Weighted incremental minimax probability machine-based method for quality prediction in gasoline blending process

Near-infrared (NIR) spectroscopy is frequently used to predict quality-relevant variables that are difficult to measure online. This technology can be applied by developing the NIR model in advance. Obtaining a high-accuracy NIR model is difficult using traditional modeling methods because process data inherently contain uncertainties and present strong non-Gaussian characteristics. Considering the difficulty in obtaining precise prediction results, biased estimation is important in producing qualified products when NIR spectroscopy is used in a feedback quality control system. The present work proposes a biased estimation model based on probabilistic representation to address the aforementioned issues. Additionally, a novel weighted incremental strategy with just-in-time learning is proposed to improve model adaptiveness. In this way, the NIR model could be established and maintained without imposing any distribution hypothesis on process data, and biased estimation could be obtained in the form of probability. The performance of the proposed method is demonstrated on an actual data set from a gasoline blending process.