基于KECA的非线性故障检测

发布时间：2018-01-28 03:53

本文关键词： 过程监测非线性故障检测核方法流形学习　出处：《浙江大学》2017年硕士论文　论文类型：学位论文

【摘要】：过程监测系统能够实时地监测生产过程,在保障工况平稳运行、改善产品质量及降低能耗等方面越来越发挥着不可替代的作用。大数据时代正在随着信息化程度不断发展以及硬件存储和计算水平的不断提升而到来,数据的极大丰富使得基于数据驱动的过程监测方法成为了近年来研究的热点。相应地,这些丰富的数据也使得基于数据驱动的过程监测方法面临着更多的挑战。本文针对工业生产过程当中的非线性特性,采用核方法和流形学习(Manifold Learning)提出了两种基于核熵成分分析(Kernel Entropy Component Analysis,KECA)的非线性故障检测方法。具体研究内容包括:(1)针对单一模型的KECA方法并不能够有效地检测工业过程当中存在的不同类型的故障的问题,提出基于集成学习和贝叶斯推论的改进KECA故障检测方法。由于不同类型的故障往往需要不同大小的核参数使得其具有较好的检测效果,本文采用相同的训练数据对不同核参数构造的KECA模型进行训练实现离线建模。在建立模型之后通过贝叶斯推论将这些模型的在线检测效果转化为概率的形式,最后将这些概率形式的检测结果根据加权方式组合形成一个最终的检测结果,给予对特定故障有较好检测效果的模型较大的权重,从而实现了对不同故障类型均具有较好的检测效果的目的。(2)考虑到KECA能够更全面地选择降维过程中数据的投影方向的优点,本文将信息熵的思想引入到最大方差展开(Maximum Variance Unfolding,MVU)当中,提出了一种基于KECA-MVU的非线性故障检测方法。利用瑞利熵来衡量由MVU学习得到的核矩阵经过数据变换之后的信息保留有效程度,根据瑞利熵最大的前几项所对应的特征向量作为数据的投影方向,实现了数据的有效压缩。最后采用线性回归的方法估计了输入数据到低维数结构的最优投影矩阵,由该投影矩阵实现过程中故障的在线检测。最后,总结了本文的主要研究成果,并阐述了未来研究工作的难点及展望。
[Abstract]:The process monitoring system can monitor the production process in real time and run smoothly in the guaranteed working conditions. Improving the quality of products and reducing energy consumption are playing an irreplaceable role. Big data era is coming along with the development of information technology and the continuous improvement of hardware storage and computing. With the abundance of data, the data-driven process monitoring method has become a hot topic in recent years. These abundant data also make the data-driven process monitoring method face more challenges. Using kernel method and Manifold learning (Manifold learning), two methods based on kernel entropy component analysis are proposed. Kernel Entropy Component Analysis. The specific research contents include: 1) the KECA method for a single model can not effectively detect the problems of different types of faults in the industrial process. An improved KECA fault detection method based on ensemble learning and Bayesian inference is proposed. Because different types of faults often require different kernel parameters, it has better detection effect. In this paper, we use the same training data to train KECA models with different kernel parameters to realize off-line modeling. After establishing the model, the on-line detection effect of these models is transformed into probabilistic form by Bayesian inference. Style. Finally, these probabilistic detection results are combined according to the weighted method to form a final detection result, which gives a larger weight to the model with better detection effect for a particular fault. Therefore, the purpose of better detection effect for different fault types is realized. (2) considering the advantage that KECA can more comprehensively select the projection direction of the data in the process of dimensionality reduction. In this paper, the idea of information entropy is introduced into the maximum Variance portfolio (MVU). In this paper, a nonlinear fault detection method based on KECA-MVU is proposed, which uses Rayleigh entropy to measure the effective degree of information retention after data transformation of kernel matrix obtained from MVU learning. According to the eigenvector corresponding to the first few terms of the maximum Rayleigh entropy as the projection direction of the data. Finally, the linear regression method is used to estimate the optimal projection matrix from the input data to the low-dimensional structure, and the on-line fault detection is realized by the projection matrix. Finally. This paper summarizes the main research results, and describes the difficulties and prospects of future research work.
【学位授予单位】：浙江大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP277

【参考文献】