【作者】 蒋玲莉；

【导师】 刘义伦； 李学军；

【作者基本信息】 中南大学 ， 机械电子工程， 2010， 博士

【摘要】 旋转机械是国民经济支撑行业中广泛使用的关键性核心设备,开展旋转机械故障诊断技术研究,对于确保此类设备安全、可靠、高效、长周期、满负荷、优质运行,避免巨大的经济损失和灾难性事故的发生,具有极大的经济、社会意义。旋转机械故障诊断过程实质是模式分析过程,以支持向量机为代表的核方法带来了模式分析领域的第三次革命,该类方法通过核函数将原始空间数据隐式映射到特征空间,在特征空间寻找线性关系,实现非线性问题的高效求解。旋转机械发生的故障常常呈现出非线性行为,核方法特别适用于旋转机械故障诊断与模式分析问题的处理。典型的核方法除支持向量机(SVM),还有核主成分分析(KPCA)、核独立成分分析(KICA)、核聚类(KC)、核Fish判别(KFD)等。论文围绕核方法及其在旋转机械故障诊断中的应用而展开,主要研究内容包括：1.面向核方法应用的旋转机械振动信号采集与预处理以机械故障综合模拟实验台为平台,设计了旋转机械典型故障振动测试实验,并采集了数据。提取时域统计特征、小波包分解和经验模态分解所得不同频段的能量、熵和能量熵特征、时序特征、关联维特征,构成11个旋转机械故障特征库,为核方法应用研究提供数据基础。针对EMD过程中易产生冗余IMF的不足,提出改进HHT方法。小波包去噪后,以EMD所得每个IMF与分解前信号的相关系数作为判断依据,剔除冗余IMF,滚动轴承故障诊断实例表明该方法可有效提取微弱信号故障特征频率。2.面向旋转机械故障诊断的KPCA核参数优选与消噪方法研究揭示了核函数及核参数对KPCA性能影响规律,发现无论选用高斯核还是多项式核,累计贡献率大于0.85的核主成分个数均随核参数的增大而递减,最终呈收敛状态；选用高斯核的KPCA,其核主成分具有更好的聚类性能,对没有任何先验知识的特征向量执行KPCA时,建议取高斯核参数σ≥25。提出了基于KPCA的信号消噪方法,克服滤波、小波消噪等常用消噪方法需要先验知识,给实际应用带来困难的不足。通过相空间重构将一维观测信号扩展为多维向量,再执行KPCA,提取核主成分,实现信号消噪,整个过程无需任何先验知识。仿真与滚动轴承振动信号消噪实例验证了方法的有效。3.基于KICA的旋转机械故障信号消噪与特征提取提出了基于KICA的信号消噪方法,通过引入适配噪声分量,将一维观测信号扩展为多维向量,再执行KICA,实现信噪分离,达到消噪目的。KICA消噪过程不受信噪比影响,具有其它消噪方法无法比拟的优势。转子不平衡振动信号的消噪实例验证了方法的有效性。定义了旋转机械KIC特征量,实测信号分析表明KIC对滚动轴承与齿轮故障均具有有较好的识别性,可作为敏感特征量用于故障诊断。4.基于(核)聚类的旋转机械故障诊断提出了EMD-模糊聚类法、基于高阶累积量的AR参数估计-模糊聚类法,滚动轴承故障诊断与性能退化评估实例验证了方法的有效性。提出了基于双谱分布区域的匹配聚类方法,有效克服了传统基于谱图频率峰值的故障诊断方法容易受混频干扰的不足,滚动轴承与齿轮的故障诊断实例验证了方法的有效性。提出了两种核聚类方法,重点研究了初始核聚类中心、及进一步核聚类中心的确定方法,实例验证了方法的有效性。5.基于SVM与多振动信息融合的旋转机械故障诊断提出了基于SVM的多振动信息融合旋转机械故障诊断方法,多个传感器的单一特征量的SVM融合可实现较高精度的滚动轴承与齿轮的故障诊断。研究了基于基座的旋转机械故障诊断,SVM对基座多传感器微弱信号执行信息融合,有效实现转子裂纹与滚动轴承的故障诊断。基座传感器安装方式具有通用性,可克服现场传感器安装不便等问题,具有极大的应用推广前景。更多还原

