Accelerating unsupervised near-field acoustic holography via frequency-domain parameter transfer

Near-field acoustic holography (NAH) enables sub-wavelength sound field reconstruction but demands repeated inverse solutions across hundreds of frequency bins in broadband engineering tests. Unsupervised physics-informed neural networks avoid pre-constructed training datasets, yet their frequency-by-frequency optimization leads to computation times that scale linearly with the number of bins. This paper proposes DP-PINN+, which exploits the analytic continuity of the NAH transfer function with respect to frequency to transfer converged network weights between adjacent bins as warm-start initializations. A normalized transfer-function distance is introduced as an offline metric for assessing transfer feasibility. Online convergence detection adaptively terminates training when the loss improvement rate drops below a preset threshold, and a drift-detection mechanism resets the transfer chain when accumulated bias exceeds tolerance. Simulations over 100–5000 Hz (491 bins) at signal-to-noise ratios of 10, 20, and 30 dB show that DP-PINN+ reduces the average per-bin computation time from 23.78 s to 2.34 s while achieving a broadband-averaged normalized mean squared error of −27.43 dB at 30 dB SNR, with a lower mean NMSE than the baseline physics-informed method and the supervised convolutional neural network. Ablation experiments indicate that parameter transfer primarily provides computational acceleration, whereas DIP mitigates noise-overfitting bias during sequential transfer at low SNR. Sound-insulation measurements on a glass-panel window demonstrate a 4.36× speedup while maintaining numerically comparable reconstruction accuracy under the investigated experimental configuration.