Balancing Novelty and Realism: A Convolutional Conditional GAN for Underwater Acoustic Channel Generation
Year: 2026
Type: conference-paper
Abstract: Underwater acoustic channels are difficult to model accurately due to their complex physical characteristics. Traditional physical-statistical hybrid models rely heavily on hard-toobtain environmental parameters and have limited generalization capabilities across different scenarios. Although existing datadriven generative methods can fit channel distributions in specific scenarios, they still have significant shortcomings in balancing physical realism with novelty in generative exploration. To address this, this paper proposes a generative model for underwater acoustic channel impulse responses based on a convolutional conditional generative adversarial network (WGAN-GP). The model uses macro-scale meteorological and hydrological parameters as conditional constraints and employs Wasserstein distance combined with gradient penalization to effectively avoid vanishing gradients and mode collapse, significantly improving training stability in high-dimensional feature spaces. Additionally, the model incorporates Feature-wise Linear Modulation (FiLM) conditional modulation and a projection discrimination mechanism, which substantially enhances the control and expressive flexibility of environmental parameters over multipath temporal features. This paper employs a five-dimensional evaluation framework encompassing representativeness, authenticity, diversity, novelty, and consistency to comprehensively assess the performance of the generative model from five perspectives: macroscopic statistics, microscopic fidelity, sample coverage, out-of-sample generation capability, and internal structural coherence. Experiments on the CIR dataset, which consists of multi-seasonal experiments in shallow water, demonstrate that the proposed model not only maintains extremely high physical fidelity in terms of timefrequency structure but also exhibits excellent scene generalization and novel sample generation capabilities. It truly achieves a balance between novelty and realism, providing effective simulation support for the design of underwater communication systems.
Authors: Ruolan Hao, Miaomiao He, Yujia Liu, Li Wei
Institution: Harbin Engineering University
Language: English
Subfield: Ocean Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7201836655
