【作者】 周杰；

【导师】 杨宇；

【作者基本信息】 湖南大学 ， 机械工程， 2023， 博士

【摘要】 齿轮的健康状态决定了机械设备的服役质量,当齿轮出现裂纹、点蚀等微弱故障且未能及时发现时,故障会随着生产的进行进一步发展,当故障发展到一定阶段,轻则导致停工维修,重则可能导致灾难性后果。因此开展齿轮的故障诊断研究具有非常重要的工程实际意义。随着科学技术的不断发展,基于信号处理的故障诊断方法日益丰富。研究人员探索了许多采用单通道信号进行机械设备的故障诊断的方法,这对于机械设备故障诊断具有重要意义。但是,当背景噪声严重或齿轮故障较早时,所采集到的振动信号的故障特征往往不够明显,单通道信号中包含的故障信息可能不足以判断齿轮是否出现故障。由于多通道信号中嵌入的故障信息更全面、更丰富,因此多通道信号处理受到了许多研究者的关注。随着传感器技术的发展,采集多通道振动信号已变得简单易行。因此,有必要对所采集的多通道信号构成多元信号进行处理,进而提取更多故障信息,实现更为全面和准确的故障分析。因此,在国家自然科学基金项目(项目编号:51975193、52275103)的资助下,论文以齿轮箱的关键零部件齿轮作为研究对象,根据不同的应用场景提出了:多元信号信息融合与信息增强方法、多元信号降噪方法、各通道功率不平衡情况下的多元信号分解方法、可用于在线监测的多元信号分解方法、模态混叠性能改进的多元信号分解方法、可提取齿轮微弱故障特征的多元信号分解方法。将所提出的方法及其理论在齿轮故障诊断中的应用可行性进行了探讨。论文的主要的研究内容和创新点如下:(1)针对多元信号的信息融合与故障特征增强问题,提出了多元局部特征尺度分解(Multivariate Local Characteristic-scale Decomposition,MLCD)与1.5维经验包络谱(1.5-Dimensional Empirical Envelope Spectrum,1.5D EES)的齿轮故障诊断方法。该方法基于局部特征尺度分解(Local Characteristic-scale Decomposition,LCD)提出了MLCD用于分析多元信号。在仔细分析MLCD模态对齐性质的基础上,提出了一种有效的多通道信息融合方法,经信息融合后故障信号分量中的故障特征明显增强。结合经验包络法和1.5维频谱的优点提出了1.5D EES。1.5D EES可以有效降低包络信号的噪声,进一步增强信号的故障特征。将MLCD-1.5D EES方法应用于齿轮故障仿真与实验信号,分析结果表明,MLCD-1.5D EES能有效融合各个通道的信号,增强齿轮的故障特征,实现了齿轮的故障诊断。(2)针对多元信号的降噪问题,提出了自适应四元数多元局部特征尺度分解(Adaptive Quaternion Multivariate Local Characteristic-scale Decomposition,AQMLCD)。该方法首先定义了一种多元信号降噪方法:四元数奇异值分解故障信息谱(Quaternion Singular Value Decomposition Fault Information Spectrum,QSVDFIS)。QSVDFIS的核心是采用循环调制强度(Cyclic Modulation Intensity,CMI)指标对多元信号中的齿轮故障信息进行评价。QSVDFIS采用CMI指标对四元数奇异值分解的有效秩阶次进行筛选以期达到最佳的降噪效果。QSVDFIS在处理多元信号时要求多元信号各个子信号之间具有关联性,MLCD所获得的多元信号分量具有模态对齐性质,该性质保证了多元信号各个子信号之间的关联性。基于QSVDFIS以及MLCD提出了AQMLCD。AQMLCD能根据多元信号分量自身的特点自适应地降低噪声,实现多元信号的信息融合,并增强信号中故障特征成分。将AQMLCD方法应用于齿轮的试验与仿真信号,结果表明,AQMLCD具有优良的降噪效果,能有效突出信号中的故障特征。(3)现有的自适应多元信号分解方法采用的是Hammersley投影策略。Hammersley投影策略采用的是均匀采样策略,该策略下所产生的投影向量在分解过程中始终不变。因此,采用Hammersley投影策略所获得的信号分量精度不够高。当多元信号出现功率不平衡现象时采用Hammersley均匀投影的自适应多元信号分解方法对多元信号进行分解时所获得的多元信号分量精度不高的现象尤为明显。此外,Hammersley投影策略需要对多元信号进行多次投影才能获得准确的多元信号分量。因此,Hammersley投影策略严重阻碍了自适应多元信号分解方法分解效率的进一步提升,有必要提出新的投影策略以解决上述问题。(1)针对现有自适应多元信号分解方法分解各通道功率不平衡情况的多元信号精度不高的问题,提出了一种新的多元信号分解方法:完全自适应投影多元局部特征尺度分解(Completely Adaptive Projection Multivariate Local Characteristicscale Decomposition,CAPMLCD)。该方法首先提出了完全自适应投影(Completely Adaptive Projection,CAP)策略。CAP投影策略采用基尼系数对多元信号的功率分布进行评价,然后对投影向量重聚类。将CAP投影策略、投影向量迭代更新思想用于改进MLCD方法提出了CAPMLCD。MEMD与MLCD在分解的过程中采用固定的、静态的投影向量进行分解,然而CAPMLCD在分解的过程中采用不断更新的、动态变化的投影向量进行分解。采用齿轮故障仿真与实验信号验证了CAPMLCD在投影次数的参数设置的不敏感性,以及分解精度方面的优越性。CAPMLCD有效的改善了各通道功率不平衡情况下的多元信号分解精度差的问题。(2)针对现有自适应多元信号分解方法分解效率低的问题,提出了一种可用于在线监测的多元信号分解方法:快速多元局部特征尺度分解(Fast Multivariate Local Characteristic-scale Decomposition,FMLCD)。该方法首先基于CAP投影策略提出了快速投影(Fast Projection,FP)策略,与CAP投影策略不同的是FP在CAP的基础上增加了投影向量的筛选机制。将FP采样策略、投影向量迭代更新思想用于改进MLCD方法提出了FMLCD。FMLCD在分解的过程中随着迭代的进行根据多元信号本身的特点对投影向量进行自适应地动态调整、动态筛选。采用齿轮故障仿真与实验信号验证了FMLCD在噪声鲁棒性、分解精度、分解效率等方面的优越性。与其他自适应多元信号分解方法相比FMLCD在分解效率上具有非常明显的优势,因此,FMLCD有望应用于在线故障监测。(4)针对现有的自适应多元信号分解方法模态混叠现象严重的问题提出了多元局部波动模态分解(Multivariate Local Fluctuation Mode Decomposition,MLFMD)。该方法首先提出一种新的局部极值点定位方法,(Second-order Differential Local Extreme Point localization,SDLEPL)。SDLEPL能有效挖掘信号中的局部隐藏信息。此外还提出了一种新的多元信号均值提取方法,多元周期积分均值(Multivariate Periodic Integral Mean Curve,MPIMC),该方法采用了积分平均的思想,积分平均思想能有效表征多元信号的局部波动。基于上述方法,以及CAP投影策略,提出MLFMD。采用齿轮故障仿真与实验信号验证了MLFMD方法的在分解效率、分解精度以及抗模态混叠方面的优越性。(5)针对齿轮微弱故障的情况下故障特征提取困难的问题,从均值曲线优化的角度出发提出了多元内禀波动特征分解(Multivariate Intrinsic Wavecharacteristic Decomposition,MIWD)。该方法首先定义了两种新的单变量均值曲线并提出了六种新的多元均值曲线提取方法。结合MEMD,MLCD以及MLFMD中所定义的多元均值曲线提取方法以及六种新的多元均值曲线提取方法定义了九种潜在多元內禀波动分量。在九种潜在多元內禀波动分量的基础上,结合局部最优思想,以及CAP投影方法,提出了MIWD。不同于MEMD,MLCD采用固定的多元均值曲线提取方法对多元信号进行分解,MIWD在分解的过程中采用局部最优的思想以适合多元信号本身的多元均值曲线提取方法对多元信号进行分解。由于MIWD在分解的各个阶段都采用最优的多元均值曲线对多元信号进行分解,因此与其他自适应多元信号分解方法相比MIWD能更好的提取齿轮的微弱故障特征。采用齿轮故障仿真与实验信号验证了MIWD在抗模态混叠能力,分解精度,分解能力以及正交性等方面的优越性。更多还原

