A Deep Learning Soft Sensor Based on Domain-Invariant Features Extraction and Online Local Adaptation

Data-driven approaches, especially those based on deep learning, have received much attention in quality prediction in the modern process industry. However, in practice, industrial data are often generated in the form of data streams. As operating conditions and process parameters change with the environment, industrial data are dynamic, resulting in degraded model prediction performance. Although the method based on deep transfer learning can automatically extract domain-invariant features in various tasks to adapt to data distribution changes. However, current deep transfer learning methods cannot handle online data streams with dynamic distribution change characteristics. To solve this problem, a deep learning soft sensor method based on domain-invariant features and online local adaptation (DIF-OLA) is proposed. The implementation of the method consists of an offline phase and an online phase. In the offline phase, domain-invariant features are extracted to build a domain-adaption model. In the online phase, a local fine-tuned update strategy is introduced to make the prediction model better adapted to the new data streams. The effectiveness of the method is verified by the dataset of an actual industrial debutanizer column process. The experimental results indicate that compared to traditional adaptation soft sensor methods, the model has improved performance metrics of MAE, RMSE and R2 by 41.14%,40.94%, and 30.94% respectively.