CV-CNN与稀疏贝叶斯学习结合的声源定位方法研究OA
CV-CNN combined with sparse bayesian learning for sound source localization
针对现有水下目标定位算法大多依赖于声源数目已知这一先验条件,但在实际应用中,由于声源数目往往无法预先获取或估计存在偏差,常导致定位精度下降乃至失效的问题.提出一种融合复数卷积神经网络(complex-valued convolutional neural networks,CV-CNN)与稀疏贝叶斯学习的声源定位方法.首先在声源数目预测阶段,利用神经网络学习传感器接收数据与声源数目之间的关系,估计未知声源的数目;随后在声源定位阶段,基于已估计的声源数目,采用离格稀疏贝叶斯学习算法完成对目标声源的定位.仿真表明,所采用的CV-CNN模型在不同信噪比条件下对混合数据集的声源数目估计准确率可达99.16％;方法在低至-5 dB信噪比时的定位均方根误差小于1°,在快拍数为100时仍能将误差保持在1°以内,表现出良好定位精度.
Most existing underwater target localization algorithms rely on the prior condition that the number of acoustic sources is known.However,in practical applications,the number of sound sources is often difficult or incorrect in advance,so it often leads to reduced positioning accuracy or even failure.Therefore,this paper presents a novel method for sound source localization based on Complex-Valued Convolutional Neural Networks(CV-CNN)and sparse Bayesian learning.First,a neural network learns the relationship between sensor data and sound source counts,then uses it to predict the number of unknown sources.Following this,for sound source localization,the sparse Bayesian learning algorithm locates the target based on the estimated number.The CV-CNN model shows 99.16％accuracy,and the localization error remaining under 1° not only underwater typical conditions but also at a low SNR of-5 dB or with only 100 snapshots.
崔晶;邢传玺;魏光春;董赛蒙
云南民族大学电气信息工程学院,云南 昆明 650500云南民族大学电气信息工程学院,云南 昆明 650500云南民族大学电气信息工程学院,云南 昆明 650500云南民族大学电气信息工程学院,云南 昆明 650500
通用工业技术
阵列信号处理深度学习离格稀疏贝叶斯学习DOA估计
array signal processingdeep learningoff-grid sparse bayesian learningDOA estimation
《云南民族大学学报（自然科学版）》 2026 (1)
107-116,10
国家自然基金(61761048)云南省基础研究专项面上项目(202101AT070132).
评论