机电工程技术2025，Vol.54Issue(7)：60-64,5.DOI:10.3969/j.issn.1009-9492.2025.07.012
基于轻量化卷积神经网络的DOA估计
DOA Estimation Based on Lightweight Convolutional Neural Networks
摘要
Abstract
The existing DOA estimation algorithm based on deep neural network can fully extract spatial features,improve the estimation accuracy of DOA and reduce the algorithm complexity.Under the condition of low signal-to-noise ratio and few fast beats,its estimation accuracy is obviously better than the traditional algorithm.However,while improving the accuracy and precision,it leads to a large number of network model parameters,high computing cost and storage cost,and it is difficult to deploy to edge computing devices with limited resources.In view of this situation,a lightweight convolutional neural network DOA estimation algorithm based on edge computing devices is proposed.Ghost convolution is used to quickly and effectively extract input covariance matrix features,SE attention mechanism is used to adaptively re-calibrate the weight of feature channels,and attention to important features is enhanced.Ghost bottleneck structure and depth can be separated.The computational complexity and the number of model parameters are reduced,and the model is lightweight.The experimental results show that the accuracy of the model is 97.82%,the size is 289.83 kB,and the number of parameters is 6123.Compared with GooleNet,ShuffleNet,MobileNetV1,MobileNetV2 and MobileNetV3 models,the proposed model improves accuracy and computational efficiency while reducing model size and parameter number,and has good performance.关键词
波达方向估计/深度学习/轻量化模型/均匀线阵/边缘计算Key words
estimation of porta-direction/deep learning/lightweight model/uniform linear array/edge computing分类
计算机与自动化引用本文复制引用
范大照,章勇,熊伟华,陈礼坤,刘素芳..基于轻量化卷积神经网络的DOA估计[J].机电工程技术,2025,54(7):60-64,5.基金项目
江西省"双千计划"长期项目(DHSQT22021003) （DHSQT22021003）
