摘要
针对基于平面麦克风阵列的声源定位方法难以估计声源深度的问题,文章提出了多任务学习残差卷积网络的声源定位和深度估计方法。提出的网络模型具有两个输出分支,分别用于实现声源定位和深度估计。该网络使用函数波束形成的成像结果作为输入特征。通过设计一种高分辨率并且无旁瓣的目标图作为网络的标签来提高函数波束形成声源识别性能,同时将声源面与测量阵列之间的距离均匀离散成不同的深度类别,根据网络输出的深度类别的概率来估计声源深度。仿真结果表明,所提方法在五种频率的测试集中定位准确率都不低于96.95%,平均距离误差小于0.003 4 m,分类准确率大于99.05%,能够准确定位声源并估计声源深度。此外,该方法在低信噪比情况下也能有效识别声源,具有良好的泛化性。
A multi-task learning residual convolutional network(MTL-ResCNN)method for sound source localization and depth estimation is proposed to address the problem that it is difficult to estimate the sound source depth by the planar microphone array-based sound source localization methods.The proposed network model has two output branches to achieve sound source localization and depth estimation,respectively.The network uses functional beamforming(FBF)imaging results as input features.A high-resolution and side-lobes-free target map is designed as the label of the network to improve the source identification performance of functional beamforming,while the distance between the source plane and the measurement array is uniformly discretized into different depth classes,and the source depth is estimated based on the probability of the depth classes output by the network.The simulation results show that in the test set of five frequencies the proposed method has a localization accuracy of no less than 96.95%,an average distance error of less than 0.0034 m,and a classification accuracy of more than 99.05%,which can accurately locate the sound source and estimate the source depth.In addition,the method can effectively identify the sound source with good generalization even under a low signal to noise ratio.
作者
耿林
张鸽
王书海
夏晨骏
谢峰
斯嘉禾
GENG Lin;ZHANG Ge;WANG Shuhai;XIA Chenjun;XIE Feng;SI Jiahe(School of Electrical Engineering and Automation,Anhui University,Hefei 230601,Anhui,China)
出处
《声学技术》
CSCD
北大核心
2024年第6期887-895,共9页
Technical Acoustics
基金
国家自然科学基金项目(51975003)
安徽省自然科学基金项目(2108085ME175)
安徽高校协同创新项目(GXXT-2021-010)。
关键词
声源定位
深度估计
多任务学习
残差卷积网络
函数波束形成
sound source localization
depth estimation
multi-task learning
residual convolutional network(MTLResCNN)
functional beamforming
