| 聂磊鑫,王海斌,李超,张永霖.高斯过程辅助的船舶辐射噪声分类算法*[J].,2025,44(1):170-181 |
| 高斯过程辅助的船舶辐射噪声分类算法* |
| Gaussian process-assisted classification algorithm for ship-radiated noise |
| 投稿时间：2023-09-08 修订日期：2024-12-31 |
| 中文摘要: |
| 现阶段船舶辐射噪声分类通常使用深度学习方法，它需要大量的数据去训练过参数化的模型。船舶辐射噪声的实测数据量一般较少，因此可以采用数据扩增策略去改善训练数据的多样性。然而相比于实测数据的时频谱图，该策略得到的增广时频谱图通常存在边缘分布偏移，忽视这一点将无法充分利用增广数据，会带来性能上的损失。为此本文提出了一种高斯过程辅助的船舶辐射噪声时频谱图分类算法。该算法在训练分类模型时，除了最小化分类器输出与标签之间的交叉熵损失，也同时借助高斯过程回归，最小化了不同数据上特征提取器输出的分布差异，进而在特征空间实现真实数据时频谱图和增广时频谱图的边缘分布对齐，这可以促进分类器的训练。公开海试数据上的实验结果表明，相较深度学习方法中现有的预训练-微调策略，在不同训练数据量下，本文所提算法都能够对近场船舶辐射噪声实现更准确的分类。 |
| 英文摘要: |
| Now, the classification of ship-radiated noise (SRN) typically employed deep learning approaches, which required much realistic data to train over-parameterized models. Such data was generally limited, so data augmentation can be used to improve the diversity of training data. However, realistic and augmented spectrums often exhibited noticeable distribution shifts. Neglecting them would prevent the utilization of augmented data. The Gaussian process-assisted classification algorithm for spectrums of SRN was proposed in this paper for solving it. In this algorithm, not only the cross-entropy loss between classifier outputs and labels but also the marginal distribution differences of features between augmented and realistic data were minimized, and the latter was implemented by Gaussian process regression. This aligned data distributions in the feature space, thereby facilitating the training of the classifier. Empirical results on open-source sea-trial data showed that the proposed algorithm outperformed the fine-tuning strategy, achieving classification for SRN in the near-field with higher accuracies. |
| DOI：10.11684/j.issn.1000-310X.2025.01.018 |
| 中文关键词: 船舶辐射噪声，时频谱图扩增，域自适应，高斯过程回归 |
| 英文关键词: Ship-radiated noise, spectrum augmentation, domain adaptation, Gaussian process regression |
| 基金项目:国家自然科学基金项目(62171440, 62301551) |
|
| 摘要点击次数: 8 |
| 全文下载次数: 2 |
|
查看全文
查看/发表评论 下载PDF阅读器 |
| 关闭 |
