【作者】 张袁元；

【导师】 李舜酩；

【作者基本信息】 南京航空航天大学 ， 车辆工程， 2013， 博士

【摘要】 车辆的振动与噪声是衡量其使用舒适性的重要方面，同时也是运行机构状态检测的重要指标。试验测试一般在车辆行驶或运动过程中进行，传感器受到不同激励源的影响，测试信号并不能完全反映所对应位置的源信息，它受到测试系统或环境噪声的影响和不同源对该位置干涉影响，从而造成测量信号之间彼此的相关性。基于直接测量得到的信号进行响应贡献度的分析，会夸大或缩小真实源的贡献。针对多相关激励背景下的振动故障诊断、源和路径识别问题，本文研究的主要内容归纳如下：(1)探讨了经验模式分解方法(EMD)，改善了其虚假分量和模式混叠的分解缺陷，以其作为车辆测试信号的前期去噪和特征提取手段。将相干分析应用于判定分解中的虚假分量，以避免虚假分量和微弱信号的误判；其次研究了信号各频率能量比例对EMD分解时虚假分量的影响，提出了抑制低频分量或加入高频分量幅值以减小虚假分量引入的分解误差；基于白噪声的EMD分解特性，提出了有色噪声频段累加方法以改善EMD分解的模式混叠(MH-EMD)，同时仿真分析了其优越性，最后将改进的EMD系统方法应用于转子模拟试验台声学信号和排气系统振动信号的特征提取，验证了新方法的可行性。(2)联合集合经验模式分解(MH-EMD)和最小均方算法(LMS)发展了一种自适应特征提取方法。以自适应均方算法为对象，研究了其固定步长、固定阶数、变步长(VS)和变阶数(VT)的算法性能，提出了在迭代过程中以比较阶数和步长变化时的最小均方误差期望为收敛方向，发展了一种联合变步长变阶数最小均方算法(VSVT-LMS)的去噪方法；结合对原信号的MH-EMD分解，使各模式分量窄带化，进而通过VSVT-LMS对每个IMF分量进行去噪，避免LMS算法对宽频信号的不稳定性，同时也避免了EMD分解的不唯一性和去噪中阈值的选择问题。最后通过对仿真和实际车辆振动信号去噪，验证了方法在工程上的可行性。(3)针对车辆振动噪声具有相关特征的测试信号，发展和完善了多种多相关分离方法。对于瞬态混合模型，探讨了一般盲源分离方法，提出了基于幅值密度比较的盲信号分离方法，在针对频率信号量较少时，可以识别单频和共频信号；改进了联合近似对角化方法，提出了均衡化的分段联合近似对角化和最优距离矩阵的联合近似对角化方法，改进方法精度高于经典方法；基于高阶累积量，取其高阶切片，结合近似对角化和联合分段对角矩阵，提出可以应用于传感器数量少于源数的分离方法，该方法较经典方法具有更好的稳定性和更高的分离精度；针对具有同频信号或相关性较大的源信号，改善小波包子带的盲分离，可以解决具有统计相关源混合观测信号的分离，以简支梁振动试验验证了方法的有效性。针对具有卷积特性的混合模型，比较了时域和频域的卷积盲分离算法，提出了一种基于互信息向量独立判断的频域算法，在算法迭代过程中，非线性函数考虑全频段的分离矩阵，避免了频域解卷积的“扰动不确定性”。基于建立的车架系统动力学模型，探讨了瞬态模型和卷积模型的应用范围，并验证了提出的新方法。(4)从理论上建立和推导了适合多子系统多输入单输出模型，建立了以比较真实和计算传递函数作为判断输入干涉方向的依据，基于理论和试验推导了加速度/力传递函数与加速度/加速度传递函数的关系；搭建了车辆局部结构的车架系统，基于相干分析进行了单向干涉模型的验证；针对某商用车的振动异常问题，测试分析了其传递路径中的激励和响应振动特征，基于双向干涉的平均方法，识别了激励源、路径的影响和贡献，从而验证了模型和方法的可行性，该方法有效提高了异常振动中的识别率和精度；对具有同样相关特征的某农用车辆噪声源进行了识别，结合一般传递路径分析方法，识别声源矛盾，并解决了该机械的噪声问题。综上所述，本文较系统地研究了车辆振动噪声中具有多相关特征的信号特征，并就其分离和识别问题进行研究，提出信号自适应特征提取方法、不同模型的盲分离算法，对不同干涉模型的路径进行了识别，进而实车验证了方法的可行性。研究结果不但有益于提高车辆故障诊断识别精度，而且在理论和方法上有益于车辆振动噪声设计以及后期优化等方面的研究。更多还原

