基于流形学习的智能诊断方法研究
发布时间：2018-10-14 10:09
【摘要】：故障诊断的实质是模式识别,主要研究内容包括信号获取、特征提取和模式分类三个方面。特征提取是故障诊断技术中最困难而又关键的环节,它直接影响故障诊断结果的准确性和故障早期预报的可靠性。因此,在复杂运行工况下,如何提取最优的低维故障特征来提高故障分类性能是一个巨大的挑战。本论文以流形学习算法为基础,深入研究了基于流形学习的特征提取与诊断技术。 针对复杂故障设备多个特征参数之间存在冗余性或不相关性,可能会增加后续分类器的时间消耗,甚至会降低故障的识别精度,提出了基于边界Fisher分析(MFA)算法的诊断模型。为了准确而全面地获取设备的故障信息,该模型采用多种信号处理方法进行分析,从多角度提取多个特征参数来表征设备的运行状态；运用MFA算法,从原始高维特征集中提取最具代表性的低维流形特征,并将所有低维特征输入K近邻分类器进行故障识别。通过对滚动轴承早期故障的诊断分析,验证了该模型的可行性和有效性。 针对机械设备故障诊断这种小样本模式识别问题,提出了正则化核边界Fisher分析(RKMFA)的特征提取算法及基于该算法的诊断模型。该模型运用RKMFA算法,直接从原始高维振动信号中提取低维流形特征,并将这些具有判别信息的少数几个流形特征输入K近邻分类器,最终识别出机械系统的故障模式。将该模型分别应用于轴承故障类型和内圈损伤程度的识别,实验结果表明RKMFA算法是一种有效的特征提取算法,同时验证了该模型的优越性。 针对机械设备故障诊断过程中获取有标签故障样本比较费时费力,提出了半监督核边界Fisher分析(SSKMFA)的特征提取算法及基于该算法的诊断模型。该模型运用SSKMFA算法直接对原始高维振动信号进行学习,通过大量廉价的无标签故障样本和少量昂贵的有标签故障样本估算故障数据的潜在流形结构,并在有标签故障样本提供的监督信息的引导下,学习出整个流形上的类别信息,从而提取具有判别性的低维流形特征,使得无标签故障样本获得良好的分类效果。将该模型分别应用于轴承故障类型以及故障严重程度的识别和齿轮箱故障类型的诊断,实验结果表明该模型能大大提高故障识别精度,同时降低算法的计算复杂度。 针对基于流形学习算法提取的低维流形特征没有明确的物理意义,导致其在故障诊断方面的理解性比较差等问题,提出了MFA分值的特征选择算法及基于该算法和支持向量机分类器的诊断模型。该模型采用多种信号处理方法对故障信号进行分析,得到一个由多个特征参数构造的原始高维特征集；运用MFA分值算法,挖掘隐藏在原始高维特征中的内在规律性,从而挑选出充分反映故障本质的敏感特征子集,将其输入到SVM分类器中,最终识别出设备的运行状态。在滚动轴承故障类型和内圈损伤程度的诊断实验中,验证了该模型的优越性。
[Abstract]:The essence of fault diagnosis is pattern recognition. The main research contents include signal acquisition, feature extraction and pattern classification. Feature extraction is the most difficult and critical link in fault diagnosis technology. It directly affects the accuracy of fault diagnosis and the reliability of early prediction. Therefore, it is a great challenge to extract the optimal low-dimensional fault feature to improve the fault classification performance under complex operating conditions. Based on manifold learning algorithm, this paper studies the feature extraction and diagnosis technology based on manifold learning. Aiming at the redundancy or non-correlation among multiple characteristic parameters of complex fault equipment, it is possible to increase the time consumption of subsequent classifier, even reduce the recognition precision of fault, and put forward the diagnosis based on the boundary Fisher analysis (MFA) algorithm. In order to obtain the fault information of the equipment accurately and comprehensively, the model adopts a variety of signal processing methods to analyze, extracts a plurality of characteristic parameters from multiple angles to characterize the running state of the equipment, and uses the MFA algorithm to extract the most representative low-dimensional manifold from the original high-dimensional characteristic set. Feature, and enter all low-dimension features into K-nearest classifier for failure The feasibility and feasibility of the model are verified by the diagnosis and analysis of the early faults of rolling bearings. Aiming at the problem of small sample pattern recognition of mechanical equipment fault diagnosis, a feature extraction algorithm of regular kernel boundary Fisher analysis (RKMFA) is proposed, and the algorithm based on this algorithm is proposed. The model applies the RKMFA algorithm to extract the low-dimensional manifold feature directly from the original high-dimensional vibration signal and inputs the few manifold features with the discrimination information into the K-nearest classifier, and finally identifies the mechanical system. The model is applied to the identification of the type of bearing failure and the degree of damage of the inner ring respectively. The experimental results show that the RKMFA algorithm is an effective feature extraction algorithm, and the model is verified at the same time. This paper presents the feature extraction algorithm of the semi-supervised kernel boundary Fisher analysis (SSKMFA) and based on the comparison of the tag fault samples in the fault diagnosis of mechanical equipment. The model uses SSKMFA algorithm to directly study the original high-dimensional vibration signals, estimates the potential manifold structure of fault data through a large number of cheap non-label fault samples and a small amount of expensive tag fault samples, and provides them with label fault samples. under the guidance of supervision information, the class information on the whole manifold is learned so as to extract the characteristic of the low-dimensional manifold with discrimination, so that the non-label fault sample is obtained. The model can be applied to the identification of bearing fault types and the severity of faults and the diagnosis of gearbox fault types respectively. The experimental results show that the model can greatly improve the accuracy of fault recognition and reduce the same time. In this paper, a feature selection algorithm for MFA score and its support based on manifold learning algorithm are presented. A diagnosis model of a vector machine classifier is used to analyze the fault signals by adopting a plurality of signal processing methods to obtain an original high-Viterbi collection constructed by a plurality of characteristic parameters, wherein the mining is hidden in the original high-Viterbi algorithm by using the MFA score algorithm. the inherent regularity in the high-dimensional features is selected to select a subset of sensitive features that adequately reflect the nature of the fault, which is input into the SVM classifier, most The operation state of the equipment is finally identified. In the diagnosis experiment of the fault type of rolling bearing and the degree of damage of the inner ring
