Multi-output deep learning-based soft sensor for real-time radial melt temperature profile prediction in polymer extrusion processes

The cross-sectional temperature distribution of the polymer melt is a primary quality indicator in polymer extrusion processes. The radial melt temperature profile near the die entry provides valuable information about the thermal homogeneity of the molten polymer. The existing physical sensors used for melt temperature profile measurement have limitations such as causing disturbances to the melt flow and poor durability due to their invasive nature. Consequently, they fail to continuously monitor the radial melt temperature distribution, inhibiting real-time quality control strategies. To address these limitations, this study proposes a non-invasive measurement technique based on a soft sensor. The soft sensor inferentially estimates the radial melt temperature profile in real-time, using readily measured extrusion parameters as inputs. A multi-output, bidirectional long short-term memory neural network, combined with an autoencoder, is employed as the mathematical model of the soft sensor. It provides real-time predictions of melt temperatures at 13 radial positions across the melt flow. This is the first deep learning-based soft sensor designed to predict the radial melt temperature profile in polymer extrusion processes. The soft sensor exhibited excellent predictive performance with a root mean square percentage error of 0.0525% on test data. Furthermore, it reported a 94–99% improvement in predictive performance at all radial positions (in terms of the normalised prediction error on test data) compared to the only other soft sensor reported in the literature for the same task, which was based on a traditional polynomial model. This novel soft sensor should be invaluable in advancing both process monitoring and control in polymer extruders.