| With the increasing demand for understanding,exploring,and utilizing the ocean,underwater wireless sensor networks have emerged as a crucial technology for supporting various underwater applications.This has garnered significant attention from both academia and industry.Unlike terrestrial radio networks,underwater networks face severe resource limitations.Due to the influence of seawater propagation characteristics,underwater acoustic communication has become a primary communication approach for establishing underwater wireless sensor networks.However,the quality of acoustic communication is highly dependent on the sound field structure,which is significantly affected by the marine environment and exhibits significant spatial and temporal variability.Additionally,the long propagation delays and narrow available bandwidth of the underwater acoustic channel pose substantial challenges to the performance of underwater wireless sensor networks.Furthermore,the robustness of underwater wireless sensor networks is also threatened by potential node failures.Effectively addressing these issues and providing fair,effective,and robust communication services for diverse underwater applications have become critical concerns for underwater wireless sensor networks.Marine environment prediction and performance optimization techniques offer promising solutions: accurate prediction of changes in the marine environment and sound field structure facilitates wellinformed network deployment,while optimized resource allocation within the constraints of underwater networks enables improved network performance,ultimately enhancing the service provision for underwater applications.Therefore,this thesis focuses on the research of two important scientific problems: marine environment prediction and performance optimization of underwater wireless sensor networks,addressing the challenges of high spatio-temporal variations in marine environments,limited energy resources of underwater nodes,difficulties in underwater network maintenance,and the complexity of underwater acoustic channels.By leveraging the representation abilities of deep neural networks,the spatio-temporal prediction abilities of recurrent neural networks and convolutional neural networks,as well as the decision-making abilities of deep multi-agent reinforcement learning in multi-user systems,this research is conducted in three aspects: complex marine environment,distributed underwater network performance optimization,and semicooperative optimization for underwater communication environments.The specific research content and innovations of this thesis are as follows:(1)Marine environment prediction method based on spatio-temporal dependencies: The performance of underwater wireless sensor networks and their communication service quality are significantly influenced by the marine environment.Accurately predicting environmental changes and acoustic field variations is crucial for achieving reasonable and effective long-term network deployment.Therefore,this thesis first conducts research on marine environment prediction tasks.To address the challenges posed by the high spatio-temporal variability and data sparsity in the marine environment,this thesis presents a deep learning-based algorithm for predicting spatio-temporal sensing data in the ocean using deep learning methods.The framework consists of a generative module and a prediction module.The generative module,implemented using deep neural networks,learns the distribution of marine environment data to generate high-resolution marine environment datasets that can be used for prediction tasks,thus addressing the challenge of data sparsity.The prediction module is based on the proposed multivariate convolutional long short-term memory network.It leverages the spatio-temporal dependencies between data and the inherent correlation between different environmental features to predict future trends,thereby addressing the challenge of high spatio-temporal variability.Experimental results demonstrate that the proposed method can accurately predict environmental changes and acoustic field variations,which can be used to guide network deployment and node transmission parameter selection.(2)Deep multi-agent reinforcement learning-based performance optimization algorithm for distributed underwater wireless sensor networks: To address the energy constraints of underwater wireless sensor network nodes and the trade-off between communication fairness and network capacity,this thesis first defines a fair reuse index that simultaneously measures the degree of communication fairness and network reuse.Maximizing network fair reuse is set as the optimization objective,taking into account constraints such as concurrent communication constraints,communication fairness constraints,node transmission power constraints,communication performance constraints,network lifetime constraints,and node energy constraints.A network performance optimization model is established.Next,a deep multi-agent reinforcement learning-based algorithm for network performance optimization is proposed,aiming to jointly optimize network communication fair reuse and network capacity by allocating node transmission power reasonably.Finally,this thesis discusses the impact of different reward function designs on node transmission behaviors and network communication performance,as well as analyzes how to choose reward functions to construct decision models across different scenarios and requirements.Experimental results demonstrate that the distributed network performance optimization algorithm proposed in this thesis achieves higher fairness and reuse than other benchmark methods in all scheduling environments,while causing only a slight decrease in network capacity.By considering communication fairness,the proposed algorithm effectively balances node energy consumption and delivery delay.(3)Deep multi-agent reinforcement learning-based semi-cooperative performance optimization algorithm for underwater wireless sensor networks: To address the difficulties in underwater network maintenance and the complexity of underwater acoustic channels,this thesis presents a semi-cooperative underwater performance optimization algorithm based on deep multi-agent reinforcement learning.By perceiving the communication environment and adaptively adjusting node transmission parameters,it achieves fair,effective,and robust communication in imperfect underwater wireless sensor networks.The proposed method trains the power allocation model in IC-UWSNs through two key processes.The first process is the internal model training process,which aims to improve the decision-making capability of the power allocation model based on local observations for solving performance optimization problems.The second process is the external advanced learning procedure,which aims to improve the ability of the power allocation model to maintain system robustness in IC-UWSNs.Compared to existing methods,the proposed approach considers real system characteristics,such as potential node malfunctions,complex underwater communication environments,and limited node energy,bridging the gap between simulation environments and real systems and improving the practicality of the method.Experimental results validate the rationality of the proposed method and the effectiveness of its key components. |
