| 引用本文: | 刘宁波， 刘新亮， 董云龙， 丁 昊， 关 键， 孙殿星. 基于VAE-WGAN的海杂波幅度-时频联合生成建模[J]. 雷达科学与技术, 2025, 23(5): 473-481.[点击复制] |
| LIU Ningbo, LIU Xinliang, DONG Yunlong, DING Hao, GUAN Jian, SUN Dianxing. Joint Magnitude-Time-Frequency Generative Modeling of Sea Clutter Using VAE-WGAN[J]. Radar Science and Technology, 2025, 23(5): 473-481.[点击复制] |
|
| 摘要: |
| 针对传统统计模型在模拟海杂波时频特性中的局限性，本文提出了一种基于改进生成对抗网络（Generative Adversarial Network，GAN）的海杂波数据生成方法。通过将复数海杂波分解为幅度和时频分量，分别输入变分自编码器-沃瑟斯坦生成对抗网络（Variational Autoencoder-Wasserstein Generative Adversarial Network，VAE-WGAN）进行训练，利用VAE的潜在空间编码和WGAN的稳定对抗训练融合生成复杂幅度分布与时变特性、并兼具幅度与相位特性的复数信号。为增强模型性能，引入梯度惩罚机制约束鉴别器Lipschitz连续性，有效缓解模式崩溃问题；集成自注意力模块强化对海尖峰（Sea Spikes）等局部强散射特征的捕捉能力，显著提升生成信号的时空相关性。实验设计覆盖2~5级海况，每级海况分别构建［64，64］、［128，128］、［256，256］三组数据集，共完成12组交叉验证，结果表明，生成数据在幅度分布、归一化频谱、时间相关性及时频特性上与实测数据高度一致，验证了模型对跨海况场景与变尺度时序数据的泛化能力。 |
| 关键词: 海杂波 生成对抗网络 时频特性 数据增强 |
| DOI：DOI:10.3969/j.issn.1672-2337.2025.05.001 |
| 分类号:TN959.72 |
| 基金项目:国家自然科学基金（No.62388102，62101583）； 泰山学者工程（No.tsqn2002211246） |
|
| Joint Magnitude-Time-Frequency Generative Modeling of Sea Clutter Using VAE-WGAN |
|
LIU Ningbo, LIU Xinliang, DONG Yunlong, DING Hao, GUAN Jian, SUN Dianxing
|
|
1. Institute of Information Fusion, Naval Aviation University, Yantai 264001, China;2. Harbin Engineering University, Harbin 150001, China 211106, China
|
| Abstract: |
| To address the limitations of traditional statistical models in simulating the time-frequency characteristics of sea clutter, a sea clutter data generation method based on an enhanced generative adversarial network （GAN） is proposed in this paper. The complex sea clutter is decomposed into amplitude and time-frequency components, which are separately fed into a variational autoencoder-Wasserstein generative adversarial network （VAE-WGAN） for training. The outputs are then integrated to synthesize complex signals with both amplitude and phase characteristics. To enhance the model performance, a gradient penalty mechanism is introduced to constrain the Lipschitz continuity of the discriminator, effectively mitigating the mode collapse. A self-attention module is incorporated to strengthen the model’s ability to capture localized strong scattering features, such as sea spikes, significantly improving the spatiotemporal correlation of generated signals. Experiments cover sea states 2~5, with three datasets of dimensions ［64,64］, ［128,128］, and ［256,256］ constructed for each sea state. Twelve cross-validation trials demonstrate that the synthetic data exhibit high consistency with measured data in amplitude distribution, normalized spectrum, temporal correlation, and time-frequency characteristics. These results validate the model’s generalization capability across varying sea states and multi-scale temporal scenarios. usually required to face the ship directly to achieve precise strike. Aiming at the challenges faced by traditional SAR imaging method under forward-looking condition, this paper proposes a forward-looking SAR ship target facade imaging method based on polar format algorithm （PFA）. This method cleverly utilizes the three-dimensional characteristics of the ship. Even if the radar operates in the forward-looking mode, although the target cannot be effectively distinguished in the azimuth direction, there is still a change in Doppler frequency in the pitch direction. Therefore, the two-dimensional high-resolution imaging of the ship can be realized in the range and pitch directions. Moreover, this method can provide more intuitive facade images of the ship, which is of great significance for identifying ship types, assessing potential threats, and carrying out precise strikes. Finally, the method is verified through simulation experiments, and clear ship facade images are obtained by using PFA. |
| Key words: sea clutter generative adversarial network time-frequency characteristics data enhancement |
