Search

Data imbalance problem hinders the quality indicators of data-driven soft sensing modeling in the process industry. This problem is caused by production complexity, which makes the pivotal variables can only be manually sampled and detected rather than directly detected by sensors. Aiming at this problem, we propose a feature-constrained encoding generative adversarial prediction network model (FCEGAPN) for quality soft sensing in the cement clinker calcination process, which is constructed by an encoder, a generator, a discriminator, and a predictor. The encoder, generator, and discriminator mine the features of multivariable time series data and produce fake samples under the constructed three-player adversarial learning mechanism. The produced multivariable time series samples are mixed with actual samples for data augmentation, so that the amount of quality indicator is increased, its feature space is enlarged, and the data imbalance problem is eliminated, whose effectiveness is validated by statistics and clustering methods. Using augmented data to train predictors ultimately improves the soft sensing performance of quality indicators. The advantages in accuracy, validity, and robustness of the proposed method are demonstrated by the experiments implemented by cement production data gained from a cement manufacturing enterprise.