鉴别性邻域嵌入方法及其在水电机组异常检测中的应用
发布时间：2019-06-27 15:57
【摘要】：可再生能源开发战略是国家十二五规划的重要组成部分。小水电是一种资源分布广、开发潜力大、环境影响小、可扩展利用的可再生能源,在国家能源发展战略上有着重大意义。在现阶段,考虑到水电机组的复杂性以及小水电站位置的苛刻性,通常采用专人值守的形式进行设备维护与异常监测。其过程不仅效率低下,而且过分依赖于工作人员的经验知识,往往具有较高的误判率,因此有必要研究机器学习理论与统计学理论并实现高性能识别算法,用于实现无人值守的小水电监测系统。邻域嵌入分析算法能够有效地进行数据分簇可视化操作,如何提升邻域嵌入算法的鉴别性能并应用于水电机组噪声源识别具有非常重要的研究价值。 本文针对依据噪声源进行的水电机组异常检测问题,分析了邻域嵌入分析算法的有监督扩展与线性投影扩展技术,设计了相应的鉴别性邻域嵌入分类算法。主要工作如下： (1)提出了基于拉斯优化的鉴别性邻域嵌入分类算法DEE。在异常检测应用任务中,现有的邻域嵌入分析算法由于缺少类别标签,识别率较为低下。DEE算法不仅能够直观地进行数据分簇可视化展示,而且通过拉斯方向提升了模型构建效率,使之得能胜任中大规模数据的鉴别任务。在3类公共数据集中验证了该算法的分簇能力与识别性能。 (2)通过引入核技巧,在保留其线性投影特性的同时将DEE进行非线性核化扩展,提出了两种不同类型的核化鉴别性邻域嵌入分类算法KDEE1与KDEE2。考虑到水电机组设备异常运行时振动噪声特征的非线性性质。在模型构建过程中,依据微分对象的不同,将核化的DEE版本称之为KDEE1和KDEE2,两者都能应用于非线性输入数据,且分簇能力与识别效率都较DEE有进一步的提升。 (3)分析了水电机组异常振动时所采集到的噪声特性及其预处理手段,将本文所提的KDEE1、KDEE2、DEE等鉴别性算法应用于实际的水电机机异常检测中。通过噪声特征子空间分簇以及识别率两种实验结果,表明核鉴别邻域嵌入分析算法具有较高的实用价值。
[Abstract]:Renewable energy development strategy is an important part of the 12th five-year Plan. Small hydropower is a kind of renewable energy with wide distribution of resources, great development potential, small environmental impact and extensible utilization, which is of great significance in the national energy development strategy. At present, considering the complexity of hydropower units and the rigour of the position of small hydropower stations, equipment maintenance and abnormal monitoring are usually carried out in the form of special personnel. The process is not only inefficient, but also too dependent on the experience and knowledge of staff, and often has a high misjudgment rate. Therefore, it is necessary to study machine learning theory and statistical theory and realize high performance recognition algorithm, which can be used to realize unattended small hydropower monitoring system. Neighborhood embedding analysis algorithm can effectively carry out data clustering visualization operation. How to improve the identification performance of neighborhood embedding algorithm and apply it to the identification of noise sources of hydropower units is of great research value. In this paper, aiming at the problem of abnormal detection of hydropower units based on noise sources, the supervised extension and linear projection expansion techniques of neighborhood embedding analysis algorithm are analyzed, and the corresponding discriminant neighborhood embedding classification algorithm is designed. The main work is as follows: (1) A discriminant neighborhood embedding classification algorithm DEE. based on Russ optimization is proposed. In the application task of anomaly detection, the existing neighborhood embedding analysis algorithms have low recognition rate due to the lack of category tags. Dee algorithm can not only intuitively display data clustering visualization, but also improve the efficiency of model construction through Russ direction, so that it can be competent for the identification task of medium and large scale data. The clustering ability and recognition performance of the algorithm are verified in three kinds of common data sets. (2) by introducing kernel technique, DEE is extended by nonlinear nucleation while retaining its linear projection characteristics, and two different types of kernel discriminant neighborhood embedding classification algorithms KDEE1 and KDEE2. are proposed. The nonlinear properties of vibration and noise characteristics of hydropower units in abnormal operation are taken into account. In the process of model construction, according to the difference of differential objects, the nucleated DEE version called KDEE1 and KDEE2, can be applied to nonlinear input data, and the clustering ability and recognition efficiency are further improved compared with DEE. (3) the noise characteristics and preprocessing methods collected during abnormal vibration of hydropower units are analyzed, and the KDEE1,KDEE2,DEE and other discriminant algorithms proposed in this paper are applied to the actual anomaly detection of hydropower machines. The experimental results of noise feature subspace clustering and recognition rate show that the kernel discriminant neighborhood embedding analysis algorithm has high practical value.
