机械加工过程中的早期故障微弱信号处理方法研究

发布时间：2018-03-26 07:21

本文选题：设备性能衰退　切入点：早期故障　出处：《哈尔滨工业大学》2016年博士论文

【摘要】：机械制造加工业是国家工业发展的基础,更是国家综合实力的重要体现,高可靠性和近零故障是有效进行机械制造加工的重要保障。早期故障诊断技术能够及时发现机械加工过程早期故障的微弱信息,根据故障发生的位置、类别以及严重程度提供有效的维修决策,对于实现高效可靠的机械加工过程具有重大意义。然而,早期故障的微弱信号具有幅值微弱、低信噪比、故障特征冗余度高以及高维非线性的特点。因此,如何从监测的低信噪比信号中,提取早期微弱故障特征,实现微弱信号主特征选择及高维非线性特征降维分析,有效实现机械加工故障的早期诊断是迫切需要解决的问题。针对早期故障的低信噪比微弱信号处理问题,在传统的经验模态分解方法基础上,提出了基于嵌入降噪的筛选迭代方法,对信号的筛选迭代分解过程进行降噪分析和准模态平滑处理,减小由端点缺失造成的扭曲现象,以解决噪声干扰和迭代误差造成的端点效应问题;同时,提出了基于相关分析的停止准则,考虑了迭代过程准模态间的局部关系,及分解模态与信号相关关系的全局特性,达到消除经验模态分解冗余模态的目的。对早期故障信号和低信噪比微弱信号的分析,以及与基于预测延拓的端点效应抑制方法的对比分析,验证了提出的方法解决经验模态分解方法端点效应和冗余模态的能力,为实现早期故障信号及低信噪比微弱信号的分析提供理论及技术支持。从空间域和频域分析了小幅值微弱信号,通过功率谱密度方法、二维小波方法和二维经验模态分解方法实现了微弱信号各频域分量的提取;并针对传统二维经验模态分解方法的冗余模态问题,提出了基于嵌入降噪迭代的二维经验模态分解方法,在迭代分解过程中消除了包络拟合造成的误差。对精密加工小幅值微弱信号的分析表明,提出的方法有效实现了晶体表面微观形貌的刀痕信息、进给方向的纹理特征及加工缺陷的特征识别,能够精确分析各形貌特征对晶体加工质量的影响,为从加工缺陷机理上改善加工质量提供技术支持。针对早期故障微弱信号的主特征选择问题,提出了基于动态遗传算法的主特征判定模型;研究高维非线性特征的动态编码机制,计算每个特征维度上特征集的适应度值,并选择该特征维度的最优特征集;根据精英保留策略动态更新特征空间的维度,迭代计算该维度的最优特征集;研究了基于受测试曲线分析(ROC)的系统敏感度辨识模型,对于设备的多元状态识别模型,分别计算每个模型的敏感度和特异性值,根据多元状态的辨识指标和特征累积频率判定系统的主特征。通过对转子和轴承性能衰退过程早期故障分析,以及与其它五种特征选择方法对比分析,表明提出的主特征选择方法能够以较小的特征维度获得较高的分类正确率。针对早期故障微弱信号高维非线性特征的降维问题,提出了考虑全局特征的互信息特征加权和流形优化的监督式局部线性嵌入方法。结合样本的类别信息,通过基于互信息分析的特征加权方法,保留全局特征的完整性并突出主要贡献的特征,实现了样本点邻域的加权选择;在此基础上,采用极大似然法实现了样本特征在低维空间的流形维度估计,和基于偏最小二乘法的测试样本流形映射。针对监督式局部线性嵌入方法获得低维流形的冗余性问题,通过基于互信息分析的流形排序方法,选择低维空间的最优流形组合。通过与其他的特征降维方法,以及不同的组合方法进行对比分析,验证了提出的改进的流形学习方法能够实现高维特征的非线性降维,并改善模型识别刀具早期故障的精度。
[Abstract]:Machinery manufacturing industry is the foundation of a country's industrial development, it is an important manifestation of the comprehensive national strength, high reliability and near zero fault is an important guarantee for effective mechanical processing. The early fault diagnosis technology can detect weak information of early fault in the machining process, according to the fault location, type and severity of providing effective maintenance decision, is of great significance for the machining process to achieve efficient and reliable. However, the weak signal with the amplitude of weak fault, low SNR, high redundancy and fault characteristics of high dimensional nonlinear characteristics. Therefore, how to monitor the low SNR signal, extracting early weak fault characteristics, analysis the main dimension of weak signal feature selection and high dimension nonlinear characteristics, is an urgent need to solve the problem of early diagnosis of mechanical fault needle effectively. Low SNR signal processing problem of early fault, based on the empirical mode decomposition method on the traditional iterative method is proposed based on embedded noise screening, screening of iterative signal decomposition process of noise analysis and quasi modal smoothing, reduce the distortions caused by the lack of the endpoint, the endpoint effect to solve the problem the noise and the error caused by iteration; at the same time, a new stop criterion is proposed based on the correlation analysis, the iterative process of quasi local relations between modes, and modal correlation and signal decomposition of the global characteristics, to eliminate redundant EMD modal purposes. For early fault signals and low SNR signal analysis the analysis and comparison of prediction and extension of extension of the endpoint effect suppressing method based on the proposed method to solve the end effect of empirical mode decomposition method and redundancy mode The ability to achieve the early fault signal and low signal-to-noise ratio to provide theoretical and technical