CIESC Journal ›› 2016, Vol. 67 ›› Issue (7): 2907-2915.DOI: 10.11949/j.issn.0438-1157.20151439

Previous Articles     Next Articles

Robust PPLS model and its applications in process monitoring

CHEN Jiayi, ZHAO Zhonggai, LIU Fei   

  1. Key Laboratory of Advanced Process Control for Light Industry Ministry of Education, Institute of Automation, Jiangnan University, Wuxi 214122, Jiangsu, China
  • Received:2015-09-11 Revised:2016-04-29 Online:2016-07-05 Published:2016-07-05
  • Supported by:

    supported by the National Natural Science Foundation of China (61573169) and the Six Talent Peaks Project of Jiangsu Province (2014-ZBZZ-010).

鲁棒PPLS模型及其在过程监控中的应用

陈家益, 赵忠盖, 刘飞   

  1. 江南大学自动化研究所, 轻工过程先进控制教育部重点实验室, 江苏 无锡 214122
  • 通讯作者: 赵忠盖
  • 基金资助:

    国家自然科学基金项目(61573169);江苏省六大人才高峰项目(2014-ZBZZ-010)。

Abstract:

A probability model can be developed by probabilistic partial least squares (PPLS) under the conditions that both principal components and errors satisfy Gaussian distribution. However, the expectation and variance of the Gaussian distribution is susceptible to outliers. As a result, the model is not robust for the real industrial process. This paper improves the robustness of the PPLS model based on the assumption that the raw data satisfy T distribution rather than Gaussian distribution. By adjusting the freedom degree of T distribution, the proposed robust probabilistic partial least squares (RPPLS) model can overcome the shortcomings of PPLS model. Furthermore, on the basis of RPPLS model, two monitoring indicators GT2 and GSPE are proposed to monitor the process state and the model changes, respectively. Comparing the monitoring performance in the TE process based on PPLS and RPPLS shows that RPPLS is more effective than PPLS in terms of the fault accuracy and the missing alarm rate.

Key words: robust probabilistic partial least squares algorithm, T distribution, parameter estimation, monitoring indices, model

摘要:

概率偏最小二乘(PPLS)模型建立的条件是主元和误差都服从高斯分布,但是高斯分布的期望和方差容易受到离群点的影响,导致模型的鲁棒性较差。针对PPLS模型的不足,提出一种鲁棒概率偏最小二乘(RPPLS)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RPPLS引入过程监控中,提出GT2和GSPE两个监控指标,分别监控过程的受控状态以及模型关系的变化。PPLS和RPPLS在TE过程监控的应用结果表明RPPLS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。

关键词: 鲁棒概率偏最小二乘算法, T分布, 参数估值, 监控指标, 模型

CLC Number: