| In the ongoing exploration of the rich marine resources,underwater wireless optical communication has great potential with the development of underwater wireless sensor networks due to its huge advantages in terms of high communication rate,low equipment cost and power consumption.Many technologies in the underwater wireless sensor networks such as routing and topology control highly rely on the location of nodes,which always change due to currents,underwater biological activity and autonomous movement of nodes,making it important to achieve precise and real-time localization between nodes.Deep learning can be applied to positioning systems with good performance,but it usually requires large-scale data sets to train network models,and the variability of the underwater environment causes the data to not follow the same distribution,and the difficulty of collecting underwater data makes it unlikely to build large databases,which can greatly reduce the localization performance of deep learning.To address the above problems,in this thesis,a distributed underwater optical localization scheme is proposed,and an underwater laser communication scenario and node structure are designed.The beam spread function is used to calculate the signal strength of the receiver and a fingerprint-based localization algorithm is proposed.Not only the 3D position of the laser but also the direction of the beam can be obtained at the receiving nodes,which makes a solid foundation for the subsequent alignment of the communication link for high-quality information transmission.Firstly,in order to achieve a non-linear mapping between the received signal strength and the position and direction of the transmitted laser,a fingerprint-based underwater optical node localization model Underwater Optical-Convolutional Neural Networks is proposed in this thesis,and an attention mechanism,SE module is introduced in the convolutional neural network to improve the localization performance of the system.Moreover,the Cramer-Rao lower bound for this localization scenario is derived to verify the effectiveness of the model for position and direction prediction.The experimental results show that UO-CNN can achieve better localization accuracy than other fingerprint-based localization algorithms,and the introduction of the attention mechanism can effectively improve the localization performance of the model when training relatively small-scale data.Then,to address the variability of underwater localization scenarios and the difficulty of collecting data,a domain adaptation-based transfer learning model Underwater Optical-Domain Adaption is proposed to solve the localization performance degradation problem due to domain offset.The maximum mean discrepancy(MMD)and CORAL functions are used to measure the difference between the features of source and target domain,respectively,and domain feature discrepancy and the prediction loss of the network model are minimized during the training process to improve the localization performance in different scenarios.The experimental results show that UO-DA can adapt to different underwater scenarios and achieve better localisation accuracy compared with other deep domain-adaptive networks,effectively improving the localization performance. |
