SIA OpenIR  > 工业信息学研究室
基于时间序列数据挖掘的旋转机械预诊断方法研究
Alternative TitleRotating Machinery Prognostics based on Time Series Data Mining
吴薇1,2
Department工业信息学研究室
Thesis Advisor胡静涛
ClassificationTM38
Keyword预诊断 时间序列数据挖掘 旋转机械
Call NumberTM38/W85/2009
Pages131页
Degree Discipline机械电子工程
Degree Name博士
2009-01-19
Degree Grantor中国科学院沈阳自动化研究所
Place of Conferral沈阳
Abstract基于状态的维护(CBM, Condition Based Maintenance)是近年来新兴的一种设备维护策略,它的基本理念是在机械设备需要维护的时候才对其进行维护,强调维护要及时、准确和经济。采用这种维护策略,能够提高工业生产的安全性和可靠性,系统地降低企业运营成本。机械设备状态预诊断是实现CBM 的核心支撑技术,对其进行深入研究,对推动CBM 的发展具有重要意义。但是,由于相关研究起步不久,目前预诊断技术还未能得到很好的实现,研究人员有必要不断地尝试各种新的有效方法来更好地解决这一问题,加快其实现方法与技术应用的成熟进程。基于此,本文从数据挖掘的角度,探索了机械设备预诊断新的解决方法和途径,深入研究和探讨了基于时间序列数据挖掘的旋转机械预诊断方法。本文的主要工作包括: 1. 结合CBM 的基本理念和应用实际的需求,对机械设备状态预诊断的基本内涵进行了系统分析。将状态评估、故障预测和剩余有效使用寿命预测三个预诊断基本功能进一步抽象,提出了包含特征提取、状态预测和模式匹配三个子问题的预诊断一般流程模式。在详细分析机械设备状态预诊断理论方法和应用技术研究现状的基础上,提出了预诊断技术研究的发展趋势及各子问题的研究侧重点。并对利用时间序列数据挖掘这一理论方法解决机械设备状态预诊断问题的可行性进行了分析。 2. 针对具有波动频繁、噪声干扰严重等特点的原始振动量时间序列无法直接用于旋转机械性能状态分析的问题,结合全息诊断信息融合分析旋转机械振动全貌的思想,提出了全息状态矩阵的概念并给出定义,用类时间轴上的多维序列表征转子系统振动全貌,以实现振动量时间序列的高级表示,为后续预测与匹配分类工作提供良好的数据源,同时增强全息诊断的信息检索和知识自动获取的能力。 3. 将旋转机械性能状态预测,归结为旋转机械设备维护应用背景下的一维数值型时间序列预测问题来进行深入研究。针对现有预测方法长期预测能力较弱,且自动化水平低的不足,提出了用于旋转机械性能状态预测的ARIMA 动态间隔预测法。该方法以动态间隔获取时间序列样本建模并预测的策略,提高了ARIMA 模型用于设备状态长期预测的准确性,并且能够实现建模与预测的自动化,满足CBM 系统的实时性要求。 4. 针对全息状态矩阵表示的旋转机械性能状态特征数据,提出了一种全息状态矩阵相似性匹配方法。结合旋转机械预诊断领域应用的特点定义了全息状态矩阵的相似性度量模型,基于全息状态矩阵近似距离三角不等式设计了剪枝搜索策略,并在此基础上设计了全息状态矩阵相似性高效准确匹配算法,不需要借助专家经验和人工识别确认,在一定阈值范围内能够实现高质量的旋转机械性能状态相似性匹配。 5. 旋转机械基本振动量特征时间序列具有海量、超高维度、短期波动频繁和大量噪声等特征,与时间序列数据挖掘传统应用的金融商业领域数据不同,直接采用传统方法会存在搜索速度大幅度降低的问题。针对这一问题,提出了基于随机投影的时间序列相似性搜索方法。该方法利用近年来新兴的随机投影统计学降维法,将原始时间序列集映射到低维空间,并利用R*树进行索引,能够在保持高准确率的同时,实现旋转机械基本振动量特征时间序列相似性快速搜索。 6. 针对现有机械设备性能状态分类方法不考虑误分类代价的问题,提出了一种代价敏感直推式旋转机械设备性能状态分类法。该方法将代价敏感分类和直推式学习的基本思想和理论相结合,采用一种代价敏感的直推式分类机制,实现了机械设备性能状态的代价敏感分类。该方法在保证较高分类准确率的基础上,明显地降低了误分类总代价。 7. 基于CBM 的基本理念,设计了旋转机械CBM 系统的基本结构,并以本文理论方法的研究成果为核心,详细设计了各模块的基本功能和处理逻辑,采用 VC#.net 与Matlab 混合编程的方式开发了一个面向大型旋转机械的CBM 系统原型,以验证本文机械设备预诊断方法研究成果的可操作性和实用性,为CBM 系统应用技术研究做出了有益的探索。
Other AbstractCondition Based Maintenance (CBM) is the future maintenance practice for equipment that is here today. CBM is a strategy aimed at extending machine life, increasing productivity, and taking machine health to the next level for the lifetime of the equipment. Unlike preventative maintenance, which is based on servicing a machine at scheduled intervals, CBM is based on specific equipment conditions including operating environment and application. Prognostics technology is an essential core element of any condition-based maintenance (CBM) system, but it is still in its infancy although some research work on developing has been done over the recent years. So it is necessary to explore new resolvents and approaches to implement prognostics in order to promote the boost the development of CBM. This thesis presents a prognostics approach for rotating machinery based on time series data mining. The main works and contributions of it are summarized as follows: 1. Summarized the basic meaning of prognostics combined with the philosophy and the practical requirement of CBM. Proposed a general prognostics flow mode composed of three sub-problems: feature extraction, condition prediction, pattern matching and classification. Based on analyzing in detail the current research situation, recommended the development trend of prognostics and the emphasized points of the tree sub-problems research. Analyzed the feasibility of using time series data mining to solve prognostics problems. 