Near-infrared spectroscopy for the concurrent quality prediction and status monitoring of gasoline blending. (August 2020)

• Record Type: Journal Article
• Authors: He, Kaixun; Zhong, Maiying; Li, Zhi; Liu, Jingjing
• Is Part Of: Control engineering practice. Volume 101(2020)
• Publication Date: 2020-08
• Subjects: Soft sensor -- Training sample selection -- Gaussian mixture model -- Semisupervised -- Gasoline blending; Automatic control -- Periodicals; 629.89
• DOI: 10.1016/j.conengprac.2020.104478
• Languages: English
• ISSNs: 0967-0661

Abstract

Abstract: Gasoline is one of the major products of oil and petrochemical industry. Blending is the final step and key to improve the efficiency of gasoline production. As an important property that can reflect the quality of gasoline products, the research octane number (RON) of final products has been widely used to evaluate the running status of gasoline blending. However, the real-time and direct acquisition of RON from this process is difficult because it have to be determined by running the fuel in a test engine with a variable compression ratio under controlled conditions. This work proposes a data-driven soft sensor based on near-infrared (NIR) spectroscopy for online RON estimation. A modified semisupervised Gaussian mixture algorithm is adopted to automatically discover meaningful modeling samples and initialize the quality prediction model. Besides, a monitoring model is integrated into the quality prediction sensor to monitor the running status and the accuracy of the NIR-based quality prediction sensor. Datasets from a numerical experiment and industrial gasoline blending are provided to reveal the effectiveness and superiority of the proposed method. Highlights: An effective probability-based method is proposed to discover meaningful modeling data for soft sensor. A two-step semi-supervised GMM model is proposed. The training process with labeled and unlabeled samples are independent of each other. A monitoring model is embedded in prediction model which can lead the adaptive updating of the prediction model. A real-world gasoline blending system is provided to illustrate the effectiveness of our proposed method.