【Abstract】 The health status of gears determines the operational quality of mechanical equipment.When gears develop subtle faults such as cracks and pitting that go unnoticed,the faults can further escalate during producti on.When the faults reach a certain stage,they can result in minor disruptions for repairs or even catastrophic consequences.Therefore,conducting research on gear fault diagnosis holds significant engineering practical significance.With the continuous development of technology,there is an increasing variety of fault diagnosis methods based on signal processing.Researchers have explored many methods for mechanical equipment fault diagnosis using single-channel signals,which is crucial for diagnosing faults in mechanical equipment.However,in cases where there is severe background noise or early-stage gear faults,the fault features captured from vibration signals are often not prominent enough,and the fault information contained in single-channel signals may be insufficient to determine if gears are faulty.Due to the comprehensive and rich fault information embedded in multi-channel signals,multi-channel signal processing has garnered attention from many researchers.With advancements in sensor technology,acquiring multi-channel vibration signals has become simple and feasible.Therefore,it is necessary to process the collected multi-channel signals as a multidimensional signal,extract more fault information,and achieve more comprehensive and accurate fault analysis.Therefore,supported by the National Natural Science Foundation of China(Project No.51975193,52275103),this paper focuses on the key component,gears,in gearboxes as the research object.Based on different application scenarios,the paper proposes the following methods: multi-signal information fusion and enhancement method,multi-signal denoising method,multi-signal decomposition method under various channel power imbalances,multi-signal decomposition method suitable for online monitoring,multi-signal decomposition method for improving mode aliasing performance,and multi-signal decomposition method for extracting subtle fault features in gears.The feasibility of applying the proposed methods and their theories in gear fault d iagnosis is discussed.The main research contents and innovation points of the paper are listed as follows.(1)A method for information fusion and fault feature enhancement of multivariate signals is proposed,which utilizes the multivariate local characteristic-scale decomposition(MLCD)and 1.5-Dimensional empirical envelope spectrum(1.5D EES).MLCD is based on the local characteristic-scale decomposition(LCD)and is used to analyze multivariate signals.An effective multi-channel information fusion method is proposed based on the analysis of the mod e alignment property of MLCD.After the fusion,the fault features in the fault signal components are significantly enhanced.Combining the advantages of the empirical envelope method and the 1.5D spectrum,the 1.5D EES is proposed to effectively reduce the noise of the envelope signal and further enhance the fault features of the signal.The MLCD-1.5D EES method is applied to gear fault simulation and experimental signals,and the results show that the MLCD-1.5D EES can effectively fuse signals from various channels,enhance gear fault features,and achieve gear fault diagnosis.