Underwater Source Localization Using Metamodel and Optimization Method

Abstract

Traditional strategies of metamodel construction have been applied to regression problems and demonstrated excellent performance in data requirements, accuracy, and noise robustness in recent years. Moreover, underwater source localization is also regarded as a regression problem. Therefore, the metamodeling approach is introduced to underwater source localization to address the lack of noise robustness, environmental mismatch, and difficulties in data acquisition in localization methods. The source localization problem is formulated as a regression modeling and optimization process, realized through the metamodel and optimization method (MMOP). Within this framework, the metamodel takes the source position as input and the feature values of normalized sample covariance matrix (SCM) of the pressure data as output. The source position is subsequently estimated by minimizing the differences between the predicted and actual feature values. The MMOP algorithm is evaluated against the feedforward neural network (FNN) algorithm and the matched field processing (MFP) algorithm through simulations and experiments. Simulation environment and experimental datasets are sourced from the SWellEx-96 experiment. Results indicate that the MMOP surpasses both FNN and MFP in localization accuracy. Additionally, the influence of signal bandwidth, array element configuration, snapshot number, and training data size on MMOP performance is analyzed, accompanied by strategies to enhance the algorithm’s performance.