【作者】 李敏；

【导师】 潘翔；

【作者基本信息】 浙江大学 ， 信息与通信工程， 2023， 硕士

【摘要】 随着自主水下航行器和无人艇等机动平台的快速发展,研究艇载短拖线阵的波达方向(Direction of Arrival,DOA)估计具有重要意义。在机动过程中,拖线阵弯曲导致目标测向的性能下降。因此,本文重点研究基于稀疏贝叶斯学习(Sparse Bayesian Learning,SBL)和卷积神经网络(Convolutional Neural Network,CNN)的阵形和DOA联合估计。在研究空域矩阵滤波抑制平台辐射噪声的基础上,用圆弧模型对短拖线阵进行建模,采用自适应弓高的稀疏贝叶斯学习(Adaptive Bow Sparse Bayesian Learning,ABSBL)算法对阵形和目标方位进行联合估计,降低阵形失配的影响。提出并采用ABSBL-Prune算法和高斯广义近似消息传递(Gaussian Generalized Approximate Message Passing,GGAMP)的ABSBL算法进行快速DOA估计,前者通过基向量剪枝实现算法加速,后者通过近似求解ABSBL中的后验均值和方差提高计算效率。两者在保持算法高分辨性能的同时,解决了ABSBL算法效率低下的问题。为进一步提升短拖线阵DOA估计的性能和算法的运行效率,研究了深度学习的DOA估计方法。设计基于多任务卷积神经网络(Multi Task Convolutional Neural Network,MTCNN)的DOA估计模型,通过训练数据学习拖线阵接收信号与波达方向之间的映射关系。拖线阵接收信号的协方差矩阵作为模型输入,DOA估计作为主任务,弓高估计作为辅助任务,进一步提升了DOA估计的性能。数值仿真验证了ABSBL-Prune、GGAMP-ABSBL和GGAMP-ABSBL-Prune算法以及MT-CNN模型用于短拖线阵高分辨测向的性能。国内南海拖线阵实验数据和意大利西海岸MAPEX2000拖线阵实验数据分别验证了各算法在机动中短拖线阵测向的有效性。更多还原

【Abstract】 With the fast development of maneuvering platforms such as autonomous underwater vehicles and unmanned surface vehicles,it is of great significance to research on direction of arrival(DOA)estimation of the short towed line array.During maneuvering,the bending of the towed line array leads to the degradation of the performance of the target DOA estimation.Therefore,this thesis focuses on the joint estimation of array shape and DOA based on sparse Bayesian learning(SBL)and convolutional neural network(CNN).Based on the study of the spatial matrix filtering to suppress the platform radiation noise,the short towed line array is modeled with a circular arc model,and the adaptive bow sparse Bayesian learning(ABSBL)algorithm is used for the joint estimation of array shape and DOA to reduce the effect of array mismatch.The ABSBL-Prune algorithm and the ABSBL algorithm based on Gaussian generalized approximate message passing(GGAMP)are proposed and used for fast DOA estimation,with the former achieving algorithm speedup through basis vector pruning and the latter solving the posterior mean and variance in ABSBL by approximating to improve the computational efficiency.Both of them solve the problem of inefficiency of the ABSBL algorithm while maintaining high-resolution performance.In order to further improve the performance of DOA estimation of short towed line array and the computational efficiency of the algorithm,the DOA estimation method of deep learning is studied.The DOA estimation model based on the multi-task convolutional neural network(MTCNN)is designed to learn the mapping relationship between the received signal of the towed line array and the DOA through the training data.The covariance matrix of the received signal of the towed line array is used as the model input,DOA estimation is used as the primary task,and bow estimation is used as the secondary task to further improve the performance of DOA estimation.The numerical simulation has verified the performance of ABSBL-Prune,GGAMP-ABSBL,and GGAMP-ABSBL-Prune and the MT-CNN model for high-resolution DOA estimation of the short towed line array.The experimental data of the South China Sea towed line array experiment and the west coast of Italy MAPEX2000 towed line array experiment has demonstrated the effectiveness of each algorithm for the short towed line array DOA estimation during maneuvering.更多还原