(2)Adaptive quaternion multivariate local characteristic-scale decomposition(AQMLCD)method is proposed for the denoising of multi variate signals.First,a multivariate signal denoising method called quaternion singular value decomposition fault information spectrum(QSVDFIS)is defined.The core of QSVDFIS is the use of the cyclic modulation intensity(CMI)index to evaluate the gea r fault information in multivariate signals.The QSVDFIS selects the effective rank order of quaternion singular value decomposition by the CMI index to achieve the best denoising effect.QSVDFIS requires the correlation between the sub-signals of multivariate signals.The multivariate signal components obtained by MLCD have mod e alignment property,which ensures the correlation between the sub-signals of multivariate signals.Based on QSVDFIS and MLCD,AQMLCD is proposed to adaptively reduce noise and enhance fault feature components in multivariate signal components.The AQMLCD method is applied to experimental and simulated gear signals,and the results show that AQMLCD has excellent denoising effect and can effectively highlight the fault features in the signal.(3)The existing adaptive multivariate signal decomposition methods use the Hammersley projection strategy,which employs a uniform sampling strategy,resulting in projection vectors that remain unchanged throughout the decomposition process.As a result,the accuracy of the signal components obtained using the Hammersley projection strategy is not high enough.This phenomenon is particularly evident when the multivariate signal exhibits power imbalance.The Hammersley projection strategy requires multiple projections of the multivariate signal to obtain accurate signal components,which severely hinders the further improvement of the decomposition efficiency of the adaptive multivariate signal decomposition method.Therefore,it is necessary to propose new projection strategi es to address these issues.(1)To address the problem of low accuracy in decomposing multivariate signals with power imbalances in existing adaptive multivariate signal decomposition methods,a new method called completely adaptive projection multivariate local characteristicscale decomposition(CAPMLCD)is proposed.This method first introduces the completely adaptive projection(CAP)strategy,which evaluates the power distribution of the multivariate signal using the Gini coefficient and then re-clusters the projection vectors.The CAP projection strategy and the iterative update of the projection vectors are used to improve the MLCD method and propose the CAPMLCD.Unlike MEMD and MLCD,which use fixed and static projection vectors for decomposition,CAPMLCD uses continuously updated and dynamically changing projection vectors for decomposition.The effectiveness of CAPMLCD in improving the accuracy of multivariate signal decomposition in cases of power imbalances across channels is demonstrated using simulated and experimental signals of gear faults.