基于变分模态分解与优化多核支持向量机的旋转机械早期故障诊断方法研究

发布时间：2018-01-04 15:28

本文关键词：基于变分模态分解与优化多核支持向量机的旋转机械早期故障诊断方法研究　出处：《重庆大学》2016年博士论文　论文类型：学位论文

【摘要】：旋转机械作为现代化工业生产中必不可少的工程设备,被广泛应用于化工、石油、冶金、电力等关系国计民生的重要领域。一旦旋转机械设备发生故障轻则致使整个系统瘫痪造成巨大经济损失,重则产生人员伤亡。研究表明旋转机械早期故障阶段具有较长的潜伏期,如果能在旋转机械设备故障萌芽即将出现、刚刚出现或故障程度比较轻微时,准确地预测故障可能发生的时间、部位以及故障类别,并据此指导旋转机械设备的保养和维修工作,将有利于对旋转机械设备故障发展进行有效控制,并保障其安全可靠地运行。旋转机械早期振动信号易受强背景噪声干扰,同时受到传输路径与信号衰减、传播介质与采集设备损耗等因素影响,进一步弱化了振动信号中包含的故障振动信号信息;旋转机械大多结构复杂,故障源信号至测点位置为非线性传播,同时运行工况不稳定、多部件耦合振动、振动干扰大等特点,使得采集到的故障振动信号具有强烈的非平稳、非线性等特点,导致故障振动信号与设备状态映射关系模糊,故障特征难以提取;旋转机械早期故障样本缺乏长期性、系统性的收集,故障样本稀缺且故障特征值和故障的类别无明确的映射关系,故障辨识难度大。论文针对旋转机械微弱故障信号增强、非线性及非平稳故障信号特征提取、小子样故障诊断等问题,深入研究基于变分模态分解及优化多核支持向量机的旋转机械早期故障诊断方法,具体研究内容如下:(1)针对旋转机械早期背景噪声干扰大故障信息微弱的问题,提出自适应最大相关峭度反卷积的微弱故障增强方法。以相关峭度作为评价指标,充分考虑早期故障振动信号中所含冲击成分的特性,通过迭代过程以实现解卷积运算;利用小波Shannon熵作为目标函数,采用变步长网格搜索法自动搜寻最优滤波器阶数以及周期;使最大相关峭度反卷积方法在达到很好的效果的同时更具有自适应性,从而有效地检测出被噪声淹没的微弱故障;(2)针对旋转机械故障信号非平稳、非线性特征提取难的问题,提出自适应变分模态分解的多频带多尺度样本熵特征集构建方法。利用不同频带上模态的多尺度样本熵组成敏感特征向量集,表征旋转机械早期故障状态特征,进而提高对旋转机械早期故障状态的辨识能力;(3)针对变分模态分解中一些关键参数选择不确定的问题,提出自适应变分模态分解方法;提出以VMD分解后各模态与原信号之间的相关性来保证分解的精度并指导最优K值的确定;VMD分解中平衡约束参数越小,所得模态分量带宽越大,易出现中心频率重叠以及模态混叠的现象;通过仿真实验分析,提出在实际应用中一般可取平衡约束参数为采样频率;研究分析了自适应变分模态分解的性能:正交性能分析、能量保存度分析、等效滤波属性分析;通过仿真实验分析得出:自适应变分模态分解在正交性能、能量保存性能方面,均优于EMD、EEMD、LMD方法;利用分数高斯噪声通过数值模拟实验对EMD、LMD、AVMD等效滤波属性分析,相比于EMD和LMD,AVMD可以更接近于小波包分解,且是一种比EMD和LMD能提供更高的时频分辨率的自适应分解方法;对比研究了AVMD与EMD性能,发现当有异常信息干扰时,AVMD仍具有很好的效果;(4)针对旋转机械早期故障样本缺乏,提出基于免疫遗传算法优化多核支持向量机的旋转机械小子样故障诊断方法。通过引入权重因子将不同核函数组合学习,基于全局核函数以及局部核函数构造多核函数,实现输入特征向量到核函数空间的快速映射,算法泛化能力更好、模型解释能力更强;利用免疫遗传算法获取多核支持向量机最优参数,克服多核支持向量机参数选择的不确定性,进而提高多核支持向量机在旋转机械小子样振动故障诊断中的稳定性以及泛化推广能力。文章最后对本文的工作进行总结,并展望下一步的研究方向。
[Abstract]:As the rotating machinery engineering equipment essential for modern industrial production, is widely used in chemical industry, petroleum, metallurgy, power and other important areas. Once the relationship beneficial to the people's livelihood rotating machinery fault light can lead to paralysis of the entire system caused huge economic losses and heavy casualties produced. The research showed that the early stage of failure of rotating machinery has a long incubation period if you can, in rotating machinery fault sprout will appear, or just appear relatively minor fault degree, accurate prediction of possible fault time, fault location and category, and then guide the rotating machinery equipment maintenance and repair work, will be conducive to the effective control of the equipment fault of rotating machinery development, and guarantee their safety reliable operation. The early vibration signal of rotating machinery is easily affected by strong background noise, at the same time by the transmission path and signal Effect of attenuation, propagation medium and acquisition equipment loss and other factors, further weakening the fault vibration signal information contained in the vibration signal of rotating machinery; most complicated structure, fault source signal to the measuring point for nonlinear propagation, and unstable running, multi components coupling vibration, vibration characteristics of high noise, the fault vibration signal the collected with strong non-stationary, non-linear characteristics, cause the fault vibration signal and device state mapping relation to fuzzy, fault feature extraction of rotating machinery early fault samples; lack of long-term, systematic collection, no explicit mappings between fault samples are scarce and fault characteristic value and fault type, fault identification difficult for weak fault signal of rotating machinery to enhance the extraction of nonlinear and nonstationary fault signal characteristics, fault diagnosis of small sample problems, in-depth research base In the early fault diagnosis of rotating machinery based on variational modal decomposition