Prediction of High Velocity Impact Damage Behavior of 3D Angle‐Interlock Woven Fabrics Based on Deep Learning

ABSTRACT 3D angle‐interlock woven fabrics (3DAWFs) have been extensively applied in aerospace and other industries because of their outstanding mechanical characteristics; thus, the investigation of their damage mechanism under impact loading has always received much attention. However, the high costs associated with experiments and the limited computational efficiency of finite element (FE) analysis have hindered the effective characterization of damage behaviors across varying parameter configurations. To address these challenges, this study presents two deep learning models. The first is a convolutional neural network (CNN) featuring an encoder‐decoder architecture enhanced by a center bias mechanism, which consists of a spatial attention module with a Gaussian enhancement mask. This model, utilizing a dataset derived from FE simulations, accurately predicts damage images of 3DAWFs, as well as deformation images that illustrate the energy absorption mechanism through the input of projectile impact angle, initial velocity, and warp/weft direction, with its mean squared error (MSE) ranging from 6.73 × 10 ⁻³ to 9.86 × 10 ⁻³ . The second model is a multi‐layer perceptron (MLP) that quantitatively predicts the projectile residual velocity based on the aforementioned input features, achieving a coefficient of determination ( R ² ) of 0.9904 and a mean absolute percentage error (MAPE) of 3.72%, and explains the influence of each feature via SHapley Additive exPlanations (SHAP) interpretability analysis. Together, the CNN and MLP models successfully predict damage morphology and mechanical properties, enabling real‐time monitoring of the impact damage behavior of 3DAWFs and offering a novel approach for a comprehensive study on the dynamic damage behavior.