support. The analysis of weak signal from the spatial domain and frequency domain analysis of the signal amplitude, power spectrum density method by wavelet method, two dimensional and two-dimensional empirical mode decomposition method to realize the extraction of weak signal of each frequency component; and the redundancy mode the problem of traditional two-dimensional empirical mode decomposition method, the decomposition method of two-dimensional empirical mode embedded noise reduction based on iteration, the iterative decomposition process eliminates the error caused by the envelope fitting. The precision machining of small amplitude analysis of weak signals show that the proposed method can effectively realize the morphology of the crystal surface mark information, texture feature recognition the defects and processing feed direction, can accurately analyze the impact of the morphology of the crystal processing quality, from the processing defect mechanism of change Provide technical support and good processing quality. For the problem of selecting main characteristics of early fault signal, puts forward the decision model of the main dynamic characteristics based on genetic algorithm; dynamic encoding mechanism of high dimensional nonlinear characteristics, the calculation of each feature dimension feature set fitness value, and choose the best feature the feature dimension to the elite collection; keep the dimensions of dynamic updating strategies of feature space, the optimal iterative calculation of this dimension is studied by analyzing the collection; test curve (ROC) system based on sensitivity identification model for multi state recognition model of the equipment, the sensitivity and specificity were calculated for each value of the model, according to the main characteristics of multi state and identification index the characteristics of the cumulative frequency of the system. Through the analysis of the rotor and bearing performance degradation process and early fault, and the other five kinds of feature selection methods for analysis, table The main feature that the proposed method is able to select the feature dimension smaller higher classification accuracy. Aiming at the problem of dimensionality reduction of early fault weak signal of high dimensional nonlinear characteristics, considering the features of global mutual information feature weighted and supervised manifold optimization method. Combined with the local linear embedding of the classification information of samples by feature weighting the method of mutual information based on the analysis of the characteristics of integrity preserved global features and highlight the main contribution of the weighting sample selection of neighborhood; on this basis, the sample characteristic manifold dimension in the low dimensional space is estimated by the maximum likelihood method, and based on the test sample manifold mapping partial least squares method. According to the method of supervision local linear embedding for the redundancy of the low dimensional manifold, manifold ranking by the method of mutual information based on the analysis of the optimal selection of low dimensional space Compared with other feature reduction methods and different combination methods, it is verified that the improved manifold learning method can achieve nonlinear dimensionality reduction of high-dimensional features, and improve the accuracy of model identification for early failure of tools.

【学位授予单位】：哈尔滨工业大学
【学位级别】：博士
【学位授予年份】：2016
【分类号】：TH16;TH17

【相似文献】