【作者】 张营；

【导师】 左洪福；

【作者基本信息】 南京航空航天大学 ， 载运工具运用工程， 2013， 博士

【摘要】 滚动轴承作为旋转机械关键部件，其性能退化或失效影响整机性能甚至导致设备非计划停机，造成经济损失甚至人员伤亡。由于滚动轴承寿命离散较大，传统的定时维修往往造成“维修不足”或者“维修过剩”，因此研究滚动轴承状态监测技术和寿命预测方法，能够变定时维修为视情维修、预测维修，是实现故障预防、保障设备安全服役的关键。振动、温度等常规监测技术能够监测到滚动轴承较严重故障，如裂纹、表面剥落等，从发现故障征兆到最后失效间隔时间短，难以有充分时间制定维修决策。本文针对常规监测技术对于早期故障发现能力的不足，设计研制了磨损区域静电传感器，探讨静电信号的主要影响因素；深入分析静电信号噪声干扰类型，提出多方法联合的去噪方法，有效提高静电监测技术早期故障识别能力；从时域、频域以及复杂度度量角度提取静电信号多维特征，进而提出多特征参数融合的轴承性能退化评估方法以及两阶段剩余寿命预测框架。研究成果对于提高机械设备关键部件的状态监测和寿命预测能力具有重要的参考价值和指导意义。论文主要研究内容如下：（1）系统总结磨损区域荷电机理，分析静电测量原理并设计研制磨损区域静电传感器，探讨了静电感应信号的主要影响因素。开展了点接触滑动摩擦胶合实验，验证静电传感器早期故障检测能力，分析载荷转速对其影响，进行了滚动摩擦故障注入实验，研究静电传感器对于不同尺寸故障的反应能力以及受载荷转速的影响，设计改装了滚动轴承疲劳加速寿命监测试验平台，比较分析静电监测与振动监测的故障反应能力。（2）由于静电感应信号混有不同类型强噪声干扰，很难采取单一方法有效抑制噪声。通过仿真分析比较了陷波器法、独立分量法以及谱插值法三种工频干扰去除方法的效果，得出谱插值法计算简单并且滤除工频及其谐波分量的同时保留了附近有效成分。通过仿真分析比较了小波去噪、EMD去噪以及奇异值差分谱三种方法用于背景噪声和随机脉冲的去除效果，得出奇异值差分谱方法信噪比最高，效果最佳。在此基础上结合静电信号特点提出一种基于谱插值法和奇异值差分谱法的联合去噪方法，通过仿真和轴承寿命实验验证了所提方法的有效性和必要性。（3）针对静电监测常规时域频域指标对滚动轴承性能退化的反映能力不足，以及相关研究的匮乏，借鉴振动监测特征参数的研究方法，进一步探索新的适用于静电监测的指标。静电信号中随机成分随着滚动轴承性能退化而不断变化，引入复杂度度量方法描述这一变化过程，分析多种复杂度度量指标对滚动轴承性能退化的反映能力，发现复杂度指标对于轴承性能早期退化更加敏感，能够有效补充现有常规静电特征参数的不足。（4）为解决单特征参数对轴承性能退化反应不灵敏或不一致的问题，提出基于谱回归-高斯混合模型（spectrum regression-gaussian mixture models，SR-GMM）的多参数融合性能评估模型。利用谱回归后正常状态数据建立GMM的基准模型，以基于贝叶斯推断的距离（bayesianinference distance，BID）表征测试数据与GMM模型的全局距离，作为反映轴承性能退化程度的定量指标。与其它两种基于GMM的监控指标以及基于支持向量描述的性能评估方法相比，所提方法能够更早地发现退化的发生。在均采用本文所提融合评估方法前提下，静电监测评估结果相比振动监测更早发现异常，静电监测能够为振动等常规监测手段提供有益补充。（5）结合滚动轴承运行的两阶段性特点，提出了多特征参数融合的随机滤波剩余寿命预测框架，在正常运行状态时以监测为主，当轴承进入缺陷运行阶段时进行寿命预测。以多特征参数融合结果BID为输入，以随机滤波模型建立BID与剩余寿命的关系，该方法解决了两个主要问题，一是失效阈值难以给定或没有标准可以遵循的问题，二是随机滤波模型需要输入数据为单维，高维数据输入时需计算其联合概率分布，增加计算成本影响在线寿命预测速度的问题。将所提方法与基于支持向量机的状态寿命评估方法，基于时间序列的寿命预测方法以隐半马尔科夫寿命预测方法进行了比较，验证本文所提方法的优越性。更多还原