【学位授予单位】：华中科技大学
【学位级别】：博士
【学位授予年份】：2013
【分类号】：TH165.3
本文编号：2270138
[Abstract]:The essence of fault diagnosis is pattern recognition. The main research contents include signal acquisition, feature extraction and pattern classification. Feature extraction is the most difficult and critical link in fault diagnosis technology. It directly affects the accuracy of fault diagnosis and the reliability of early prediction. Therefore, it is a great challenge to extract the optimal low-dimensional fault feature to improve the fault classification performance under complex operating conditions. Based on manifold learning algorithm, this paper studies the feature extraction and diagnosis technology based on manifold learning. Aiming at the redundancy or non-correlation among multiple characteristic parameters of complex fault equipment, it is possible to increase the time consumption of subsequent classifier, even reduce the recognition precision of fault, and put forward the diagnosis based on the boundary Fisher analysis (MFA) algorithm. In order to obtain the fault information of the equipment accurately and comprehensively, the model adopts a variety of signal processing methods to analyze, extracts a plurality of characteristic parameters from multiple angles to characterize the running state of the equipment, and uses the MFA algorithm to extract the most representative low-dimensional manifold from the original high-dimensional characteristic set. Feature, and enter all low-dimension features into K-nearest classifier for failure The feasibility and feasibility of the model are verified by the diagnosis and analysis of the early faults of rolling bearings. Aiming at the problem of small sample pattern recognition of mechanical equipment fault diagnosis, a feature extraction algorithm of regular kernel boundary Fisher analysis (RKMFA) is proposed, and the algorithm based on this algorithm is proposed. The model applies the RKMFA algorithm to extract the low-dimensional manifold feature directly from the original high-dimensional vibration signal and inputs the few manifold features with the discrimination information into the K-nearest classifier, and finally identifies the mechanical system. The model is applied to the identification of the type of bearing failure and the degree of damage of the inner ring respectively. The experimental results show that the RKMFA algorithm is an effective feature extraction algorithm, and the model is verified at the same time. This paper presents the feature extraction algorithm of the semi-supervised kernel boundary Fisher analysis (SSKMFA) and based on the comparison of the tag fault samples in the fault diagnosis of mechanical equipment. The model uses SSKMFA algorithm to directly study the original high-dimensional vibration signals, estimates the potential manifold structure of fault data through a large number of cheap non-label fault samples and a small amount of expensive tag fault samples, and provides them with label fault samples. under the guidance of supervision information, the class information on the whole manifold is learned so as to extract the characteristic of the low-dimensional manifold with discrimination, so that the non-label fault sample is obtained. The model can be applied to the identification of bearing fault types and the severity of faults and the diagnosis of gearbox fault types respectively. The experimental results show that the model can greatly improve the accuracy of fault recognition and reduce the same time. In this paper, a feature selection algorithm for MFA score and its support based on manifold learning algorithm are presented. A diagnosis model of a vector machine classifier is used to analyze the fault signals by adopting a plurality of signal processing methods to obtain an original high-Viterbi collection constructed by a plurality of characteristic parameters, wherein the mining is hidden in the original high-Viterbi algorithm by using the MFA score algorithm. the inherent regularity in the high-dimensional features is selected to select a subset of sensitive features that adequately reflect the nature of the fault, which is input into the SVM classifier, most The operation state of the equipment is finally identified. In the diagnosis experiment of the fault type of rolling bearing and the degree of damage of the inner ring
【学位授予单位】：华中科技大学
【学位级别】：博士
【学位授予年份】：2013
【分类号】：TH165.3
【参考文献】
相关期刊论文 前4条
1 阳建宏;徐金梧;杨德斌;黎敏;;基于主流形识别的非线性时间序列降噪方法及其在故障诊断中的应用[J];机械工程学报;2006年08期
2 杨叔子;设备诊断技术的现状与未来[J];设备管理与维修;1995年11期
3 张熠卓;徐光华;梁霖;;基于非线性流形学习的喘振监测技术研究[J];西安交通大学学报;2009年07期
4 梁霖;徐光华;栗茂林;张熠卓;梁小影;;冲击故障特征提取的非线性流形学习方法[J];西安交通大学学报;2009年11期
相关博士学位论文 前1条
1 刘永斌;基于非线性信号分析的滚动轴承状态监测诊断研究[D];中国科学技术大学;2011年
，本文编号：2270138
本文链接：https://www.wllwen.com/kejilunwen/jixiegongcheng/2270138.html
