3D inversion techniques for mapping porosity across composite material layers

Abstract

Carbon fibre reinforced polymers (CFRP) are extensively used across various industries, particularly aerospace and other engineering fields, thanks to their exceptional structural and mechanical properties. However, these materials are prone to defects such as delamination, broken fibre, resin cracks, fibre misalignment and void inclusions such as porosity. Among these, porosity remains the most common defect, significantly impairing the performance and durability of CFRP structures by altering their structural integrity and mechanical properties. Traditional non-destructive testing (NDT) methods, including Ultrasonic Testing (UT), Radiographic Testing (RT), Computed Tomography (CT), Thermographic Inspection, and Eddy Current Testing (ECT), are commonly employed to map and characterise porosity. However, these methods often face challenges such as high costs, time consumption, limited spatial coverage, and the inability to comprehensively inspect large structures or capture the full extent of porosity distribution. To address these challenges, the present study investigates the application of Convolutional Neural Networks (CNN) to ultrasound data for mapping porosity across CFRP layers. By leveraging CNN, this research introduces a novel machine learning-based approach to analyse ultrasound imaging data and predict porosity distribution. In this study, the CNN is setup to incorporate the pre-trained VGG-16 architecture and optimised via the OPTUNA framework, is proposed to facilitate fast and accurate porosity assessment in CFRP layers. The CNN is trained and validated on two distinct datasets: simulated ultrasound signals generated by a Finite Element (FE) model, whereas the experimental data is from an ultrasound immersion tank setup. Four samples with different porosity levels across their layers are built and tested in the present work. The CNN model achieved 96.49% training accuracy and 100% validation accuracy on the simulated dataset, while demonstrating 99.66% training accuracy and 99.22% validation accuracy on the experimental dataset. These results highlight the ML model robustness and adaptability in handling both synthetic and experimentally acquired ultrasonic signals for porosity inspection. The CNN consistently identified porosity patterns across different layers, underscoring its efficacy as a non-destructive testing tool with potential for automated data interpretation and streamlined quality control in composite manufacturing processes. Notwithstanding these promising outcomes, this research emphasises the importance of extensive, high-quality datasets to mitigate the risks of overfitting and improve the generalisability of the CNN model. The findings contribute to the broader domain of machine learning (ML)-driven inspection 4 methods, demonstrating that advanced ML algorithms can significantly enhance the detectability and quantification of critical defects in CFRP structures. Consequently, the proposed methodology lays the groundwork for the integration of ML-based non-destructive evaluation systems, with implications for optimising CFRP production, reducing inspection times, and bolstering the reliability of advanced composite components in aerospace, automotive, and other high-performance industries.