Feature Extraction Method of Active Echo Highlight Structure Based on Transformer Network

Physical feature extraction of underwater acoustic targets is crucial for target detection and identification, which is able to transform the physical properties of targets into quantifiable feature parameters. Structure extraction is performed by statistically analyzing some characteristics in the target echo signal, such as the number of highlights, intensity distribution, time interval. Traditional highlight structure extraction methods rely on manually designed signal processing techniques. However, these methods often face challenges in maintaining extraction accuracy under complex environmental noise. Deep learning methods effectively solve these problems by constructing multilayer neural networks to automatically learn feature representations from the data. To address the above issues, this paper proposes a feature extraction method for active echo highlight structure based on a Transformer network. The underwater acoustic target-echo highlight model has been designed. Furthermore, a deep learning model was constructed based on Transformer to achieve patch embedding and feature extraction of the input samples. Moreover, a joint loss function, including the regression loss of echo highlight structure, the regression loss of target scale, and the time extension correction loss, was designed to improve the convergence effect of the model and the accuracy of feature extraction. The model training results show that the prediction error of the number of highlights can be controlled within 5, and the prediction error of the target scale is controlled within 12 meters. In addition, the model is tested on actual underwater acoustic target echo data, and achieves a good prediction performance of the highlight structure on real target echoes.