【Abstract】 Rolling bearing is one of the key components in the rotating machinery, once the failure of therolling bearing occured will affect the performance of the whole machine, or cause unplanned outage,which will make for economic losses or even the heavy casualties. As the greate discreteness of thelife of the rolling bearing, traditional maintenance often causes “inadequate maintenance” or “Excessmaintenance”. Therefore it is great significance to study the condition monitoring and life predictionmethods of the rolling bearing, which are the crucial technicals of the fault prevention and securityservice of the devices. In view of the traditional monitoring techniques for the inadequacy of ability ofearly failure found, the wear site electrostatic sensor and the corresponding life prediction frameworkis designed and proposed in this paper. The research results have important reference values andguiding significances for the improvement of condition monitoring and life prediction ability ofmechanical equipment.The main contents of this paper are as follows:(1) The wear site charging mechanism is revaled in this paper, and we design the electrostaticsensor based on the electrostatic induction principle. Considering the rolling bearing workingpeculiarity as a mixture of rolling and sliding, sliding bearing steel point contacts for early detectionof scuffing experiments, rolling friction fault injection experiments and rolling bearing lifeexperiments are carried out, which are used to verify the electrostatic sensor sensitivity for earlyfailure and different degree of faults. The results show that the electrostatic sensor could detectabnormal signals before complete failure and provide more effective information of the rolling beaingdegradation.(2) As electrostatic induction signals are mixed with strong different types of noises, it is difficultto adopt single method to suppress noises effectively. Three power frequency interference removemethod, digital notch filter, independent component analysis and spectrum interpolation areintroduced and compared by simulation, and then three method, wavelet denoising, EMD denosingand difference spectrum of singular values, are introduced to remove the background noises and therandom pulses. On this basis, a noval joint denoising method based on the spectrum interpolation anddifference spectrum of singular values is proposed to denoising of the electrostatic signals. Thesimulation and experiments show that the proposed method is effective and necessary.(3) For the traditional time and frequency domain index is insufficient to reflect the performance degradation of the rolling bearing, new indicators are explored for electrostatic monitoring. The natureChange of the electrostatic signals of rolling bearing life cycle is the random component changingduring the degradation process. Complexity just right can reflect this change so several complexitymeasures are proposed in this paper. The results show that complexity measures can reflect initialdegradation more sensitive than general indices.(4) In view of single feature is insensitive or inconsistency on the performance degradation of thebearing, a multi-parameter fusion method based on spectrum regression-gaussian mixture models(SR-GMM) is proposed in this paper. Several dimension reduction algorithms are compared and theSR mothod will be better to discover the data structure and has faster computing speed, and than thedata set of normal state is used to built the baseline GMM, a new index, bayesian inference distance(BID), is used to characterize the global distance of the test data to the GMM model, which is usedas a quantitative indicator to reflect the bearing performance degradation degree. This indicator usesbayesian inference combined priori and posteriori probability, which could earlier response to datachanges. Comparing to the other index, bayesian inference probability (BIP) and negative loglikelihood probability (NLLP), BID is advantageous. And the GMM-BID method can finddegradation occurred much earlier than the support vector domain description (SVDD) method. Tofurther illustrate the advantages of electrostatic monitoring, it is compared with the vibrationmonitoring, it is found that the electrostatic monitoring is more sensitive to the early fault and canprepared enough time for the decision-making.(5) Stochastic filtering life prediction model is brought into the field of electrostaticmonitoring.In view of data input of stochastic filtering model need to be one-dimensional, when theinput is high-dimensional, it need to calculating the joint probability distribution, which willinfluence the speed of calculation, a multi-feature fusion life prediction framework based onstochastic filtering is proposed, which the BID results are used as the input data, and establish therelationship between BID and residual life by stochastic filtering. The accuracy of the predictionresults is verified by the experiments. Meanwhile, the proposed method is compared with the statelife assessment method based on support vector machine (SVM), life prediction based on time seriesand hidden semi-Markov model (HSMM), the results show the superiority of the proposed method.更多还原