基于三维N型卷积神经网络和频域注意力-亥姆霍兹正则化的近场声源重建方法

Near-field acoustic reconstruction method based on three-dimensional N-shaped convolution neural network and frequency focal-KH regularization

[目的]针对全息面、低采样率条件下近场声源重建误差较大的问题,提出一种高分辨率、低误差的平面声源表面法向振速重建的深度神经网络框架。[方法]首先,建立用于近场声源重建问题的三维N型卷积神经网络框架(包含预编码器),通过提取空间声场频域内的特征,以弥补空间信息的稀疏性;然后,提出频域注意力机制,设计包含频域注意力–归一化重建均方误差、亥姆霍兹正则项的损失函数,以自适应增加频域内难训练样本的损失权重,从而提升声源在高频和本征特征区间的重建精度;最后,通过Matlab对COMSOL Multiphysics软件进行二次开发,建立矩形薄板声振模型的训练集和测试集,开展对比验证。[结果]对比结果表明,该方法在验证集上100~2000 Hz内的平均重建误差仅为4.96%,重建精度明显高于SRCNN和PV-NN。[结论]该研究成果可以降低近场声源重建实船应用中的全息面采样点数量,同时可保证较高的声源面法向振速重建精度。

[Objectives]Low sampling rates on reconstruction surfaces cause high reconstruction error in near-field acoustic holography.Therefore,a deep learning-based approach which is applicable to planar sound sources and high-precision reconstruction with low sampling rates is put forward.[Methods]A three-dimensional N-shaped convolution neural network for near-field acoustic reconstruction is established to extract features in the frequency dimension in order to make up for sparse sampling in the spatial dimension.A frequency focal mechanism,namely an adaptive frequency weight focus mechanism,is put forward to improve reconstruction precision in the natural frequency and high frequency.Moreover,this paper also raises frequency-scaled focal loss and frequency-scaled focal Kirchhoff–Helmholtz(KH)loss,which are considered regularization.To validate the proposed methods,datasets are created with COMSOL Multiphysics and Matlab.[Results]The mean error range of 100–2000 Hz of the algorithm proposed in this paper is only 4.96%,higher than those of SRCNN and PV-NN.[Conclusions]The proposed method is verified as having the potential to reconstruct the accurate velocity fields of sound sources under low sampling rates.