机械状态流形特征增强理论及监测诊断方法研究

发布时间：2018-01-07 04:12

本文关键词：机械状态流形特征增强理论及监测诊断方法研究　出处：《中国科学技术大学》2017年博士论文　论文类型：学位论文

【摘要】：本论文以机械设备状态的精确辨识与故障的有效诊断为研究目标,着眼于流形学习对本质特征信息上的挖掘能力,通过分析基于状态维护的状态特征、模式特征、信号特征三种特征信息的特点,在理论上提出了参照化流形、深度化流形以及稀疏化流形三种流形增强学习新模式,在内容实现了对状态差异特征、模式敏感特征、信号模态特征的流形增强学习,建立了一种流形特征增强学习的设备状态监测与故障诊断的研究体系,并分别就上述三种流形增强学习的若干理论和问题进行了深入研究。在设备状态特征增强学习中,基于机械设备具有长期健康运行、非健康状态同健康状态存在着相对较大差异的物理特点,对监测样本建立以相同数据为参照样本的实时比对模型,利用流形学习获取该模型参照化流形特征空间,通过分析空间聚类迁移变化来刻画设备状态的退化情况。该参照化流形空间聚类分布有效地突出了监测状态相比于健康状态的变化信息,揭露了设备状态的差异特征。基于状态特征构造方式和流形空间的不同,分别发展了多元统计特征的流形空间聚类和无特征的拉斯曼流形流形基空间聚类两种状态监测方案。并进一步提出了参照化流形空间聚类控制图,实现对设备状态退化的多阶段量化分析以及早期退化时刻的放大和报警。在故障模式特征增强学习中,基于参照化流形对于差异性特征增强的作用,建立多单元比对模型对原始特征形成扩展学习、进一步利用多层流形学习实现对特征的级联式多层学习。通过这种多层化学习方式,建立了如同多层神经网络特征学习模式,在这种特征扩展再学习过程中完成对模式敏感特征的增强学习。本文利用这种级联式深度化流形学习方式,提出了多比对模型的两层流形特征增强算法,有效的增强了不同故障类型之间的差异性、增强了模式特征敏感性,对于实际故障识别显示了与传统的流形特征识别方法的优越性。在信号模态特征增强学习中,利用时频流形能够提取瞬态特征本质信息的优势,引入稀疏分析原理,完成两者优势互补,重新建立一种稀疏化流形特征学习新模式,实现对信号稀疏模态特征的深度挖掘与再学习,克服传统时频流形学习在原始瞬态特征失真以及无法去除强噪声的缺陷。基于这种稀疏化流形分析的思想,分别提出了时频匹配稀疏和包络移不变稀疏的流形模态特征增强方法,实现了对原始信号波形特征的保持与恢复,有利于精确的故障模态诊断研究。综上所述,本论文研究了流形特征增强学习在CBM中的理论方法,包括状态监测预警、故障模式识别、故障信号诊断三个方面,提出了参照化流形、深度化流形以及稀疏化流形特征增强学习方法,进行了一次较为完整的特征增强学习的研究。和传统方法对比,流形特征增强可以实现更有效、更精准、更敏感的状态监测与故障诊断分析,对系统化的机械设备CBM研究具有重要意义。
[Abstract]:In this paper, the effective diagnosis accurate identification and fault state of the machinery and equipment as the research object, focusing on the manifold learning the essential characteristics of information mining capacity, through the analysis of characteristics, state maintenance based on mode characteristics, characteristics of signal characteristics of three kinds of feature information, in theory, puts forward a reference of the depth of the manifold manifold. And the sparse manifold three manifold reinforcement learning mode, characteristics of the state is realized by a difference in content, pattern sensitive features, manifold signal modal characteristics of reinforcement learning, a manifold learning feature enhancement system of state monitoring and fault diagnosis of equipment, and these three manifolds enhanced several theories and problems learning is studied. In the equipment characteristics of reinforcement learning, mechanical equipment has a long-term healthy operation based on non health status with the healthy state of existence Relatively large differences in physical characteristics, to establish monitoring samples with the same data for real-time alignment model reference samples, learning to obtain the reference models of manifold feature space using manifold, through the degradation of spatial clustering analysis to characterize the change of equipment condition. According to the distribution of cluster manifold space effectively highlights the change of state monitoring information compared to the healthy state, reveals the differences of equipment condition features. Feature structure and manifold space based on the different were the development of a manifold space clustering multivariate statistical feature and non feature based spatial clustering Larsemann manifold two state monitoring scheme. And put forward the reference of manifold space clustering control chart, realization the degradation of device status and quantitative analysis of multi stage amplification and early degradation time alarm. In fault pattern feature enhancement study According to the study, to enhance the diversity of manifold features based on the original features of the formation of an extended learning based multi unit alignment model, further use of manifold learning to achieve multilayer cascade multilayer characteristics of learning. Through this kind of multilayer learning methods, such as the establishment of a multilayer neural network with the characteristics of learning mode, in this feature expansion again learn to complete the reinforcement learning pattern of sensitive features in the process. Using this kind of cascade depth of manifold learning method, and puts forward two layers of the model than the manifold feature enhancement algorithm, the difference between the effective enhancement of different fault types, enhanced pattern features of sensitivity for the actual fault identification shows the superiority and manifold features the traditional recognition methods. In signal modal characteristics of reinforcement learning, using time-frequency manifold can extract the transient characteristics of essential information advantage, Introduce the principle of sparse analysis, complete complementary advantage, the re establishment of a new model of sparse manifold learning characteristics, implementation and study depth of excavation of sparse modal characteristics, to overcome the traditional time-frequency manifold learning defects in the original distortion transient characteristics and unable to remove the strong noise. The analysis of this sparse manifold based on the idea. When the frequency matching are proposed and sparse envelope method to enhance shift invariant sparse modal manifold, implement maintain and restore the original signal waveform characteristics, which is helpful to the study of modal fault diagnosis accurately. To sum up, this paper studies the characteristics of reinforcement learning theory of manifold method in CBM, including the state monitoring and early warning fault pattern recognition, fault diagnosis, signal three aspects, put forward the reference of the depth of the manifold, manifold and manifold feature sparse reinforcement learning method, in a A relatively complete research on feature enhanced learning. Compared with traditional methods, manifold feature enhancement can achieve more effective, more precise and sensitive state monitoring and fault diagnosis analysis, which is of great significance for the research of systematic mechanical equipment CBM.

【学位授予单位】：中国科学技术大学
【学位级别】：博士
【学位授予年份】：2017
【分类号】：TH17

【相似文献】