Reconstruction of vibration noise in plate structures based on Data-Physics driven model and transfer learning

The identification of vibration and reconstruction of sound fields of plate structures are important for understanding the vibroacoustic characteristics of complex structures. This paper presents a data-physics driven (DPD) model integrated with transfer learning (DPDT) for high-precision identification and reconstruction of vibration and noise radiation of plate structures. The model combines the Kirchhoff-Helmholtz integral equation with convolutional neural networks, leveraging physical information to reduce the need for extensive data. By embedding transfer learning, it enhances generalization across different structures. Two plate models of different sizes and publicly experimental data were used to evaluate the model's performance. Results show that the DPDT model achieves superior prediction accuracy stability, and faster convergence compared to the DPD model, with high R2, normalized cross-correlation, and low normalized mean squared error values, demonstrating its robustness and efficacy in reconstructing sound fields even with limited data points. This approach demonstrates significant potential for practical engineering applications, particularly in bridge vibration and noise control.