基于迁移学习的小样本垂直阵目标距离估计方法  

Target distance estimation of few-shot vertical array based on transfer learning

作　　者：姚琦海 汪勇[1,2] 杨益新 YAO Qihai;WANG Yong;YANG Yixin(School of Marine Science and Technology, Northwestern Polytechnical University, Xi′an 710072, China;Shaanxi Key Laboratory of Underwater Information Technology, Xi′an 710072, China)

机构地区：[1]西北工业大学航海学院,陕西西安710072 [2]陕西省水下信息技术重点实验室,陕西西安710072

出　　处：《哈尔滨工程大学学报》2022年第6期761-769,共9页Journal of Harbin Engineering University

基　　金：国家自然科学基金项目(11974286,61971353).

摘　　要：针对仅有少量水声数据海域的目标距离估计问题,本文以声场复声压的实部和虚部为特征,构建迁移学习模型,在利用卷积神经网络对预选海域大量水声数据进行预训练的基础上,对探测海域小样本水声数据进行再训练,从而实现小样本水声数据下的水下声源距离估计。利用SWellEX-96实验无强干扰的S5航次数据和有强干扰的S59航次数据进行了方法的验证,对实验中的浅源和深源实现距离估计,对比了匹配场处理、传统卷积神经网络和迁移学习3种方法的水下目标声源距离估计性能。实验结果表明:基于卷积神经网络的迁移学习模型在无强干扰和有强干扰2种环境中均可有效实现距离估计,且估计性能明显优于传统卷积神经网络和匹配场处理方法。To address the problem of target distance estimation in the sea area with only a small amount of underwater acoustic data,this paper takes the real and the imaginary parts of the complex sound pressure of the sound field as the characteristics,constructs a transfer learning model,and retrains the small sample underwater acoustic data in the exploration sea area on the basis of pretraining a large amount of underwater acoustic data in the preselected sea area by using the convolution neural network to realize the underwater sound source distance estimation under the small sample underwater acoustic data.The S5 voyage data in the SWellEX-96 experiment without any strong interference and S59 voyage data with a strong interference are used to verify the performance of the method.Distance estimation is achieved for shallow and deep sources in the experiment.The range estimation performance of the underwater target sound source by matching field processing,traditional convolution neural network,and transfer learning methods are compared.Results show that the transfer learning model based on a convolution neural network can provide good distance estimations in both environments with and without strong interference.Moreover,the estimation performance is significantly better than traditional convolutional neural networks and matching field processing.

关 键 词：迁移学习 小样本 垂直阵 距离估计 强干扰 卷积神经网络 SWellEX-96实验 匹配场处理 

分 类 号：TB533[理学—物理]