Crack detection in laser-welded aluminum alloy based on the integration of generative adversarial networks

Abstract

Monitoring weld quality in high-speed laser welding is crucial due to the complex dynamics of defect formation. Temperature-based sensors, such as infrared (IR) cameras and pyrometers, provide valuable insights into crack formation by capturing temperature distributions. However, these sensors face limitations in resolution and frequency, particularly under high-speed conditions. This study addresses these challenges by integrating a deep-learning model based on generative adversarial networks (GANs) for video frame interpolation (VFI), enhancing both resolution and frequency. This enables precise temporal synchronization between high-speed and IR camera data, facilitating robust, high-resolution crack detection. The developed CNN model effectively predicts defect occurrences in self-restraint crack test specimens of 6014-T4 aluminum during laser oscillation welding, demonstrating the feasibility of using GANs to augment input data and generate high-quality synthetic images. Both IR and high-speed camera images captured essential crack characteristics, while VFI interpolation enhanced the frame rate to 2000 fps, achieving an average peak signal-to-noise ratio (PSNR) of 39.01 dB. Confusion matrix analysis revealed high prediction accuracy, exceeding 99% across all models. The study concludes that GANs can identify significant data regions to support real-time crack detection in high-speed laser welding, with optimal pixel-to-image ratios proposed based on experimental findings.