A novel adaptive algorithm with near-infrared spectroscopy and its application in online gasoline blending processes

• A novel adaptive algorithm is designed to establish and maintain the NIR model.
• A new outlier detection method is used to remove the outliers during modeling process, both online and offline.
• The improved recursive strategy is applied to update NIR model with new available sample.
• The locally weighted strategy is used to update the weights of samples in an NIR model.

Near-infrared (NIR) spectroscopy has been widely used to estimate product qualities. Although numerous studies on NIR modeling methods have been conducted, few papers have reported the online application of an NIR spectrometer in the gasoline blending process. This study presents a novel adaptive modeling method to establish an NIR model for the gasoline blending process. This method is based on the local learning and recursive modeling framework. Based on the framework, the proposed method can adjust the model structure from two aspects: (i) in sampling intervals, the model is updated with a local learning strategy, and the weights of the training samples can be gradually adjusted; and (ii) when new reference samples become available, the new data pairs are selected and added to the training data set based on an effective evaluation mechanism. The high performance of the proposed algorithm was demonstrated through a spectroscopic data set from a real gasoline blending process. The research octane number (RON), as the most important properties of gasoline, was estimated. Several modeling methods such as recursive partial least squares (RPLS), partial least squares (PLS), and locally weighted PLS were utilized for comparison. The results show that the proposed approach produce more accurate results than the traditional RPLS and locally weighted PLS algorithms.