【学位授予单位】：浙江工业大学
【学位级别】：硕士
【学位授予年份】：2014
【分类号】：TV734;TV738
本文编号：2506936
[Abstract]:Renewable energy development strategy is an important part of the 12th five-year Plan. Small hydropower is a kind of renewable energy with wide distribution of resources, great development potential, small environmental impact and extensible utilization, which is of great significance in the national energy development strategy. At present, considering the complexity of hydropower units and the rigour of the position of small hydropower stations, equipment maintenance and abnormal monitoring are usually carried out in the form of special personnel. The process is not only inefficient, but also too dependent on the experience and knowledge of staff, and often has a high misjudgment rate. Therefore, it is necessary to study machine learning theory and statistical theory and realize high performance recognition algorithm, which can be used to realize unattended small hydropower monitoring system. Neighborhood embedding analysis algorithm can effectively carry out data clustering visualization operation. How to improve the identification performance of neighborhood embedding algorithm and apply it to the identification of noise sources of hydropower units is of great research value. In this paper, aiming at the problem of abnormal detection of hydropower units based on noise sources, the supervised extension and linear projection expansion techniques of neighborhood embedding analysis algorithm are analyzed, and the corresponding discriminant neighborhood embedding classification algorithm is designed. The main work is as follows: (1) A discriminant neighborhood embedding classification algorithm DEE. based on Russ optimization is proposed. In the application task of anomaly detection, the existing neighborhood embedding analysis algorithms have low recognition rate due to the lack of category tags. Dee algorithm can not only intuitively display data clustering visualization, but also improve the efficiency of model construction through Russ direction, so that it can be competent for the identification task of medium and large scale data. The clustering ability and recognition performance of the algorithm are verified in three kinds of common data sets. (2) by introducing kernel technique, DEE is extended by nonlinear nucleation while retaining its linear projection characteristics, and two different types of kernel discriminant neighborhood embedding classification algorithms KDEE1 and KDEE2. are proposed. The nonlinear properties of vibration and noise characteristics of hydropower units in abnormal operation are taken into account. In the process of model construction, according to the difference of differential objects, the nucleated DEE version called KDEE1 and KDEE2, can be applied to nonlinear input data, and the clustering ability and recognition efficiency are further improved compared with DEE. (3) the noise characteristics and preprocessing methods collected during abnormal vibration of hydropower units are analyzed, and the KDEE1,KDEE2,DEE and other discriminant algorithms proposed in this paper are applied to the actual anomaly detection of hydropower machines. The experimental results of noise feature subspace clustering and recognition rate show that the kernel discriminant neighborhood embedding analysis algorithm has high practical value.
【学位授予单位】：浙江工业大学
【学位级别】：硕士
【学位授予年份】：2014
【分类号】：TV734;TV738
【二级参考文献】
相关期刊论文 前1条
1 王金水;张东站;施秀升;赖兴瑞;;基于免疫系统和模糊逻辑的自适应网络入侵检测(英文)[J];心智与计算;2009年01期
，本文编号：2506936
本文链接：https://www.wllwen.com/kejilunwen/shuiwenshuili/2506936.html
