A Soft Sensor Method based on Unsupervised Multi-layer Domain Adaptation for Batch Processes

In batch processes, soft sensors frequently face the problem of irregular distributions between current and past data owing to variations in operating circumstances, as well as and poor model performing resulting from the absence with labels in the current data. This paper proposes a soft sensor method that is founded on dynamic multi-layer domain adaptation (DMDA). The method being proposed first training a convolutional neural network model with a substantial quantity of labeled data in the source domain, and subsequently use the obtained parameters as the beginning stage for the target model. Then, by utilizing multi-kernel maximum mean discrepancy (MK-MMD) and conditional embedding operator discrepancy (CEOD), the multi-layer convolutional neural network can effectively measure the difference in the overall (marginal) and specific (conditional) distributions between the source and target domains. Furthermore, the incorporation of an adaptive factor facilitates the dynamic adjustment of distribution weight, enabling precise fine-tuning of the target model. Finally, a regression model is established using the distribution-adapted historical data to achieve unsupervised soft sensor modeling. The substrate concentration in different fermentation tanks of the chlortetracycline fermentation process can be predicted through the utilization of the proposed approach. The experimental findings indicate that this method can accomplish tank-to-tank knowledge transfer, and significantly outperform traditional transfer learning-based soft sensor methods.