Enhancing underwater single snapshot DOA estimation for limited dataset with modified knowledge distillation

In recent years, the progress in DOA estimation using deep learning algorithms has attracted significant attention. However, their heavy reliance on extensive datasets poses a critical limitation, particularly in underwater settings where data collection is arduous. Furthermore, the lack of temporal correlation and statistical properties inherent in single-snapshot information lead to low accuracy in single-snapshot DOA estimation. To confront the above hurdles, this paper introduces an approach to improve the accuracy of underwater single snapshot DOA estimation with limited underwater datasets. By modifying the process structure and model characteristics of knowledge distillation (KD), we construct a new distillation structure that can bridge the gap between single snapshot data and multi-snapshot data sharing identical labels, achieving a breakthrough in compressing multi-snapshot data. This enhances the neural network's capacity to process both few-snapshot datasets and single-snapshot datasets. In addition, we designed novel input features to reduce the difficulty of CNN fitting by extracting the real and imaginary parts of the analytical signals, and integrated the array structure information to improve the generalization ability of our network in different scenarios. Besides, based on these innovations, we build a mapping framework between synthetic and real underwater datasets. This work involves second-order joint training of KD and transfer learning, which can help deal with small samples. The experiment results of our method show significant improvements in underwater DOA estimation accuracy, coupled with a marked reduction in overfitting risks associated with limited datasets. This work not only advances the application of deep learning in challenging underwater scenarios but also lays a foundation for future data-driven inference strategies.