Deep Learning Approach for Predicting the Physical Properties of Air-Jet Textured Yarn with PET/PTT Bicomponent Fiber

Accurate prediction of air-jet textured yarn (ATY) properties is essential for product development and quality control in textile manufacturing. This study proposes a deep learning-based regression model to predict the properties of ATY produced with PET/PTT bicomponent fiber as the core yarn. The model’s performance was compared with traditional statistical regression methods. The dataset included key process parameters—denier, overfeed, air pressure, and processing speed—and their corresponding physical properties: tenacity, initial modulus, length instability, and loop density gap. Hyperparameter tuning, regularization, and K-fold cross-validation were employed to enhance model performance and reduce overfitting. Mean absolute error convergence analysis was also used to determine the optimal number of training epochs. Results showed that the multilayer perceptron deep learning model consistently outperformed statistical regression models, achieving R² values above 0.7 for initial modulus, length instability, and loop density gap. Prediction for tenacity showed limited improvement due to weak feature-property correlations. Importantly, the deep learning model improved predictive accuracy for the loop density gap, a property with minimal linear correlation to process variables, though with higher variance—indicating that additional data could improve stability. These findings demonstrate the potential of deep learning as a powerful tool for predicting complex material properties in textile processes, particularly when dealing with nonlinear relationships among multiple process factors.