A data augmentation approach combining time series reconstruction and VAEGAN for improved event recognition in Φ-OTDR

Abstract

This paper introduces a data augmentation method based on Time Series Reconstruction (TSR) and Variational Auto-encoder Generative Adversarial Network (VAEGAN) to address the problem of low event recognition accuracy in Φ-OTDR systems caused by scarce samples. TSR method generates new feature data by performing a temporal domain transformation on the Mel spectrograms and the VAEGAN network is utilized to augment the background information. The TSR&VAEGAN can greatly improve the data diversity while keep the feature authenticity. Experiment results show that the proposed approach can improve the classification accuracy of minor class from 88% to 94% when only 10 real minor samples are applied. This method can effectively enhance the event recognition capability of Φ-OTDR systems in scenarios with limited samples.