Combining fundamental knowledge and latent variable techniques to transfer process monitoring models between plants

In this paper we explore the issue of the transfer of process monitoring models between different plants that exploit the same manufacturing process to manufacture the same product. Given a source plant A and a target plant B, the objective is to use the data available from plant A to monitor the operation of plant B, until a sufficient amount of data entirely representative of the operation in plant B is collected to allow building a process monitoring model based on these data only.Two different model transfer methodologies are proposed, which depend on the nature of the measured process variables (namely, on whether they are common between the two plants or not). Both the proposed approaches combine fundamental engineering knowledge on the system (derived from mass or energy balances) with latent variable modeling techniques (namely, principal component analysis and joint-Y partial least-squares regression). Both approaches are based on adaptive algorithms, which make them practical for online use, and are tested on a benchmark problem related to the scale-up of the monitoring model for an industrial spray-drying process. Results show that both proposed procedures provide robust and prompt fault detection, even when very few data are available from plant B.

► Two approaches are proposed to transfer process monitoring models between plants. ► The approaches combine fundamental process knowledge with latent variable methods. ► Data available from a source plant are exploited to monitor a target plant. ► An application to the scale-up of an industrial spray-drying process is presented. ► Reliable fault detection is achieved even when very few samples are available.