Hybrid data augmentation and VGGish-BiLSTM-Attention model fusion for small-sample φ-OTDR event recognition

To address the poor generalization of phase-sensitive optical time domain reflectometry ( φ -OTDR) systems caused by the scarcity of critical vibration event samples, this paper proposes a few-shot event recognition approach named VGGish-BiLSTM-Attention with hybrid augmentation (VBA-HA). The framework first applies a hybrid data augmentation strategy that combines basic augmentations, including noise injection and time stretching, with generative augmentation based on a conditional generative adversarial network. This strategy effectively expands the training set and enhances sample diversity at low cost. A serially fused VGGish-BiLSTM-Attention (VBA) model is then constructed. The model leverages a pre-trained VGGish network to extract deep features from Mel spectrograms, employs a bidirectional long short-term memory (BiLSTM) network to capture forward and backward temporal dependencies in vibration signals, and incorporates an attention mechanism to adaptively focus on discriminative feature segments critical for classification. Experimental results on a public φ -OTDR dataset show that under a few-shot condition with only 70 training samples per class, the proposed method achieves an average recognition accuracy of 96.2% and an F1-score of 96.1%, significantly outperforming conventional methods with the same amount of data. By deeply integrating data augmentation and model innovation, this study provides a data-efficient solution for few-shot φ -OTDR event recognition and substantially reduces the reliance of such systems on large-scale labeled datasets.