Data Enhancement of Underwater High-Speed Vehicle Echo Signals Based on Improved Generative Adversarial Networks

Abstract

Underwater target recognition is currently one of the hottest topics in computational intelligence research. However, underwater target recognition tasks based on deep learning techniques are difficult to conduct due to the shortage of acoustic echo signal samples, which results in poor training performance for existing deep learning models. Generative adversarial networks (GANs) have been widely used in data enhancement and image generation, providing a novel strategy for dealing with challenges in the research field mentioned above. To address the insufficiency of echo signal data for underwater high-speed vehicles, this paper proposes an underwater echo signal data enhancement method that uses an improved GAN based on convolution units for small sample sizes. First, we take pool test data as the training sample input and carry out data standardization, data interception, and copy-related processing work. Secondly, this paper proposes an improved generative adversarial network underwater (IGAN-UW) model to generate underwater echo signals. Finally, a CNN model combines the generated data with the original data to conduct classification training for underwater targets. Experimental results show that the IGAN-UW model is suitable for the generation of highly realistic original echo signals in cases with small sample sizes, providing a new approach to the active detection and recognition of underwater targets.