2. Defined the holospectrum condition matrix using pseudo multidimensional time series to express the rotation panorama of the rotator system borrowing ideas from of information fusion analysis for the rotating machinery of holospectrum technique. The holospectrum condition matrix achieves the normal vibration time series representation and can provide high quality data source for the following analysis steps. Meanwhile, it combines the holospectrum technique with the information searching and automatic knowledge acquisition capabilities of data mining, which may boost the perfection and development of holospectrum technology. 3. Rotating machine condition prediction may come down to the one-dimensional numerical time-series forecasting problems. Proposed an ARIMA predicting with dynamic intervals approach to predict the future machine status. This approach adopts the modeling and predicting strategy with dynamic intervals and achieves high accuracy automatic long term machine condition prediction. 4. Using pseudo multidimensional time series to express the rotation panorama of the rotator system, the holospectrum series similarity matching algorithm proposed achieves the condition pattern matching for rotating machine based on the searching strategy of the combination of the weaker triangle inequality. The method can effectively achieve a high quality automatic rotating machine condition pattern similarity matching. 5. The normal vibration time series is a kind of high frequency data. A fast algorithm of similarity pattern matching based on random projection for high frequency time series is proposed. In order to achieve the high-level representation of time series, this algorithm uses the random projection method to map the original time series to the lower space. Then, the spatial data index structure such as R* tree is built using the high-level representation of the original time series. It is a fast similarity searching algorithm with high accuracy for high frequency time series. 6. Most research on machinery condition pattern classification pursue to minimize the error rate without considering the misclassification cost. A new machinery condition pattern classification method based on a cost-sensitive transduction inference is presented. A Cost-sensitive transduction classification machine is proposed based on the Kolmogorov algorithm randomness theory and the minimum expected misclassification cost principle. This method can reduce the misclassification cost effectively with high classification accuracy. 7. Designed overall functions of the CBM prototype system for rotating machinery and the architecture of the prototype system, based on the in-depth understanding of CBM philosophy. Designed the basic functions and the processing logic of the modules, taking the research productions of this thesis as the core. Implemented tools are VC#.net and Matlab. This prototype system validated the maneuverability and the practicability of the productions of this dissertation research for prognostics, and provided an example for the further study and implementation of the practical follow-up project in CBM applications.
Language中文
Contribution Rank1
Document Type学位论文
Identifierhttp://ir.sia.cn/handle/173321/390
Collection工业信息学研究室
Affiliation1.中国科学院沈阳自动化研究所
2.中国科学院研究生院
Recommended Citation
GB/T 7714
吴薇. 基于时间序列数据挖掘的旋转机械预诊断方法研究[D]. 沈阳. 中国科学院沈阳自动化研究所,2009.
Files in This Item:
File Name/Size DocType Version Access License
10001_20051801471901(2195KB) 开放获取--Application Full Text
Related Services
Recommend this item
Bookmark
Usage statistics
Export to Endnote
Google Scholar
Similar articles in Google Scholar
[吴薇]'s Articles
Baidu academic
Similar articles in Baidu academic
[吴薇]'s Articles
Bing Scholar
Similar articles in Bing Scholar
[吴薇]'s Articles
Terms of Use
No data!
Social Bookmark/Share
All comments (0)
No comment.
 

Items in the repository are protected by copyright, with all rights reserved, unless otherwise indicated.