Predicting transfer learning suitability in ANN-based control of FOPDT industrial processes

The application of Artificial Neural Networks (ANN) in industrial control has become a popular topic of research in recent years. The adoption of strategies showing satisfactory results in other domains, such as Transfer Learning, have been proposed to overcome scarce data limitations. However, there is a lack of studies specifically addressing the requirements of control environments, where applying unsuitable ANN-based controllers can have critical consequences. In this work, we conduct an analysis of Transfer Learning focusing on the control of First-Order plus Dead-Time (FOPDT) processes. In particular, we first provide an overview of state-of-the-art Transfer Suitability Metrics (TSM) along with an analysis of their applicability to control. To do that, we define two transference scenarios that can be found in practice. Based on the insights extracted from the analysis, we propose a novel learning-based metric aimed at estimating the transfer deterioration when applying a data-based controller to a target domain. This metric enables the quantification of transfer suitability, so that a low deterioration value indicates that training a new neural network specifically for this process would yield similar performance. The proposed metric shows a good performance, and a simplified version is also proposed to offer a balanced trade-off between complexity and predictive accuracy. • Comparative analysis of transfer suitability metrics, with special emphasis on regression. • Scenarios defined to characterize TL in industrial control along with the applicability of TSM. • Novel metric to predict transfer deterioration granted by information-theoretic interpretation. • Data-based TSM metrics validated within the FOPDT domain, including a transfer deterioration analysis. • Simplified version of novel metric with good complexity vs. predictive accuracy trade-off.