旋转机械转子—滚动轴承复合故障诊断方法研究

发布时间：2019-03-27 10:51

【摘要】：旋转机械转子-滚动轴承系统中,复合故障发生的概率大于单一故障,且发生复合故障时,准确识别缺陷类型,以便提早安排维修计划,防止意外停机是关键,所以研究复合故障诊断更有意义。复合故障诊断方法,除了能够诊断出复合故障的情况外,还应向下兼容,具备诊断单一故障的能力。本文通过试验探索了转子不平衡、不对中和滚动轴承6种复合故障的振动特征,研究发现:复合故障振动信号含有各单一故障的特征信息,同时也存在耦合的特征信息,而且复合故障对单一故障特征有很大影响。在轴承座上拾取的振动信号十分复杂,但总体可概括为滚动故障信号、确定性信号和噪声,微弱故障信号可能被其他信号掩盖,难以提取故障特征。针对这一问题文中提出了基于LMD-MED的故障特征提取方法,LMD作为MED的预处理方法,MED处理前4个PF分量PF分量,来消除噪声和增强微弱冲击信号。文中通过SKF6205滚动轴承滚动体早期缺陷试验数据,证明该方法有效提取出了故障特征,并且峰值占比和信噪比较传统方法分别增加了150%和18.3%。针对LMD-MED方法无法直接确定MED步长和PF分量的问题,提出一种基于峭度-步长及信噪比-分量选取的改进LMD-MED方法。首先,搜索LMD-MED处理后的峭度拐点,以此确定MED步长;然后,在Hilbert包络信号中分别计算滚动轴承故障频率处的信噪比,确定最大信噪比对应的PF分量和故障类型;最后,对选定分量作Hilbert包络分析。文中通过试验,说明了该方法在单一滚动轴承故障诊断及复合故障中滚动轴承故障诊断方面都有良好效果。利用这些试验数据,将EMD-MED方法和LMD-MED方法进行对比研究,结果发现:LMD的分量能量较EMD集中,模式混叠较轻,同时LMD将部分高频噪声分解到了目标分量中;而EMD却可以分解出目标分量中的大部分噪声,信噪比较高。为构建复合故障特征集,分别在信号的低频段和滚动轴承一阶共振区频段提取时频域特征。文中利用小波包分解对信号进行分解和重构,重构低频段信号,提取2个频域特征和8个时域特征;重构滚动轴承一阶共振区信号,提取EMD处理后IMF1分量的6个频域特征和8个时域特征;将这些特征组成复合故障特征集。在训练集中,采用kPCA提取该特征集的核主元,利用PSO方法优化SVM惩罚参数和RBF核函数参数,训练SVM模型。文中通过转子不平衡、不对中和N205滚动轴承12种故障试验证实了该方法的有效性,训练集3折交叉验证分类精度和测试集分类精度分别达到99.44%和99.58%,故障识别精度较高。
[Abstract]:In the rotary mechanical rotor-rolling bearing system, the probability of the composite fault occurrence is greater than a single fault, and when the composite fault occurs, the type of the defect is accurately identified, so that the maintenance plan can be arranged early so as to prevent the accidental shutdown from being critical, so the research of the composite fault diagnosis is more meaningful. The composite fault diagnosis method is compatible with the ability to diagnose a single fault, in addition to the ability to diagnose the complex fault. In this paper, the vibration characteristics of the six kinds of composite faults of the rotor unbalance, the non-centering and the rolling bearing are explored. The results show that the composite fault vibration signal contains the characteristic information of each single fault, and the characteristic information of the coupling is also present. And the composite fault has a great influence on the single fault characteristic. The vibration signal picked up on the bearing seat is very complex, but can be generally summarized as a rolling fault signal, a deterministic signal and a noise, and the weak fault signal may be masked by other signals, making it difficult to extract the fault features. In this paper, a fault feature extraction method based on LMD-MED is proposed. LMD is used as the pre-processing method of MED. The PF component of four PF components before MED process is used to eliminate the noise and enhance the weak impact signal. In this paper, the early-defect test data of the rolling element of the rolling bearing of the SKF6205 rolling bearing proves that the method can effectively extract the fault characteristic, and the peak-to-noise ratio and the signal-to-noise ratio are increased by 150% and 18.3%, respectively. An improved LMD-MED method is proposed for LMD-MED method which can not directly determine the MED step size and the PF component, and proposes an improved LMD-MED method based on the kurtosis-step size and the signal-to-noise ratio-component selection. Firstly, searching the inflection point of the rolling bearing after the LMD-MED processing to determine the MED step size; then, respectively calculating the signal-to-noise ratio at the fault frequency of the rolling bearing in a Hilbert envelope signal, and determining the PF component and the fault type corresponding to the maximum signal-to-noise ratio; and finally, performing a Hilbert envelope analysis on the selected component. The test shows that the method has good effect in the fault diagnosis of single rolling bearing and the fault diagnosis of rolling bearing. Using these test data, the EMD-MED method and the LMD-MED method are compared and studied. The result shows that the component energy of the LMD is concentrated in the EMD, the mode aliasing is light, and the LMD decomposes part of the high-frequency noise into the target component, and the EMD can decompose most of the noise in the target component, The signal-to-noise ratio is high. In ord to construct a composite fault feature set, that frequency domain characteristic is respectively extracted at the low frequency section of the signal and the first order resonance region frequency band of the rolling bearing. In this paper, the wavelet packet decomposition is used to decompose and reconstruct the signal, the low-frequency segment signal is reconstructed, the two frequency-domain characteristics and the 8 time-domain characteristics are extracted, the first-order resonance region signal of the rolling bearing is reconstructed, and the six frequency-domain characteristics and the 8 time-domain characteristics of the IMF1 component after the EMD processing are extracted; These features are combined into a composite fault feature set. In the training set, the kernel principal component of the feature set is extracted by the kPCA, and the SVM model is trained by using the PSO method to optimize the SVM penalty parameter and the RBF kernel function parameter. In this paper, the validity of this method is confirmed by the unbalance of the rotor and the 12 fault tests of the neutral and N205 rolling bearings. The accuracy of the three-fold cross-validation classification and the classification accuracy of the test set are 99.44% and 99.58%, respectively, and the fault recognition accuracy is high.
【学位授予单位】：中国矿业大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TH133.3

【参考文献】