A Novel In-Line Polymer Melt Viscosity Sensing System of Integrated Soft Sensor and Machine Learning

Real-time monitoring of melt viscosity is one of the many challenges faced in the polymer extrusion process. Viscosity is one of the important metric reflecting material properties during the plastic process. However, viscosity is an indicator that can be evaluated by calculating the evaluation value through temperature, pressure, screw rotation speed, and so on but cannot be directly measured by physical sensors. Melt viscosity should be calculated at the exact melt temperature. Temperature sensors cannot accurately measure the temperature of the melt in the barrel. Soft sensing technique is the best solution for estimating material properties. It just needs some physical sensors and physical formulas. The proposed viscosity soft sensor consists of physical sensors, a temperature estimator, and a simulation analysis software for calculating material properties. Four physical temperature signals, one physical pressure signal, and the simulation properties data are used as the dataset for the temperature estimator. An ensemble machine learning model of temperature estimators consists of random forests (RFs) and convolutional neural networks (CNNs). Melt viscosity, shear stress, and shear rate are calculated by physical formulas. Experimental results show that the proposed temperature estimator predicts that the temperature will reduce the mean absolute error (MAE) from 6.08 to 2.86. Compared with the current work, the prediction error rate of the soft sensor is also reduced from 4% to 1.1%. The proposed soft sensor can be used to better predict polymer melt temperature. Finally, according to the predicting results, the material property scatter chart of viscosity can be precisely plotted under the specific melt temperature. In the past, the melt viscosity could only be measured offline. The proposed method can be added as a plug-in to the existing process to achieve real-time viscosity monitoring.