Sparse-sampled Near-field Acoustic Holography Based on 3D-CNN and Dense Block

A new method of sparse-sampled near-field acoustic holography (DB-SSNAH) is proposed to improve the accuracy of vibration velocity regression. The proposed method uses an upgraded stacked autoencoder (SAE) architecture and a CNN architecture based on dense block. First, a SAE is used to pick up and revert the features of vibration distribution. Then, DBCNN which is based on dense block is used to return sampled pressure to the vibration distribution. An aluminum plate is used to simulate. Theoretical vibration and sound pressure data generated by excitation forces at different positions of the plate are used for training. The result shows that the proposed method decreases the error between theoretical vibration and predictive values of neural networks compared with JTCSA-NAH especially at high frequencies. This further demonstrates the superiority of the proposed method.