基于核函数理论的改进FDA间歇过程故障诊断研究

发布时间：2018-04-08 20:50

本文选题：啤酒发酵　切入点：过程监控　出处：《哈尔滨理工大学》2016年硕士论文

【摘要】：有效的过程监控是工业生产过程中安全生产的重要保证,而且在改善产品质量和经济效益方面起到了重要作用。统计过程监控是以多元统计为理论基础的一种基于数据驱动的方法,通过对工业过程监控数据的处理和分析,获得工业过程的运行情况,在线检测和诊断过程中的异常状况,根据异常情况作出相应对策而指导整个系统运行、保证生产效率。本文结合间歇生产过程特点,对基于Fisher判别分析的过程监测方法进行了不同程度的发展和改进。主要研究内容及工作如下:1.对于复杂的间歇过程,在实现故障诊断过程中很可能碰见奇异矩阵的现象。为了解决这种奇异矩阵的问题,本文提出了一种基于奇异值分解的改进核Fisher判别分析算法,本文在Fisher判别分析方法的故障检测基础上,首先利用核函数将原始数据从原始空间映射到高维空间,然后利用奇异值分解的方法,将处理后的数据投影到一个分解后的非奇异正交矩阵中,最后利用Fisher判别分析算法来实现过程监测与故障诊断。2.工业过程数据往往来自多个数据源或异构的数据集,基于采用单个核函数的形式处理类似数据的效果不是很理想,本文基于多核学习的理论,采用多个基核组合形式,构成组合核函数的形式,与传统的Fisher判别分析相结合,提出一种组合核Fisher判别分析算法,通过啤酒发酵实验验证了算法的有效性。3.本文根据Fisher判别分析(有监督全局算法)的不足,利用核方法理论,提出了一种能够挖掘样本数据的全局欧氏分布结构和局部流行分布结构的核局部Fisher判别分析故障诊断算法,该算法充分利用局部保持投影与Fisher判别分析的优点,对于采样数据进行完全的信息挖掘,通过啤酒发酵过程,验证了该算法的优越性。
[Abstract]:Effective process monitoring is an important guarantee of production safety in industrial production, and plays an important role in improving product quality and economic benefits.Statistical process monitoring is a data-driven method based on multivariate statistics. Through the processing and analysis of industrial process monitoring data, the operation of industrial process, on-line detection and diagnosis of abnormal conditions in the process can be obtained.According to the abnormal situation to make corresponding countermeasures to guide the operation of the whole system, to ensure production efficiency.Based on the characteristics of batch production process, the process monitoring method based on Fisher discriminant analysis has been developed and improved in different degrees.The main contents and work are as follows: 1.For complex intermittent processes, singular matrices are likely to be encountered in the process of fault diagnosis.In order to solve the problem of singular matrix, an improved kernel Fisher discriminant analysis algorithm based on singular value decomposition is proposed in this paper.Firstly, the original data is mapped from the original space to the high-dimensional space by kernel function, and then the processed data is projected into a decomposed nonsingular orthogonal matrix by using the singular value decomposition method.Finally, Fisher discriminant analysis algorithm is used to realize process monitoring and fault diagnosis.Industrial process data often come from multiple data sources or heterogeneous data sets. The effect of processing similar data based on single kernel function is not ideal.Combined with the traditional Fisher discriminant analysis, a combined kernel Fisher discriminant analysis algorithm is proposed. The effectiveness of the algorithm is verified by beer fermentation experiments.In this paper, according to the deficiency of Fisher discriminant analysis (supervised global algorithm), a kernel local Fisher discriminant analysis fault diagnosis algorithm based on global Euclidean distribution structure and local popular distribution structure is proposed by using kernel method theory.The algorithm makes full use of the advantages of local preserving projection and Fisher discriminant analysis, and makes complete information mining for the sampling data. The superiority of the algorithm is verified by the beer fermentation process.
【学位授予单位】：哈尔滨理工大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：TP277

【参考文献】