【Abstract】 Rotating machinery are the key equipments in supporting industry of national economy. It has great economic and social significance to study on the rotating machinery fault diagnosis, for it can ensure a safe, reliable, efficient, long period, full load and quality environment for equipment operation, avoid enormous economic loss and disastrous accident. The process of rotating machinery fault diagnos is a pattern analysis process, and the kernel method represented by Support Vector Machine brought the third revolution of pattern analysis. This method implicit maps original spatial data to feature space through kernel function, seeks for linear relationship in feature space to achieve the efficient solving of nonlinear problem. The rotating machinery faults are usual nonlinear behaviour, and the kernel method is applicable to rotating machinery fault diagnosis and modal analysis especially.Besides SVM, there are some typica kernel methods such as KPCA, KICA, KC, KFD and so on. This paper expanded around the kernel method and its application on rotating machinery fault diagnosis, the main research contents include:1. Acquisition and preprocessing of rotating machinery vibration signal aimed kernel method applicationIt experimented on vibration test of typical rotating machinery faults and collected datas based on Machinery Fault Simulator. Through extracting time domain features, energy, entropy, energy entropy, sequential characteristic and correlation dimension of different frequency bands by wavelet decomposition and empirical mode decomposition, it constituted eleven feature libraries, and provide a data base for applied research of kernel function. According to the deficiency that EMD is easy to produce redundancy IMF, it proposed a improved HHT method. After wavelet packet de-noising, it took correlation coefficients which between each IMF and original signal as the judgment basis, then rejected the redundancy IMF. The fault diagnosis example of rolling bearing shows that this method can extract weak signal fault feature effectively. 2. Optimization of KPCA kernel parameter and research of de-noising method aimed rotating machinery fault diagnosIt revealed the regular patterns of kernel function and kernel parameter on KPCA performance, discovered both gaussian kernel and polynomial kernel, when accumulative ratio above 0.85, the number of kernel principal component decreases with the increase of kernel parameter, and finally takes on convergence status. The KPCA which selected gaussian kernel has a better clustering performance on kernel principal component. It suggests choosing gaussian kernel parameterσ≥25 when conducting KPCA on feature vectors has no prior knowledge.It proposed a signal de-noising method based on KPCA, which overcomed the deficiency of common de-noising methods such as filter and wavelet de-noising need prior knowledge in practical application. It expanded one-dimensional observed signal as the multi-dimension vector through the phase space reconstruction, then extracted kernel principal component by KPCA to achieve signal de-noising, and the whole process needn’t any prior knowledge. It verified the effectiveness of this method by simulation and vibration signal de-noising example of rolling bearing.3. Fault signal de-noising and feature extraction of rotating machinery based on KICAIt proposed a signal de-noising method based on KICA, this method expanded one-dimensional observed signal as the multi-dimension vector through introducing the adaptation noise component, then executed KICA to implement noise separation, achieved the goal of de-noising. The de-noising process of KICA won’t be influenced by signal-to-noise ratio, and the other de-noising methods can’t be compared with it. It verified the effectiveness of this method by signal de-noising of unbalanced rotor.It defined the KIC characteristic quantity of rotating machinery, measured signal analysis showed that KIC can identify the fault of bearing and gear well, and it could be the sensitive characteristic quantity in fault diagnosis.4. Rotating machinery fault diagnosis based on (kernel) clustering It proposed a EMD-fuzzy clustering method and a AR-parameter estimation fuzzy clustering method, and verified the effectiveness of these methods by examples of frolling bearing fault diagnosis and evaluate the degree of performance degradation.It proposed a clustering match method based on the bi-spectrum distribution area, this method can effectively overcomed the deficiency that traditional fault diagnosis methods based on the peak of frequency spectrum are easy interfered by mixing, and verified this method’s effectiveness by rolling bearing and gear fault diagnosis examples.It proposed two kinds of kernel clustering algorithms, and key researched the determination method of initial kernel clustering center and further kernel clustering center, verified the effectiveness of this method by examples.5. Rotating machinery fault diagnosis based on SVM and multi-vibration information fusionIt proposed a multi-vibration information fusion rotating machinery fault diagnosis method based on SVM, it can achieve high precision rolling bearing and gear fault diagnosis through SVM fusion of single characteristic quantity from multi sensors.It studied on the rotating machinery fault diagnosis based on base, SVM fused weak signals from base multi-sensors, which effectively achieved the fault diagnosis of rotor crack and rolling bearing. Installing sensor on the base has generality which can overcome some on site questions such as inconvenience of sensor installation, and has a great application prospect.更多还原