Modeling extruder’s behavior as multivariate time series for deep learning-driven forecasting

Abstract

he concept of a digital twin (DT) plays a crucial role in the Recyplast-Demo research project, aiming to study the extrusion process of recycled plastic. A digital twin refers to the creation of a virtual and dynamic representation of a real system, in this case, the extrusion machine. The process involves feeding plastic pellets into the machine, which are then propelled forward by a motor-driven screw inside a heated barrel. The pellets melt and transform into molten plastic as they pass through a shaping die, ultimately acquiring their final form. Currently, the motor speed and heater temperature settings are set to constant values throughout the extrusion operations. It is assumed that the properties of the plastic pellets remain stable, leading to consistent setpoints once adjusted by machine operators. Those operators mainly monitor material output pressure, temperature, and motor torque to ensure optimal extrusion performance. However, the use of recycled plastic introduces variability in material properties, challenging the assumption of stability and constant setpoints. Therefore, it becomes crucial to dynamically adapt the motor speed and heater temperature based on factors such as material output pressure, temperature, and motor torque. This is where the digital twin becomes valuable, as it provides real-time insights into the behavior of the plastic extruder. By understanding the extruder’s real-time behavior, we can maintain consistent product quality despite variations in input material properties. Additionally, the digital twin enables accurate evaluation of the extruder’s performance under different conditions and contexts. In this study, we have conducted an analysis of the extruder’s behavior, treating the measurements as multivariate time series data. This analysis allowed us to gain a deeper understanding of the extruder’s behavior and develop an intelligent model for forecasting the extrusion process. Moreover, based on this model, a recommendation system can be developed to provide valuable insights and suggestions in the next works.