Data Augmentation Method Based on Improved GAN for Underwater Acoustic Signals

Abstract

Underwater acoustic target recognition based on deep learning encounters significant constraints due to the scarcity of labeled samples, which often leads to model overfitting and diminished generalization performance. To tackle this critical challenge, this study introduces a novel data augmentation framework grounded in an enhanced Generative Adversarial Network (GAN), specifically designed for LOFAR spectrograms. Initially, the original time-domain underwater acoustic signals are transformed into LOFAR spectrograms through STFT and undergo targeted preprocessing to enhance feature distinctness. Subsequently, an improved conditional WGAN-GP model is devised, integrating a U-Net generator to preserve fine-grained signal characteristics and a multi-scale discriminator to comprehensively evaluate the authenticity of generated samples. To ensure the reliability of the augmented data, an independent classification network is employed to monitor and filter the generated samples. Experimental evaluations utilizing AlexNet, VGG16, and ResNet18 demonstrate the efficacy of the proposed method: the framework generates high-fidelity LOFAR spectrograms, effectively expands the dataset size, and substantially enhances the classification accuracy and generalization capability of deep learning models. This work provides a practical and effective solution to the challenge of insufficient labeled underwater acoustic samples in target recognition tasks.