Domain adaptation transfer learning soft sensor for product quality prediction

Abstract For multi-grade chemical processes, often, limited labeled data are available, resulting in an insufficient construction of reliable soft sensors for several modes. Additionally, the current soft sensors built in a specific mode cannot be directly extended to accurately predict the product qualities of other modes. In this paper, inspired by the idea of transfer learning, a domain adaptation extreme learning machine (DAELM) is developed to establish a simple soft sensor model suitable for multi-grade processes with limited labeled data. Additionally, an efficient model selection strategy is developed to select its model parameters. By utilizing and transferring the useful information from different operating conditions to the existing soft sensor, the prediction domain is enlarged and the prediction accuracy is enhanced. The prediction results of two multi-grade chemical processes demonstrate the advantages of DAELM as compared to the current popular soft sensors (e.g., extreme learning machine).