CycleGAN-Based Data Augmentation for Enhancing Classification Accuracy in Φ-OTDR Systems

This paper proposes a data augmentation method based on CycleGAN to address the problem of classification accuracy being affected by noise and data scarcity in
$\Phi$
-OTDR systems. The method converts
$\Phi$
-OTDR signals into Mel-spectrograms, uses CycleGAN to generate new samples for scarce data categories, and then trains an AlexNet classifier on the augmented dataset. Experiments show this approach can effectively improve classification performance. Additionally, the CycleGAN-based style transfer helps retain original features, providing a new solution for Φ-OTDR's data challenges.