Research on joint training strategy for 3D convolutional neural network based near-field acoustical holography with optimized hyperparameters

Abstract

Near-field acoustical holography (NAH) method based on 3D convolutional neural network and stacked autoencoder (CSA-NAH) can eliminate aliasing error or inverse ill-posedness of NAH methods with sparse sampling rate of measurement, ensuring the accuracy of NAH while reducing the measuring cost. However, the CSA-NAH method’s hyperparameters are not optimized, also the error amplification caused by networks interception and integration together with parameters freezing is neglected. To improve the accuracy of CSA-NAH, this paper modifies the implementation framework of it, and proposed a joint training CSA-NAH (JTCSA-NAH). Subsequently, the feasibility of JTCSA-NAH is verified by numerical example and the results show that the average reconstruction error on the near-field acoustic pressure decreases from 3.4% to 2.41%. Finally, JTCSA-NAH method is applied to the examples compared with other similar or latest NAH methods, the better performance verifies the applicability of the proposed method.