Combining Virtual Sample Generation Based Data Enhancement and Multi-objective Optimization Based Selective Ensemble for Soft Sensor Modeling

Abstract

Soft sensor modeling technology realizes the real-time estimation of difficult-to-measure variables by constructing the mathematical model between secondary variables and primary variable. Nevertheless, sufficient and high-quality training samples are difficult to obtain owing to the high cost of data acquisition and low sampling rate. To solve this, a soft sensor modeling method, combining virtual sample generation based data enhancement and multi-objective optimization based selective ensemble (DESE), is proposed. First, a supervised variational autoencoder (SVAE) is constructed by introducing quality variable. Second, a generative model is built through the combination of SVAE and Wasserstein GAN with gradient penalty (WGAN-gp). Third, SV-WGANgp is trained on each sample subset, which is obtained by resampling, and a fixed number of virtual samples are generated. A set of base models is established for the expanded original samples subsequently. Finally, the multi-objective optimization method is utilized to prune these models, which satisfy both accuracy and diversity requirements. After integrating the selected base models, the final prediction results are obtained. Experimental results verify that, compared with the other three popular generation models, DESE significantly improves the prediction performance of soft sensor model by supplementing the original samples.