【Abstract】 The vibration and noise of vehicle is a most important Evaluation index of comfort and also isvery important to monitor the running state of mechanism. Generally, the testing of vehicle iscarried out in the movement process, so a testing signal may not directly reflect the sourceinformation of the corresponding position when sensor is affected by different excitation source. Itis influenced by the environment noise, test systems or interference effects of different sources tothe position, resulting in correlation between measurement signals. To analyze signal responsecontribution rate, the real source contribution rate is exaggerated or reduced. For vibration faultdiagnosis, source and path identification under multi-correlation background, the main contents ofthis paper are summarized as follows:(1) The main characteristics of Empirical mode decomposition (EMD) was discussed in thispaper. The false component and mode mixing in decompositing process was improved, which wasapplied to do denoising and extract feature of vehicle testing signal. The method of using thecoherence analysis to avoid the misjudgment of weak signal and false component was proposed.From the perspective of analyzing frequency-energy of signal components, the influence of falsecomponent generating in EMD decomposition was researched and the new regularity thatincreasing high frequency component or reducing low frequency component can reduce the ratio offalse component was discovered. By studying the EMD decomposing characteristics of white noise,the method of multiple stacking colored noise to improve mode mixing of the EMD was proposed(MH-EMD, EMD with mutil-high frequecy), which simulation effect is obviously improvedcompared to the classic Ensemble Empirical Mode Decomposition (EEMD). The system methodwith improved EMD was applied to extract signal features from rotating machinery and exhaustsystem, the new method was verified to be feasible and effective.(2) A new adapted denoising method was proposed in paper, as is the joint of ensembleempirical mode decomposition (MH-EMD) and least mean square algorithm (LMS). The algorithmperformance of the LMS with fixed step and fixed filter order, variable step and variable order werestudied, it was proposed that convergence direction of algorithm with both order and step changingin iterative process is up to the expectation of least mean square error, a kind of jointing order andstep LMS (VSVT-LMS) was developed, and Using MH-EMD, the original signal was decomposed,so that every mode component would be narrowband, and then denoising by VSVT-LMS. The instability of the LMS for wideband signals, but also the EMD threshold selection problem isavoided effectively. Simulation and vehicle signal were analyzed and it shows that the new methodhas a good adaptive characteristics and accuracy, and the feasibility of the method was verified inengineering.(3) For coupled or related characteristics of testing signal on vehicles, some separationmethods were proposed. First, the classical methods of blind source separation were discussed.Based on the comparison of the amplitude density, a blind signal separation method was proposedto the transient mixed model. If the frequency of the signal is less, the method can be used toeffectively identify the single frequency and common frequency signal. The joint approximatediagonalization method is improved, the segments joint approximate diagonalization withequalization and optimal distance matrix joint approximate diagonalization method were proposed,the improved methods can achieve higher precision. Then, using the high order section of higherorder cumulants, and combining with the approximate diagonalization and joint block diagonalmatrix, which method was used better when sensor number is less than the source number.Combining the improved wavelet subbands with the blind source separation, the separation of theobservation signal with the characteristics of statistics relevance source could be effectively solved;For the mixed model with convolution properties, comparising separation algorithms of the timedomain and frequency domain, a frequency domain algorithm based on independent judgment ofmutual information vector was proposed, nonlinear function including the separation matrix of allthe frequency band was considered to avoid the disturbance uncertainty of frequency-domaindeconvolution in the iterative process. Based on the system dynamic model of the frame, theapplication range of transient model and the convolutive model was studied and new methods wereverified.(4) The model of multiple input and single output which is good for subsystem wasestablished and derived from the theory. Then, by constrasting to calculation transfer function andreal transfer function, the analysis method of input interference direction was proposed, and therelation of the acceleration with force transfer function and acceleration with acceleration transferfunction were derived from the theory and experiment. A bench of frame system was built, basedon coherent analysis, the one-way interference model was verified, and for the vibration problem ofa commercial vehicle, the vibration characteristics of excitation and response in the transmissionpaths were tested and analyzed. Base on the method of bidirectional interference average, thecontribution of excitation source and transfer path were effectively identified, which verify the feasibility of the model and method. Others, the recognition rate and accuracy of abnormalvibration diagnosis were improved by using new method. Combined with the general transfer pathanalysis method, which can be applied to effectively identify sound source contradiction and solvethe problem of the mechanical noise for the relevant characteristic noise source on the farm vehicle.In conclusion, the characteristics of a multi correlated subsystem were systematically studiedand the issue of its separation and identification was also studied in this paper. The method ofsignal self-adaptive feature extraction, the algorithm of blind source separation for different modelsand the mathematical expressions of different interference model were proposed.These methodsare verified in the vehicle.The results have an important theoretical significance and engineeringapplications for improving fault diagnosis accuracy and vehicle NVH performance, and guidingvehicles NVH designs.更多还原