利用深层卷积网络自适应增强学习的水声目标线谱提取方法

Line spectrum extraction of underwater acoustic target using deep convolution network and adaptive enhancement learning

提出了一种使用自适应增强学习的深层卷积神经网络方法对水声目标线谱进行提取。该方法利用构造的多尺度ConvNeXt算子建立滑动窗深层卷积神经网络模型(SwDCNN),设计涵盖损失函数、学习率更新和模型迭代优化的自适应增强学习准则并用于模型训练。仿真和海试数据验证结果表明,所提方法有以下优点:(1)卷积算子和模型结构参数按线谱提取需求配置,可以增强LOFAR谱图特征高性能多分辨力挖掘能力;(2)大规模数据下的模型训练可实现渐进式精确拟合,有助于提升模型收敛效果;(3)模型可有效提取低信噪比、中断、弯曲漂移、粗细不均、邻近成簇、密集分布等复杂情况下的线谱,在查全率、查准率、虚警率、线谱位置精度(LLA)和线谱幅值精度(LAA)等指标上均优于文中其他深度神经网络方法;(4)和传统及其他文中所用的深度神经网络方法相比,线谱最小可检测信噪比分别降低超过5 dB和2 dB,实际复杂场景线谱提取能力更强,综合效果更好。

A deep convolution neural network method using adaptive enhancement learning is proposed to extract underwater acoustic target line spectrum.Firstly,a multi-scale ConvNeXt operator is constructed.Then a sliding window deep convolution neural network model(SwDCNN) for extracting line spectrum from Lofargram is established based on the operator.Finally,an adaptive enhancement learning criterion covering loss function,learning rate update and model iterative optimization is designed and used for SwDCNN model training.The numerical simulations and sea trial data show that the proposed method has four advantages:(1) The structure parameters of SwDCNN model are configured according to the line spectrum extraction requirements,which can enhance the multi-resolution mining ability of Lofargram features.(2) Under the condition of large-scale data,the model training effect of progressive accurate fitting can be achieved,and the model convergence effect can be improved.(3) The line spectrum under complex conditions such as low SNR,interruption,bending drift,uneven thickness,clustered nearest neighbors and dense distribution can be effectively extracted by the SwDCNN model.In terms of recall,precision,false alarm,LLA(Line Location Accuracy) and LAA(Line Amplitude Accuracy),the SwDCNN model is superior to other deep neural network methods used in this paper.(4) Compared with the traditional and other common deep neural network methods,the line spectrum minimum detectable SNR of the SwDCNN model is reduced by more than 5 dB and 2 dB respectively.The SwDCNN model has stronger line spectrum extraction ability and better comprehensive effect in actual complex scenes.