SCI和EI收录∣中国化工学会会刊

›› 2016, Vol. 24 ›› Issue (8): 1038-1046.DOI: 10.1016/j.cjche.2016.05.015

• Process Systems Engineering and Process Safety • Previous Articles     Next Articles

Calibration of soft sensor by using Just-in-time modeling and AdaBoost learning method

Huan Min, Xionglin Luo   

  1. Department of Automation, China University of Petroleum, Beijing 102249, China
  • Received:2016-01-05 Revised:2016-04-18 Online:2016-09-21 Published:2016-08-28
  • Supported by:
    Supported by the National Basic Research Program of China (2012CB720500).

Calibration of soft sensor by using Just-in-time modeling and AdaBoost learning method

Huan Min, Xionglin Luo   

  1. Department of Automation, China University of Petroleum, Beijing 102249, China
  • 通讯作者: Xionglin Luo
  • 基金资助:
    Supported by the National Basic Research Program of China (2012CB720500).

Abstract: Soft sensor is an efficacious solution to predict the hard-to-measure target variable by using the process variables. In practical application scenarios, however, the feedback cycle of target variable is usually larger than that of the process variables, which causes the deficiency of prediction errors. Consequently soft sensor cannot be calibrated timely and deteriorates. We proposed a soft sensor calibration method by using Just-in-time modeling and AdaBoost learning method. A moving window consisting of a primary part and a secondary part is constructed. The primary part is made of history data from certain number of constant feedback cycles of target variable and the secondary part includes some coarse target values estimated initially by Just-in-time modeling during the latest feedback cycle of target variable. The data set of the whole moving window is processed by AdaBoost learning method to build an auxiliary estimation model and then target variable values of the latest corresponding feedback cycle are reestimated. Finally the soft sensor model is calibrated by using the reestimated target variable values when the target feedback is unavailable; otherwise using the feedback value. The feasibility and effectiveness of the proposed calibration method is tested and verified through a series of comparative experiments on a pH neutralization facility in our laboratory.

Key words: Process control, Measurement, Soft sensor, Calibration, Deterioration, Moving window, Just-in-time, AdaBoost

摘要: Soft sensor is an efficacious solution to predict the hard-to-measure target variable by using the process variables. In practical application scenarios, however, the feedback cycle of target variable is usually larger than that of the process variables, which causes the deficiency of prediction errors. Consequently soft sensor cannot be calibrated timely and deteriorates. We proposed a soft sensor calibration method by using Just-in-time modeling and AdaBoost learning method. A moving window consisting of a primary part and a secondary part is constructed. The primary part is made of history data from certain number of constant feedback cycles of target variable and the secondary part includes some coarse target values estimated initially by Just-in-time modeling during the latest feedback cycle of target variable. The data set of the whole moving window is processed by AdaBoost learning method to build an auxiliary estimation model and then target variable values of the latest corresponding feedback cycle are reestimated. Finally the soft sensor model is calibrated by using the reestimated target variable values when the target feedback is unavailable; otherwise using the feedback value. The feasibility and effectiveness of the proposed calibration method is tested and verified through a series of comparative experiments on a pH neutralization facility in our laboratory.

关键词: Process control, Measurement, Soft sensor, Calibration, Deterioration, Moving window, Just-in-time, AdaBoost