the C4.5 fault decision tree is generated according to the sample feature. The node information of the decision tree is fed back to the signal processing algorithm in order to reduce the unnecessary attribute calculation and feature extraction and reduce the complexity of the algorithm to form a closed loop diagnosis system. Experimental results show that the correct diagnosis rate of the algorithm is 96.
【学位授予单位】：华中农业大学
【学位级别】：硕士
【学位授予年份】：2011
【分类号】：TH165.3

【相似文献】