Novel virtual sample generation method based on data augmentation and weighted interpolation for soft sensing with small data

Data-driven soft sensing modeling plays an increasingly important role in the prediction of key variables in the process industry. Since data is an essential part of modeling, how to obtain sufficient samples to build more accurate soft sensors becomes a formidable challenge. In this paper, a virtual sample generation method based on data augmentation and weighted interpolation (DAWI-VSG) is proposed to expand the soft sensing dataset with high-quality samples. First, the original dataset is decomposed by singular value decomposition (SVD), and the features are extracted and then synthesized into a matrix to obtain new samples that can approximate the original sample set. Second, the two sample sets are merged, and outliers are detected with the improved Fast Angle-based Outlier Detection (FastABOD), which makes the data more uniformly distributed by weighted interpolation between the outliers. In addition, XGboost is utilized for predicting the outputs of the virtual samples. To verify that the effect of the proposed DAWI-VSG, simulations of the numerical function and the actual chemical process Pure terephthalic acid (PTA) were performed, and correlation analysis was introduced as a measure of whether the generated samples are consistent with the real ones. The results showed that the proposed DAWI-VSG can boost the predictive power of soft sensing by generating higher quality and more reasonable samples compared to other advanced methods.