水声目标定位是国际水声学研究的重点和热点，也是国民经济建设和国防建设的重要技术。文本结合深度学习技术对浅海环境下的水声目标定位方法进行了探索。
海洋环境的复杂多变和水声数据采集的困难给利用深度神经网络（DNN）进行水声目标定位带来了不小挑战。利用DNN进行水声目标定位时，其模型性能受训练数据量影响严重。当训练数据较少时，DNN往往得不到充分训练，容易过拟合训练数据，使模型泛化能力下降。此外，在目前利用DNN进行水声目标定位的方法中，绝大多数方法需要训练数据有标签，但在实际中，阵列接收到的水声数据通常没有对应标签。本文针对上述问题，主要进行了以下研究：
1. 针对目标海域仅有少量带标签实验数据情况下的水声目标定位问题，设计了一种基于网络迁移并结合标签分布的水声目标定位方法（记为LD-TL）。该方法将在仿真数据上预先训练好的一维卷积神经网络（1D-CNN）模型的网络结构和部分参数迁移到目标域模型中，并结合标签分布思想，采用标签分布向量来描述环境的不确定性，使用少量实验数据对目标域模型进行微调，然后利用网络的输出更新标签分布向量，并将更新后的标签分布向量作为样本标签对网络进行监督训练。实验结果表明，LD-TL能够在仅具有少量实验数据情况下工作良好，与匹配场定位方法和其他基于DNN的方法相比能够显著提高水声目标定位性能。此外，LD-TL相较于其他方法具有更好的鲁棒性，且能大幅降低训练时间，极大地方便了在实际情况下的应用。
2. 针对目标海域无带标签实验数据情况下的水声目标定位问题，设计出了两种结合标签分布思想的水声目标定位算法，分别是基于对抗迁移的水声目标定位算法和基于映射迁移的水声目标定位算法。两种方法均聚焦于仿真数据并充分挖掘仿真数据和实验数据之间的关联，分别利用对抗的手段和映射的手段，对齐仿真数据和实验数据在新特征空间中的表示，再结合网络输出和标签分布，对模型进行进一步训练，得到最终的水声目标定位模型。实验结果表明，所提出的方法其定位性能优于匹配场定位方法，结合标签分布思想后可以在不同程度上提高模型的定位性能。

Underwater source localization is one of research emphases in international underwater acoustic research, and it is also an important technology for national economic construction and national defense construction. Combined with deep learning technology, the underwater source localization method in shallow sea environment is explored in this research. 
It is a difficult task for underwater source localization by deep neural network (DNN) due to the complex and changeable marine environment and the difficulty of underwater acoustic data acquisition. The performance of DNN is seriously affected by the amount of training data, when there is a lack of training data, DNN will have problems with overfitting as well as generalization abilities in practical applications. In addition, most of the current underwater source localization methods using DNN require labeled training data. However, in practice, the experimental data received by the array are usually unlabeled. In this paper, the researches mainly focus on the above-mentioned problems:
1. Aiming at the problem of only a small amount of labeled experimental data in the target sea area, a label distribution-guided transfer learning (LD-TL) for underwater source localization is proposed, where a one-dimensional convolutional neural network (1D-CNN) is pre-trained with the simulation data generated by an underwater acoustic propagation model and then fine-tuned with very limited amount of experimental data. In particular, the experimental data for fine-tuning the pre-trained 1D-CNN are labeled with label distribution vectors instead of one-hot encoded vectors, moreover, the label distribution vectors are updated by the output of the 1D-CNN. Experimental results show that the performance of underwater source localization with very limited amount of experimental data is significantly improved by the proposed LD-TL when compared with MFP, and other DNN-based methods. Moreover, results of sensitivity analysis and computational complexity show that the proposed LD-TL is robust against the environmental mismatch and can greatly reduce the training time, which is greatly convenient for practical application. 
2. For the problem of no labeled experimental data in the target sea area, two methods combined with label distribution are proposed for underwater source localization, which are adversarial-based transfer method and mapping-based transfer method, respectively. Both methods focus on simulation data and fully exploit the similarity between simulation data and experimental data. By using adversarial and mapping technology, the representations of simulation data and experimental data in the new data space are aligned to find transferable features. Then, considering the output of network and label distribution, the model is trained again to obtain the final localization model. Experimental results show that the proposed methods achieve better localization performance than MFP, and the localization performance of proposed model can be improved to different degrees by considering label distribution.