基于高阶统计量的齿轮传动系统故障特征提取方法研究

发布时间：2018-04-29 13:05

本文选题：齿轮传动系统 + 故障特征提取　；参考：《华北电力大学》2013年博士论文

【摘要】：本文以齿轮传动系统故障振动信号为研究对象,在分析振动机理、故障模式基础上,从高斯性、非高斯性角度划分振动信号的各种成分,然后利用振动信号的双谱幅值提取到非高斯性强度、双谱熵这两类故障特征,最后讨论了新高阶统计量特征的优化与实用问题,主要研究成果如下： (1)通过研究齿轮传动系统主要零部件的失效模式、振动机理、故障表现,将振动信号划分为与故障部件有关的非高斯确定性成分、与故障部件无关的非高斯确定性成分、高斯随机成分、对称非高斯随机成分、非对称非高斯随机成分。当齿轮或轴承发生故障时,其振动信号中的确定性成分、随机成分会产生明显变化,于是信号的非高斯性也随之改变,尤其是因故障而增强的边带成分,它会导致信号的非高斯性、二次非线性增强,信号的双谱分析结果对此十分敏感。另外,双谱能抑制掉振动信号中的高斯随机成分、对称非高斯随机成分,保留非对称非高斯随机成分信息,这有利于降低噪声、非故障随机振动的干扰。 (2)当齿轮传动系统发生故障时,振动信号的非高斯成分在双频域内的分布强度与形态均随之改变。基于双谱幅值信息,分别提取到量化描述振动信号非高斯成分强弱变化的非高斯性强度特征值,以及量化描述振动信号非高斯成分在双频域内分布形态变化的双谱熵特征值。按照双频域内的不同定义区间,共提取到六种特征值,分别是基于主定义域区间的NGIPD、HB-PD,基于任意二维区间part的NGIpart、HB-part,和基于多分区的NGIregion(k)、HB-region(k)。这些新的高阶统计量特征既保留了双谱分析的优点,而且弥补了双谱幅值切片等常规高阶统计量特征值的不足。不同定义区间的非高斯性强度、双谱熵特征值各有特点：NGIPD与HB-PD分别包含了双谱幅值的全部强度、分布形态信息,不含双谱对称冗余信息；当二维定义域区间part内包含有较丰富故障信息时,NGIpart与HB-part表征故障的能力较好,但需要人为观察并设定合适的双频域二维区间；振动信号的故障信息在双频域内分布不均,特征值NGIregion(k)、HB-region(k)在某些分区内故障特征提取效果较好,某些分区效果较差,这正体现了双频域分区特征值对双谱内容、故障信息具有较强的“聚焦能力"。 (3)针对非高斯性强度、双谱熵特征值存在的不足,进行了特征压缩与信号滤波这两类故障特征优化。双谱分区特征值会产生高维特征空间,给实际应用造成麻烦,利用主分量分析、核函数主分量分析从线性及非线性角度进行特征压缩,可将绝大部分故障信息浓缩于低维主分量空间或低维核主分量空间,尤其在利用Fisher准则进行特征优选预处理后,新的主分量、核主分量特征值具有较高的故障区分能力。由于齿轮啮合频率及其谐波成分通常能量较强,且易受非故障因素影响,于是采用Gabor滤波与信号重构,在时频域空间精准滤除振动信号的啮合频率及其谐波成分,结果表明,重构信号的非高斯性强度特征对齿轮故障敏感程度提高,有利于后续故障诊断。另外,为研究新特征值在齿轮传动系统状态监测与故障报警中的使用方法与效果,以原始振动信号的HB-PD、Gabor滤波重构信号的NGIPD为例进行故障特征趋势分析,并根据“3σ准则”设定故障阈值,不论报警时间还是诊断准确率,都取得理想效果,为工程实际应用提供了新思路与新方法。
[Abstract]:This paper takes the fault vibration signal of the gear transmission system as the research object. On the basis of the analysis of the vibration mechanism and the fault mode, the various components of the vibration signal are divided from the Gauss nature and the non Gauss angle. Then the two types of fault features of the non Gauss strength and the bispectrum entropy are extracted by the bispectral amplitude of the vibration signal. Finally, the new high order statistics are discussed. The main research results are as follows:
(1) by studying the failure mode, vibration mechanism and fault performance of the main parts of the gear transmission system, the vibration signals are divided into non Gauss deterministic components related to the fault parts, non Gauss deterministic components independent of the fault parts, Gauss random components, symmetric non Gauss random components and asymmetric non Gauss random components. When or when a bearing fails, the deterministic component in the vibration signal will change obviously, so the non Gauss character of the signal is also changed, especially the edge band component enhanced by the fault. It will lead to the non Gauss character of the signal, the two nonlinear enhancement, and the bispectrum analysis of the signal is very sensitive. In addition, bispectrum The Gauss random component in the vibration signal can be suppressed, the asymmetric non Gauss random component is symmetric, and the asymmetric non Gauss random component information is retained, which helps to reduce the noise and the interference of the non fault random vibration.