(2)To address the problem of low efficiency in existing adaptive multivariate signal decomposition methods,a new method called fast multivariate local characteristicscale decomposition(FMLCD)is proposed,which can be used for online mon itoring.This method first introduces the fast projection(FP)strategy based on the CAP projection strategy,but with an added screening mechanism for the projection vectors.The FP sampling strategy and iterative update of the projection vectors are used to improve the MLCD method and propose the FMLCD.FMLCD adaptively and dynamically adjusts and screens the projection vectors based on the characteristics of the multivariate signal during the decomposition process as the iterations progress.The superiority of FMLCD in terms of noise robustness,decomposition accuracy,and efficiency is demonstrated using simulated and experimental signals of gear faults.Compared to other adaptive multivariate signal decomposition methods,FMLCD has a significant advantage in decomposition efficiency and is therefore expected to be applied to online fault monitoring.(4)Multivariate local fluctuation mode decomposition(MLFMD)is proposed to address the problem of serious mode mixing in existing adaptive multivariate sig nal decomposition methods.This method first introduces a new method for locating local extreme points,called second-order differential local extreme point localization(SDLEPL),which can effectively explore local hidden information in the signal.In addition,a new multivariate signal mean extraction method,multivariate periodic integral mean curve(MPIMC),is proposed,which adopts the idea of integral averaging to effectively characterize the local fluctuations of multivariate signals.Based on these methods,as well as the CAP projection strategy,MLFMD is proposed.The efficiency,accuracy,and anti-mode mixing ability of the MLFMD method are verified through gear fault simulation and experimental signals.(5)To address the difficulty in extracting fault features from weak gear faults,a novel approach called multivariate intrinsic wave-characteristic decomposition(MIWD)is proposed from the perspective of mean curve optimization.This method defines two new types of univariate mean curves and introduces six new methods for extracting multivariate mean curves.By combining the multivariate mean curve extraction methods defined in MEMD,MLCD,and MLFMD,as well as the six newly proposed methods,nine potential multivariate intrinsic wave componen ts are defined.Based on these nine components,MIWD is developed by incorporating the local optimality concept and CAP projection method.Unlike MEMD and MLCD,which use fixed multivariate mean curve extraction methods for decomposition,MIWD adapts the extraction method of multivariate mean curves to the characteristics of the multivariate signals by employing the local optimality concept during the decomposition process.Due to the optimal multivariate mean curve extraction at each stage of decomposition,MIWD outperforms other adaptive multivariate signal decomposition methods in extracting weak fault features from gears.The superiority of MIWD is demonstrated through simulations and experimental signals in terms of resistance to mode mixing,decomposit ion accuracy,decomposition capability,and orthogonality.更多还原