基于可预测偏最小二乘算法的复杂工况过程的监控技术

发布时间：2018-01-08 23:26

本文关键词：基于可预测偏最小二乘算法的复杂工况过程的监控技术　出处：《上海交通大学》2015年硕士论文　论文类型：学位论文

【摘要】：随着信息采集、传输、存储和处理技术的不断发展,工业过程中有大量反映生产过程和设备运行的数据被采集和存储,如何有效地利用这些离线和在线数据,提取出能够反映工业过程特征的信息,用于工业过程的监控,以此保证设备安全运行,提高生产效率和产品质量,成为了目前故障检测与诊断领域的重点研究内容。基于数据驱动的过程监控技术应运而生,已成为未来过程监控技术的重要发展方向。可预测元分析(Fore CA)作为新兴的数据降维技术,它能够找到一个最优的转换方式将多变量时间序列分解为一个可预测空间和一个白噪声空间。分解得到的可预测性因子能够反映过程的本质特征,这主要是由于该算法是从时间序列的角度出发,将过程数据的序列相关性考虑在内,且其将时域特性变换到频域,在频域利用信息熵来衡量不确定性,从而保证较好的可预测性。鉴于Fore CA算法的上述优点,本文将其引入过程监控领域,并将其用于回归,与偏最小二乘(PLS)算法相结合,提出一种基于可预测偏最小二乘(Fore PLS)的故障检测模型,并利用其回归预测性能进行多故障诊断,最后将其与时间序列分析方法相结合实现对缓变故障的预测。为探索复杂工况下的故障检测、诊断与预测方法做出了有益的尝试。下面具体介绍一下本文的主要工作:(1)将Fore CA算法用于回归并与PLS方法相结合,提出了可预测偏最小二乘(Fore PLS)方法。该算法能够提取出过程数据特征空间中与质量变量相关的可预测性特征。(2)将提出的Fore PLS方法用于故障检测,构建基于Fore PLS的故障检测模型,并根据Fore PLS算法的特点构造了CUSUM统计量和SPE统计量,用来进行故障的检测。(3)提出了基于DFore PLS回归预测的多故障诊断方法,为了解决多类分类中的不平衡分类问题,将主动学习引入故障诊断领域,有目的地挑选边界附近最有“信息量”的样本进行训练,避免了冗余样本对分类器精度的影响,提高了分类器对故障样本的识别能力,同时也提高了分类器的训练效率。(4)将Fore PLS模型与向量自回归模型结合,提出了一种针对缓变故障的基于时间序列的故障预测方法。能够有效防止这类缓变故障对系统带来的损失,同时可以避免频繁更换部件,提高了生产效率。
[Abstract]:With the continuous development of information collection, transmission, storage and processing technology, a large number of industrial processes reflect the operation of the production process and equipment data acquisition and storage, how to effectively use these offline and online data, extract the feature information can reflect the industrial process, to monitor the industrial process, in order to ensure the safe operation of equipment. Improve the production efficiency and product quality, has become the current field of fault detection and diagnosis. The key research contents emerged process monitoring technology based on data driven, has become an important direction for future development of process control technology. Predictable element analysis (Fore CA) as a new dimensionality reduction technique, it can find an optimal conversion the multivariate time series is decomposed into a predictable space and a white noise space. The decomposed predictability factor can reflect the process of the The character, this is mainly because the algorithm is starting from the perspective of time series, the serial correlation of process data into account, and the time domain is transformed into frequency domain, in the frequency domain using the information entropy to measure the uncertainty, so as to ensure good predictability. In view of the advantages of Fore CA algorithm, the the introduction of process monitoring field, and used regression and partial least squares (PLS) algorithm are combined to propose a prediction based on partial least squares (Fore PLS) fault detection model, and by using the regression prediction performance of multi fault diagnosis, and the time series analysis method combined to predict slow in order to explore the fault. The fault detection under complex condition, diagnosis and prediction method is a beneficial attempt. The main work of this paper introduces in detail below: (1) Fore CA and PLS algorithm for regression A combination method proposed can predict the partial least squares (Fore PLS) method. The algorithm can extract the predictable characteristics associated with quality variables in the feature space of process data. (2) Fore PLS the proposed method is used for fault detection, fault detection module Fore construction based on PLS, and construct CUSUM statistics according to the statistic characteristics of Fore and SPE PLS algorithm, used for fault detection. (3) propose a method for fault diagnosis of DFore based on PLS regression prediction, in order to solve the multi class classification of unbalanced classification problem, the active learning into the fault diagnosis field, to select the most "near the boundary information" the training samples, to avoid the influence on the precision of the classification of the redundant samples, improve the recognition ability of the classifier for fault samples, but also improve the classifier training efficiency. (4) the Fore PLS model and vector auto Based on regression model, a fault prediction method based on time series for slowly varying faults is proposed. It can effectively prevent such slow failures from causing damage to the system, and avoid frequent replacement of components and improve production efficiency.

【学位授予单位】：上海交通大学
【学位级别】：硕士
【学位授予年份】：2015
【分类号】：TP277

【参考文献】