Data Augmentation and Application of DAS Based on Generative Adversarial Networks

Distributed fiber optic acoustic sensing system (DAS) shows great potential in perimeter security with its high accuracy, high sensitivity and wide range monitoring capability. Aiming at the data scarcity problem of DAS in perimeter security caused by the low probability of acquiring abnormal event samples and the interference of environmental noise, a data augmentation approach for LMA-DRAGAN utilizing generative adversarial network is proposed. The optimised LSKblock module and the improved MobileViT attention mechanism are effectively combined to achieve dynamic sampling of feature maps, which effectively retains the integrity of the original information. And an adaptive learning rate adjustment strategy is employed to balance the parameter update frequency of the generator and the discriminator, thereby helping to stabilize the training process. The study demonstrates that on the dataset containing five types of perimeter security events, the method proposed achieves MAE value of 0.1658, CS value of 0.9542, RMSE value of 0.2260, and IS value of 2.5112, which provides higher diversity while maintaining the realism compared to the conventional augmentation methods. Validated by AlexNet classification model, the recognition accuracy of the enhanced dataset reaches 97.97%, which is a 19.53% improvement over the unenhanced baseline. This study provides a reliable enhancement strategy for DAS perimeter security under small sample conditions, which is conducive to its wider engineering applications.