随着海洋强国战略的提出和科学技术的飞速发展革新，海洋研究在国家战略层面的重要性日益提升。水下声源定位作为海洋与水声工程领域的关键问题，一直都受到研究人员的关注。

匹配场处理（Matched Field Processing, MFP）由于环境失配等原因会造成一定的定位误差。而利用传统的深度学习方法进行水下声源定位往往需要大量的带标签数据，在某些海洋环境中或者某些定位场景下，无法获得足够大量的数据。另外新实验场景下海洋环境参数发生改变，导致基于原有海洋环境建立的水下声源定位模型性能大幅降低甚至完全失效，重新训练模型又需要新的海洋环境新任务下的大量带标签数据，但收集带标签数据的成本往往很高，而直接使用少量的带标签数据进行定位的效果通常比较差。因此限制了传统深度学习方法在海洋与水声工程中的应用。针对以上问题，本文提出了基于元学习的水下声源定位方法，并结合动态权重和标签分布，旨在提高水下声源定位性能。主要研究内容和创新如下：

1．针对在水下声源定位任务中，由于带标签数据不足导致定位性能不佳问题，运用元学习（Meta Learning）技术和动态权重（Dynamic Weight）思想，提出了动态权重元学习水下声源定位方法（DWML）。该方法利用仿真数据构造不同的水下声源定位任务，将真实的海洋实验数据定位任务视为新任务，在新任务中仅使用少量带标签数据进行水下声源定位。在元学习方法的基础上，结合动态权重思想，提出了动态权重调节函数，该调节函数能够自适应地给不同定位任务赋予不同的损失权重，以提高定位性能。在所提方法基础上又进一步结合标签分布思想，运用标签分布向量代替传统的One-hot标签，进一步提高定位性能。通过实验证明，与传统的匹配场处理方法以及其他深度学习方法相比较，本文所提方法在仅有少量带标签数据的新任务中表现优异。

2．针对水下声源定位研究中仅有少量带标签数据和无标签数据导致定位性能不佳问题，提出了半监督元学习水下声源定位方法。考虑到带标签数据很少，而无标签数据比带标签数据更容易获得，本研究提出了一种半监督元学习模型（SSML）用于水下声源定位。SSML由参数化编码器、非参数化原型细化过程和距离函数组成。SSML带有压缩和激励注意模块，基于注意力机制，编码器能够更好地提取特征以生成原型，再利用无标签数据来改进原始原型，从而提高水下声源定位的准确性。通过具体实验证明了所提出的SSML方法的有效性，并与其他方法进行了比较，实验结果证明了该方法的优越性。

With the proposal of the strategy of marine power and the rapid development and innovation of science and technology, the importance of marine research at the national strategic level is increasing. Underwater source localization, as a key issue in the field of ocean and acoustic engineering, has always been the focus of researchers.

Matched Field Processing (MFP) can cause certain positioning errors due to environmental mismatch and other reasons. However, traditional deep learning methods for underwater source localization often require a large amount of labeled data, which may not be available in certain marine environments or under certain positioning scenarios. In addition, changes in ocean environmental parameters in new experimental scenarios can significantly reduce or even invalidate the performance of underwater source localization models based on the original ocean environment. Retraining the model requires a large amount of labeled data for the new marine environment and task, but the cost of collecting labeled data is often high, and the effect of using a small amount of labeled data for positioning is usually poor. Therefore, traditional deep learning methods are limited in their application in ocean and acoustic engineering. To address these issues, this paper proposes a meta-learning-based underwater source localization method, combined with dynamic weights and label distribution, aiming to improve the performance of underwater source localization. The main research content and innovations are as follows：

1. In order to mitigate poor localization performance caused by insufficient labeled data in a new underwater source localization task, a method utilizing Dynamic Weight Meta-Learning (DWML) is proposed. DWML uses simulation data to construct different underwater source localization tasks, and regards the real ocean experiment data localization task as a new task. In the new task, only a small amount of labeled data is used for underwater source localization. Based on the meta-learning method and inspired by  dynamic weight, a dynamic weight adjustment function is proposed, which can adaptively assign different loss weights to localization tasks to improve the performance. On the top of the proposed method, label distribution vectors are incorporated to replace one-hot labels, giving the localization performance further improvement. Experiments validate that the proposed method performs well in new tasks with only a small amount of labeled data. Compared with matching field processing methods and other deep learning methods, the localization performance attains a better level.

2. On tackling the localization performance impacted by only a few labeled and unlabeled data in underwater source localization research, a semi-supervised method based on meta learning is proposed. Considering that there are only several labeled data but unlabeled data is easier to obtain than labeled data, this study proposes a semi-supervised meta-learning model (SSML) for underwater source localization. SSML consists of parametric encoder, nonparametric prototype refinement process and distance function. SSML is equipped with squeeze-and-excitation attention modules. Based on the attention mechanism, the encoder can better extract features to generate prototypes, and then use unlabeled data to improve the original prototypes to better locate underwater sources. Specific experiments and comparison with other methods have been designed for SSML to verify its effectiveness. The experimental results manifest the superiority to conduct underwater source localization.