Deep Learning-Assisted Digital Image Techniques for In-situ Damage Characterization of Non-crimp Fabric Reinforced Reactive Thermoplastic Composites

Abstract

Infusible unidirectional non-crimp fabric (UD-NCF) glass fiber-reinforced reactive thermoplastic composites have emerged as promising candidates for primary load-bearing structures such as wind turbine blades due to their excellent mechanical performance, recyclability, and compatibility with liquid composite molding (LCM) processes. Their application in extreme service environments, including cyclic loading and low temperatures down to -50°C, places significant demands on structural integrity. These conditions highlight the need for a comprehensive understanding of the mechanical response, damage evolution, and failure mechanisms at -50°C to ensure long-term durability and reliability. Despite growing interest in thermoplastic composite systems, a critical gap remains in the literature, specifically the limited understanding of the damage behavior and failure characteristics under such demanding conditions. This research aims to develop an automated, AI-driven, in-situ non-destructive digital imaging technique for damage characterization of an infusible UD-NCF glass fiber-reinforced reactive thermoplastic acrylic (NCF-GF/acrylic) composite material. The study is structured around four main objectives, each addressed through a dedicated research task. In Task 1, the mechanical performance and failure characteristics of NCF-GF/acrylic were investigated and compared to epoxy-based counterparts (NCF-GF/epoxy) at room (22°C) and low (–50°C) temperatures. Both materials exhibited increased stiffness and strength at LT under longitudinal, transverse, and in-plane shear loading conditions; however, NCF-GF/acrylic showed distinct damage features, including greater supporting fiber yarn splitting and breakage, reduced tow fiber bridging, and a transition from ductile to brittle behavior. This task provided key insights into the damage and failure mechanisms of NCF-GF/acrylic and demonstrated its potential as a viable alternative to epoxy-based systems for wind energy applications in cold climates. An in-situ digital imaging technique was developed in Task 2 to characterize damage in NCF-GF/acrylic cross-ply laminates subjected to tensile loading at LT. A custom algorithm was developed to automatically detect the initiation and growth of 90° tow cracks, matrix cracks, and 0° tow cracks through image stacking, shift-correction, and thresholding. The laminates exhibited four stages of deformation/damage, including linear elastic, onset/growth of 90° fiber tow cracks, onset/growth of 0° fiber tow cracks, and progressive failure of 0° fiber tows. Although at LT, the effective laminate strength and stiffness increased by 4% and 13%, respectively, damage initiated sooner and propagated at a higher rate leading to a 60% increase in crack density at saturation. The digital imaging technique proved to effectively detect local damage in the glass fiber/thermoplastic laminates, which led to a deeper understanding of their low-temperature deformation response, damage characteristics, and damage tolerance. Through Task 3, a novel integrated digital imaging technique was developed to automatically characterize damage and simultaneously monitor stiffness degradation in NCF-GF/acrylic cross-ply laminates subject to tension-tension cyclic loading. Under a peak stress of 50% UTS, the laminate exhibited four distinct stages of stiffness degradation, including a significant drop during the first cycle due to 90⁰ fiber tow crack initiation, a gradual decrease due to crack multiplication and propagation up to saturation, a stable phase comprising localized delamination crack propagation, and a sudden drop prior to specimen failure. Under a peak stress of 75% UTS, both 0⁰ and 90⁰ fiber tow cracks developed during the first stage followed by an accelerated growth rate and reduced 90⁰ tow crack density at saturation. The integrated digital imaging technique proved to effectively correlate damage events with stiffness degradation, leading to a deeper understanding of the fatigue behavior of glass fiber/thermoplastic composites. A deep learning based convolutional neural network was integrated with digital imaging technique to characterize damage in NCF-GF/acrylic composite laminates in Task 4. Images from prior experimental studies were preprocessed and annotated with six feature-related classes, which were subsequently divided into training and testing sets. A U-Net architecture incorporating skip connections, dropout, and weighted Dice loss function was employed. Results showed that smaller batch sizes achieved the lowest weighted Dice loss and the highest Dice coefficients. Notably, the Dice coefficient improved significantly from ~0.736 for balanced class weights to ~0.840 for highly biased class weights, representing a 14.1% increase in the prediction accuracy of 90⁰ tow cracks. Early stopping at appropriate epochs effectively minimized overfitting while optimizing accuracy. The model demonstrated strong capability in identifying damage initiation, progression, and interaction, and was seamlessly integrated with automated crack counting algorithms for quantitative damage assessment. Furthermore, the dataset provides a robust foundation for future research on advanced deep learning methodologies for damage characterization in non-crimp fabric composites.