鲁棒PPLS建模及其在过程监控中的应用

发布时间：2018-03-11 04:15

本文选题：鲁棒　切入点：PPLS算法　出处：《江南大学》2017年硕士论文　论文类型：学位论文

【摘要】：近年来,工业生产中安全事故频发,为了减少财产损失和人员伤亡,保证生产的安全平稳运行和产品质量,可靠准确的过程监控至关重要。现代生产过程中保存了大量的数据,给基于数据驱动的统计监控方法的应用和发展提供了条件和基础。在传统过程监控方法中,通常假设样本数据服从正态分布,且样本的采样率一致。但是,实际变量数据分布复杂并且采样周期多样,给过程的监控带来了困难。本文针对实际工业中离群值,多采样率,动态特性等实际问题,对概率偏最小二乘(Probabilistic Partial Least Squares,PPLS)模型进行推广,提出了PPLS鲁棒建模方法,并将其应用到过程监控中,主要研究内容如下:1、针对工业过程中离群值问题,分析PPLS模型的不足,提出一种鲁棒PPLS方法,用拖尾更宽的T分布代替高斯分布,通过调整自由度参数,使模型对含离群点数据的拟合效果更好。更进一步,将鲁棒PPLS引入过程监控中,构建GT2和GSPE两个监控指标,通过监控主元和残差两个子空间,判断系统运行状况。PPLS和鲁棒PPLS在TE过程监控的应用结果表明鲁棒PPLS不仅能更准确检测故障的产生,而且能更有效降低故障的漏报率。2、针对工业过程中多采样率问题,基于半监督方法,提出一种半监督鲁棒PPLS方法,将采样率不一致的完整数据分成少数标记样本和大量未标记样本,然后分别用这两种样本数一致的数据建立鲁棒PPLS模型,通过充分挖掘大量未标记数据中的有用信息来提高模型的准确性。另一方面,通过建立GT2和GSPE_x和GSPE_y三个监控指标,半监督鲁棒PPLS完成了对主元空间、过程变量的噪声空间和质量变量的噪声空间的监控。通过对半监督鲁棒PPLS和下采样鲁棒PPLS在TE过程监控应用中比较,结果表明半监督鲁棒PPLS比降采样鲁棒PPLS效果更好。3、针对实际工况下过程动态特性问题,基于状态空间方式扩展过程数据矩阵,提出一种动态鲁棒PPLS的数据建模方法。动态鲁棒PPLS不仅考虑了变量之间的关联性,而且还提炼出变量在时间维度上的动态性,从而能对过程进行准确描述提高模型精度。此外,基于动态鲁棒PPLS,引入GT2和GSPE两个指标,通过充分融合过程动态和静态的有用信息实现对动态过程准确及时地监控。在TE过程中的应用研究,反映了动态鲁棒PPLS能更准确有效的监控过程故障的发生。
[Abstract]:In recent years, safety accidents occur frequently in industrial production. In order to reduce the loss of property and casualties, ensure the safe and stable operation of production and product quality, reliable and accurate process monitoring is very important. It provides the condition and foundation for the application and development of data-driven statistical monitoring method. In the traditional process monitoring method, it is usually assumed that the sample data is normally distributed, and the sample sampling rate is the same. It is difficult to monitor the process because of the complicated distribution of the actual variable data and the variety of sampling period. This paper aims at the practical problems such as outlier value, multi-sampling rate, dynamic characteristic and so on in the actual industry. The probabilistic Partial Least SquaresPPLS model is generalized, and a robust modeling method of PPLS is proposed and applied to process monitoring. The main research contents are as follows: 1. Aiming at the problem of outliers in industrial process, the deficiency of PPLS model is analyzed. In this paper, a robust PPLS method is proposed, in which Gao Si distribution is replaced by T distribution with a wider tail. By adjusting the parameters of degree of freedom, the model can fit outliers better. Furthermore, robust PPLS is introduced into process monitoring. Two monitoring indexes, GT2 and GSPE, are constructed. By monitoring the principal component and residual subspace, the application results of system operation. PPLS and robust PPLS in te process monitoring show that robust PPLS can not only detect the occurrence of faults more accurately. Moreover, it can reduce the failure rate more effectively. Aiming at the problem of multi-sampling rate in industrial process, a semi-supervised robust PPLS method is proposed based on semi-supervised method. The complete data with inconsistent sampling rate are divided into a few labeled samples and a large number of unlabeled samples, and then the robust PPLS model is established by using the data with the same number of samples, respectively. The accuracy of the model is improved by fully mining useful information from a large amount of unmarked data. On the other hand, by establishing three monitoring indexes, GT2, GSPE_x and GSPE_y, semi-supervised robust PPLS completes the principal component space. The noise space of process variable and the noise space of quality variable are monitored. The comparison of semi-supervised robust PPLS and down-sampling robust PPLS in te process monitoring application is carried out. The results show that semi-supervised robust PPLS is more effective than down-sampling robust PPLS. The process data matrix is extended based on state space to solve the problem of process dynamic characteristics under actual working conditions. A data modeling method for dynamic robust PPLS is proposed. Dynamic robust PPLS not only considers the correlation among variables, but also abstracts the dynamics of variables in time dimension, which can accurately describe the process and improve the accuracy of the model. Based on dynamic robust PPLS, two indexes, GT2 and GSPE, are introduced to realize accurate and timely monitoring of dynamic process by fully integrating dynamic and static useful information. It reflects that dynamic robust PPLS can more accurately and effectively monitor the occurrence of process failures.
【学位授予单位】：江南大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：X924;TP277

【参考文献】