To address the challenges in deep learning-based underwater acoustic source range estimation caused by high-dimensional discrete labels and scarce intra-class samples that constrain model feature learning,this study p...To address the challenges in deep learning-based underwater acoustic source range estimation caused by high-dimensional discrete labels and scarce intra-class samples that constrain model feature learning,this study proposes a siamese neural network-based feature extraction method for target range estimation.First,a dataset containing distancelabeled positive/negative sample pairs was constructed using simulated data,followed by the design and training of a siamese neural network to extract range-discriminative features.Subsequently,siamese neural network feature extraction-based convolutional neural network(S-CNN)and siamese neural network feature extraction-based residual neural network(S-ResNet)were developed through transfer learning strategy.Simulation results demonstrate the method’s effectiveness in enhancing range-sensitive feature representation:S-CNN/S-ResNet outperformed baseline models without feature extraction(NS-CNN/NSResNet),with S-ResNet exhibiting superior robustness across varying training sample sizes,signal-to-noise ratios,and environmental uncertainties.The SWellEX-96 experiment validation confirmed the significant advantages of the proposed method over conventional matched-field localization techniques.Notably,the S-ResNet achieved 10%higher confidence probability and 2%lower mean percentage error compared to the S-CNN.展开更多
基金supported by the Taishan Scholars Program,the Shandong Provincial Natural Science Foundation(ZR2024QD082)the Qingdao Natural Science Foundation(23-2-1-103-zyyd-jch).
文摘To address the challenges in deep learning-based underwater acoustic source range estimation caused by high-dimensional discrete labels and scarce intra-class samples that constrain model feature learning,this study proposes a siamese neural network-based feature extraction method for target range estimation.First,a dataset containing distancelabeled positive/negative sample pairs was constructed using simulated data,followed by the design and training of a siamese neural network to extract range-discriminative features.Subsequently,siamese neural network feature extraction-based convolutional neural network(S-CNN)and siamese neural network feature extraction-based residual neural network(S-ResNet)were developed through transfer learning strategy.Simulation results demonstrate the method’s effectiveness in enhancing range-sensitive feature representation:S-CNN/S-ResNet outperformed baseline models without feature extraction(NS-CNN/NSResNet),with S-ResNet exhibiting superior robustness across varying training sample sizes,signal-to-noise ratios,and environmental uncertainties.The SWellEX-96 experiment validation confirmed the significant advantages of the proposed method over conventional matched-field localization techniques.Notably,the S-ResNet achieved 10%higher confidence probability and 2%lower mean percentage error compared to the S-CNN.