(2) when the gear transmission system fails, the intensity and shape of the non Gauss component of the vibration signal are changed in the dual frequency domain. Based on the bispectrum amplitude information, the non Gauss intensity characteristic values are extracted to quantify the non Gauss intensity changes of the vibration signals, and the quantized description of the non Gauss components of the vibration signals in two pairs is quantized. The bipectral entropy eigenvalues of the morphological changes in the frequency domain. According to the different definition intervals within the dual frequency domain, six eigenvalues are extracted, which are based on the NGIPD, HB-PD, NGIpart, HB-part, and HB-region (k) based on arbitrary two-dimensional interval part, respectively. These new higher-order statistics features are guaranteed. The advantages of bispectral analysis are retained, and the inadequacy of the eigenvalues of conventional high-order statistics, such as bispectral amplitude slices, is made up. The characteristics of the non Gauss entropy of different defined intervals have their own characteristics: NGIPD and HB-PD contain all the intensity of the bispectral amplitude, the distribution form of information, and no bispectral redundancy information; When the packet part contains a lot of fault information, the ability of NGIpart and HB-part to characterize the fault is better, but it is necessary to observe and set suitable two frequency domain. The fault information of the vibration signal is distributed unevenly in the dual frequency domain, the eigenvalue NGIregion (k), and HB-region (k) is better in extracting the fault features in some partitions. The effect of some partitions is poor, which reflects the dual frequency domain partition eigenvalue has strong "focusing ability" for bispectrum content and fault information.
(3) in view of the inadequacy of the non Gauss intensity and the existence of the eigenvalues of the bispectrum entropy, the feature compression and signal filtering are optimized for the two types of fault features. The eigenvalues of the bispectrum partition can produce high dimensional feature space and cause trouble to the practical application. By using the principal component analysis, the kernel function principal component analysis is compressed from the linear and nonlinear angles. Most of the fault information is concentrated in the low dimensional principal component space or the low dimensional kernel principal component space. The new principal component and the kernel principal component eigenvalue have high fault differentiation ability, especially when the Fisher criterion is used to perform the feature optimization preprocessing. As a result, Gabor filtering and signal reconstruction are used to accurately filter the meshing frequency and harmonic components of the vibration signal in the time and frequency space. The results show that the non Gauss strength characteristics of the reconstructed signal can improve the sensitivity of the gear fault and be beneficial to the follow-up fault diagnosis. In addition, the new eigenvalue is studied for the state monitoring of the gear transmission system. With the method and effect of using the fault alarm, the fault feature trend is analyzed with the NGIPD of the HB-PD and Gabor filtering reconstruction signal of the original vibration signal, and the fault threshold is set according to the "3 Sigma" criterion. No matter the alarm time or the diagnostic accuracy, the ideal effect is obtained. The new idea and new method are provided for the practical application of the project.

【学位授予单位】：华北电力大学
【学位级别】：博士
【学位授予年份】：2013
【分类号】：TH132.41;TH165.3

【相似文献】