Machine learning assisted mechanical properties prediction of fine denier polyester fiber

Abstract

In recent years, more and more attention has been attracted by fine denier polyester fibers in electronic information industry due to their unique structural and excellent mechanical properties. As known, the mechanical properties of fine denier polyester fibers have a close relationship with their melt-spinning process parameters, thus, it is worth further exploring the influence of process parameters on fiber mechanical properties (namely breaking strength, elongation at break and CV of breaking strength). This study aims to develop a novel prediction model based on the artificial neural network (ANN) to predict the key mechanical properties of fine denier polyester fibers. For this purpose, firstly, different specifications of fine denier polyester fibers with different mechanical properties are produced by adjusting the key melt-spinning parameters, namely spinning temperature, take-roll speed and metering pump speed. Then the experimental data is collected, analyzed and augmented with effect to be used for modeling. Next, a novel prediction model based on ANN is developed by importing the two-streams structure and two attention matrices to improve the feature abstraction for mechanical properties prediction. Lastly, different predictive performance indicators, namely root mean square error (RMSE), mean absolute percent error (MAPE) and the coefficient of determination (R2) are imported to evaluate the proposed model. For the total of three mechanical properties, the predicted results of the proposed model in terms of RMSE, MAPE and R2 are respectively 0.3499, 0.0906 and 2.9516.