Soft Sensor Development Based on Unsupervised Dynamic Weighted Domain Adaptation for Quality Prediction of Batch Processes

Batch processes often encounter insufficiency of modeling data and differences of data distributions between different tanks and batches, which pose great challenges for accurate data-driven soft sensor modeling. Transfer learning has become a promising solution for addressing such problems; however, the existing transfer learning soft sensors still have problems, such as not considering the quality of source domain samples, insufficient utilization of hidden feature information, and only considering single distribution alignment. To address these issues, a dynamic weighted multilayer domain adaptation network (DWMDAN) is proposed for online estimation of key variables in batch processes. The training of the model consists of two stages: Stage I, the feature extractor and the first domain discriminator are utilized for adversarial training to compute the importance weights of the source domain samples, and the whole model is updated by combining weighted regression with global domain adaptation. Stage II, a hybrid domain adaptation mechanism is further designed to fine-tune the target model based on importance-weighted information. Specifically, adversarial and distribution domain adaptation methods are combined to adapt the marginal and conditional distributions between domains at multiple network layers of the model, where an adaptive factor is introduced to dynamically adjust the distribution weights. The superiority and effectiveness of the proposed method are demonstrated by an industrial fed-batch chlortetracycline (CTC) fermentation process. The experimental results show that the proposed method can extract richer and more reliable domain-invariant features for tank-to-tank knowledge transfer, thus delivering more accurate quality predictions.