应用知识蒸馏的深度神经网络波束形成算法

Beamforming algorithm for deep neural network using knowledge distillation

摘要:

自适应波束形成技术广泛应用于雷达领域的旁瓣抗干扰中。当回波数据量增多时，传统的波束形成算法无法进行快速处理，而应用深度神经网络模型通过数据的预训练则可以快速地进行波束形成，因此根据波束形成原理设计深度神经网络，并利用知识蒸馏的方式对深度神经网络进行压缩，使压缩后的模型既有原始模型良好的泛化性能而且又有更快的计算速度。仿真结果表明，相比于传统的 LMS 算法，在实验环境下，未经模型压缩的深度神经网络自适应波束形成算法的计算速度提高了约 7 倍，基于模型压缩的深度神经网络自适应波束形成算法的计算速度提高了约 20 倍。

Abstract:

Adaptive beamforming technology is widely used in sidelobe anti-interference in the radar field. When the amount of echo data increases, the traditional beamforming algorithm cannot perform fast processing, and the deep neural network model can quickly perform beamforming through data pre-training. Therefore, this paper designs a deep neural network according to the beamforming principle. The deep neural network is compressed by means of knowledge distillation, so that the compressed model has both good generalization performance and faster calculation speed. The simulation results show that compared with the traditional LMS algorithm, the computational speed of the adaptive beamforming algorithm for deep neural networks without model compression is improved by about 7 times and the computational speed of the adaptive beamforming algorithm based on model compression is improved by about 20 times in the experimental environment.