Multi-task learning with state propagation for quality forecasts in polymer extrusion lines

Abstract

Modern polymer extrusion processes usually consist of several process steps, exhibit dynamic nonlinear behavior and depend on many external factors, for instance ambient conditions. Consequently, disturbances cannot be anticipated and extrusion lines are usually operated in a reactive manner leading to process fluctuations, quality deviations and increased material consumption. Unfortunately, existing approaches are either focused on single stage solutions which do not consider the process structure or rely on static models neglecting the dynamic behavior. Therefore, we propose a multi-task learning framework of individual (nonlinear) autoregressive state-models with exogenous inputs connected via a causality graph and jointly trained via physics-informed state propagation. Additionally, training the model on an extended forecasting horizon allows to capture and forecast the dynamic process behavior for a prolonged period. Moreover, the modular architecture can be adapted for any continuous production process with sequential and parallel material streams as well as feedback loops of existing controllers. The proposed method is validated with data from an actual pipe extrusion line and compared to end-to-end black-box models. The results show superior performance with accurate forecasts up to 120 min with a mean absolute percentage error of 0.09% for the pipe diameter representing a reduction of 13.9% to 22.3% compared to different baseline models. Furthermore, the effect of spurious correlations is mitigated through enforced physical dependencies improving the robustness of the resulting model. This allows to change the operating mode from reactive to proactive and to reduce the need for trial and error optimizations which leads to increased efficiency in complex extrusion lines.