Transfer learning soft sensor for product quality prediction in multi-grade processes

For multi-grade chemical processes, current data-driven soft sensor models built in a specific operating condition cannot be directly applied to predict product qualities of other conditions. A simple transfer learning method namely domain adaptation extreme learning machine (DAELM) is presented to construct a soft sensor model in multi-grade processes with limited labeled data. In addition, an effective strategy is developed to fast select the model parameters. By exploring and utilizing useful information from different operating conditions, the prediction accuracy can be improved. Compared with traditional soft sensors, the prediction results of two multi-grade chemical processes validate the advantages of DAELM.