Keywords: topology optimization, acoustic-structural system, artificial neural network, transfer learning, MMC, sound quality.

Sound quality is an important measure of the sound performance of acoustic devices. However, multi-frequency calculations in sound quality optimization can lead to poor solvability of the optimization problem. Data-driven approach is an effective way to solve multi-frequency computing problems. However, as the acoustic device products are updated, the structure or environment will change. The original neural network model will no longer be applicable and the prediction accuracy will be severely reduced. The new optimization task requires the collection of new data samples and the training of new neural networks. The extensive data collection process and iterative optimization process further reduce the solvability of the sound quality optimization problem. For the sound quality optimization of acoustic devices, this paper proposes a data-driven acoustic-structural topology optimization design method that can quickly and accurately predict the acoustic frequency response and significantly improve the computational efficiency problem. Deep transfer learning is also introduced to achieve fast and accurate prediction of acoustic frequency response using a small amount of sample data in a new structure/environment. The main contributions of this paper are as follows: (1) A new agent model based on deep neural network (DNN) is proposed to replace the complex finite element model, combined with Movable Morphable Components (MMC) method with a small number of design variables to achieve sound quality optimization of acoustic devices. (2) Deep transfer learning is introduced for the DNN training, which realizes the rapid and accurate prediction of sound pressure frequency response in new tasks by using a small amount of sample data, and enhances the adaptability of DNN agent model in different optimization tasks. (3) Numerical examples demonstrate that the proposed method can reduce the data dependence. In the multi-layer iterative optimization task, using small sample data for multiple transfer learning can achieve efficient optimization design and greatly reduce the design complexity.

download the full-text of this paper (PDF, 11 pages, 485 Kb)

go to the previous paper
go to the next paper
return to the table of contents
return to the volume description