Cycle-consistent generative adversarial networks for damage evolution analysis in fiber-reinforced polymers based on synthetic damage states

Abstract

Analyzing computed tomography (CT) scans is challenging and time-consuming due to their high complexity. Machine learning, particularly in the form of segmentation techniques, has emerged as the state-of-the-art approach for defect detection in parts and materials. However, the lack of pixel-accurate labeled training data remains a significant challenge. This paper presents a damage state transformation approach based on a cycle-consistent generative adversarial network (CycleGAN) using fatigue damage states of fiber-reinforced polymers. The generated synthetic data is visually almost indistinguishable from real data. Introduced damage can be determined by calculating the damage removed during the transformation from a high-damage state to a low-damage state. Using multiple transformation steps in detecting and distinguishing different damage states the effectiveness is demonstrated. In addition, the virtual addition of damage to undamaged specimens is investigated. The results show that certain damages exhibit chaotic generation across successive slices while maintaining semantic connections in specific regions across multiple slices. Overall, this research presents a valuable approach for improved self-supervised damage detection and characterization in CT scans, with potential applications in materials analysis and structural health monitoring.