【作者】 李向欣；

【导师】 李理；

【作者基本信息】 哈尔滨工程大学 ， 水声工程， 2024， 硕士

【摘要】 匹配场处理方法被广泛应用于海洋环境参数反演、水下目标定位等领域。匹配场技术的核心思想是首先对拷贝声场和实测声场数据做匹配相关,进而使用全局优化算法来获得环境参数以及目标位置等信息。然而在参数量或者参数划分精度增加时,匹配场处理的计算量和所需存储空间会呈现指数级增加,其导致的硬件压力问题在高实时性需求和存储受限的应用场景下尤为突出。近年来深度强化学习算法在路径决策等全局优化问题上,相比传统算法在计算耗时上有明显的优势,本文将其与匹配场处理问题结合,用于匹配场中的全局优化过程,此外本文还引入了存储空间优化策略来降低匹配声场的计算量,以减轻硬件存储的压力。针对海洋环境参数的反演,本文构建了基于深度强化学习模型(Deep-Q learning Network,DQN)的参数反演模型,对沉积层厚度、沉积层密度、沉积层上端声速、沉积层下端声速、海底衰减系数,和海底声速六个参数进行反演。本文使用简正波模型构建声场仿真数据集并进行模型训练,之后使用Elba岛海试数据对模型的有效性进行了验证。结果表明,与传统的遗传算法相比,本文所提的基于DQN的参数反演模型在获得相似的精度前提下,耗时仅为遗传算法1/3;通过M-H采样对反演结果后验概率密度的分析表明,使用DQN求得的反演参数与遗传算法寻优得到的参数值均在密度分布最集中的70%概率区间内,两种算法得到的参数估计平均值基本重合,进一步验证了反演结果的可靠性。在此基础上,本文利用基于敏感性指数的多步迭代反演策略,依据敏感度不同,将全部反演参数形成的高维空间分割为两个低维子空间,通过将在原始的高维度空间进行搜索转换为依次在两个低维子空间进行串行搜索的方式降低计算量。实验表明,该策略将声场计算耗时从882 min降低到21 min,同时基本保证了反演精度不变,证明该策略通过降低声场计算量,在反演中带来的时间优化能力相比于DQN在全局优化过程的性能提升要更加显著。针对水下目标被动定位问题,本文构建了基于DQN的目标定位模型,并将上文的海底环境参数估计值作为先验信息来对声源目标深度和距离进行估计,该目标定位问题转化为训练DQN模型在二维空间中找到抵达目标的最优路径,并在仿真和海试实验数据上对模型进行了验证,通过实验证明在同样的计算精度下所提的DQN模型能够将定位耗时降低到原来的1/45。此外本文提出网格空间压缩策略,将目标寻优过程分解为多步进行,划分网格的精度由粗到细进行逐步搜索,从而降低网格声场的计算量,实验表明该策略在不影响定位精度的前提下将声场计算量降低到全网格搜索的1/23,有效提升了计算效率,证明了该方法的有效性。更多还原

【Abstract】 The matching field processing method is widely used in the fields of Marine environmental parameter inversion and underwater target location.The core idea of the matching field technology is to first match the data of the copied sound field and the measured sound field,and then use the global optimization algorithm to obtain the environmental parameters and target location information.However,when the number of parameters or the precision of parameter partitioning increases,the computational amount and storage space required for matching field processing will increase exponentially,resulting in hardware pressure problems that are particularly prominent in the application scenarios with high real-time requirements and limited storage.In recent years,the deep reinforcement learning algorithm has obvious advantages in computing time compared with the traditional algorithm in global optimization problems such as path decision making.In this paper,it is combined with the matching field processing problem to apply to the global optimization process in the matching field.In addition,this paper also introduces the storage space optimization strategy to reduce the calculation amount of the matching sound field,so as to reduce the pressure of hardware storage.Aiming at the inversion of Marine environmental parameters,a parameter inversion model based on Deep-Q learning Network(DQN)was constructed in this paper to invert six parameters: thickness of sediment layer,density of sediment layer,sound velocity at the upper end of sediment layer,sound velocity at the lower end of sediment layer,submarine attenuation coefficient,and submarine sound velocity.In this paper,the normal wave model is used to build a sound field simulation data set and train the model.Then,the validity of the model is verified by the sea test data of Elba Island.The results show that compared with the traditional genetic algorithm,the proposed parameter inversion model based on DQN reduces the optimization time by 66% under the premise of obtaining similar accuracy.The analysis of the posterior probability density of the inversion results by M-H sampling shows that the inversion parameters obtained by DQN and the parameters optimized by genetic algorithm are both in the probability interval of 70% with the most concentrated density distribution,and the estimated average values of the parameters obtained by the two algorithms are basically identical,which further verifies the reliability of the inversion results.On this basis,this paper uses a multi-step iterative inversion strategy based on sensitivity index.According to different sensitivities,the high-dimensional space formed by all inversion parameters is divided into two low-dimensional subspaces,and the computation is reduced by converting the search in the original high-dimensional space into serial search in two low-dimensional subspaces.Experiments show that this strategy reduces the sound field calculation time from 882 min to21 min,and basically guarantees the same inversion accuracy.It proves that this strategy can improve the time optimization ability in the inversion more significantly than the performance of DQN in the global optimization process by reducing the amount of sound field calculation.Aiming at the passive location problem of underwater targets,this paper constructs a target location model based on DQN,and uses the estimated value of the above submarine environmental parameters as prior information to estimate the depth and distance of the target of the sound source.The target location problem is transformed into training the DQN model to find the optimal path to the target in two-dimensional space.The model is verified on the simulation and sea test data,and the experiment proves that the proposed DQN model can reduce the positioning time to 1/45 of the original one under the same calculation accuracy.In addition,this paper proposes a grid space compression strategy,which divides the target optimization process into multiple steps,and searches the grid precision step by step from coarse to fine,thus reducing the computational load of the grid sound field.Experiments show that this strategy can reduce the computational load of the sound field to 1/23 of that of the whole grid search without affecting the positioning accuracy,effectively improving the computational efficiency.The effectiveness of this method is proved.更多还原