and optimization of multiple kernel support vector machine, the specific contents are as follows: (1) for rotating machinery early background noise interference problems of weak fault information, the fault adaptive maximum correlation kurtosis deconvolution enhancement method. Related to the kurtosis as the evaluation index, fully considering the characteristics of early fault vibration signal contained in the impact component, through the iterative process to achieve the deconvolution method; using wavelet Shannon entropy as the objective function, the automatic search for the optimal filter order and cycle using variable step grid search method; the maximum correlation kurtosis deconvolution method is more adaptive to achieve good results at the same time thus, effectively detect weak fault submerged by the noise; (2) according to the fault signal of rotating machinery non-stationary, nonlinear feature extraction problem,. The adaptive variational modal decomposition of multi band multiscale sample entropy feature construction method. Sensitive feature vector set using multi-scale sample entropy of different frequency modes, characterization of rotating machinery early fault characteristics, and improve the ability to identify early fault of rotating machinery condition; (3) according to the variation in mode decomposition the problem of determining the choice of some key parameters, an adaptive variational modal decomposition method; the correlation between VMD after the decomposition of the original signal mode and to ensure the accuracy of determining decomposition and guide the optimal K value; VMD decomposition equilibrium constraint parameter is smaller, the modal components more bandwidth, center frequency and prone to overlap the modal aliasing phenomenon; through simulation analysis, put forward in the practical application of general equilibrium constraint parameters for the sampling frequency; analysis of the adaptive variational modal decomposition Performance: the orthogonal performance analysis, energy saving analysis, analysis of equivalent filter attributes; through the simulation analysis results show that the adaptive variational modal decomposition in orthogonal performance, energy saving performance, are better than those of EMD, EEMD, LMD; through numerical simulation experiments on EMD, LMD using fractional Gauss noise analysis, AVMD equivalent filter properties compared to EMD, AVMD and LMD, can be more close to the wavelet packet decomposition, which is an adaptive than EMD and LMD can provide higher time-frequency resolution decomposition method; comparison of AVMD and EMD performance, found that when there are abnormal interference information, AVMD still has a very good effect; (4) according to the early fault of rotating machinery like lack of samples, proposed methods for fault diagnosis of rotating machines. The optimization of multi kernel support vector machine based on immune genetic algorithm. By introducing a weighting factor of different kernel function combination learning, global kernel function based on And the local kernel structure of multi kernel function, fast mapping input vectors to the kernel space, better generalization capability, the ability to explain the model by using immune genetic algorithm gets stronger; multi kernel support vector machine optimized parameters, parameter selection of support vector machines to overcome the uncertainty of multi kernel support, and improve the stability of the multi kernel support vector machine in the vibration fault diagnosis of rotating machinery in small sample and generalization ability. Finally, a summary of this paper, and the prospect of the next research direction.

【学位授予单位】：重庆大学
【学位级别】：博士
【学位授予年份】：2016
【分类号】：TH17

【参考文献】