Physics-informed neural network enhanced acoustic source identification via cross-spectral correlation of microphone array signals

Acoustic source identification techniques have gained attention due to their wide range of applications in noise localization, equipment monitoring, and other fields. While traditional data-driven machine learning models utilize signal correlations from array measurements, their reliance on large training datasets (often exceeding millions of samples) and lack of physical interpretability hinder practical applications. To address these limitations, this paper proposes a neural network approach that incorporates a physical prior to achieve efficient acoustic source identification through the synergistic optimization of cross-spectral analysis and physical constraints. Specifically, this method encodes the acoustic energy conservation law and Rayleigh resolution limit into the loss function to construct a physics-informed neural network model, which significantly reduces the demand for training data. Numerical simulations show that the average absolute error of 0.4154 × 10 -2 and mean square error of 0.0815 × 10 -2 of the proposed method are reduced by 23.2 % to 40.2 % compared to the traditional data-driven model when using only 50,000 samples, which represents 5 % of the training data. Notably, the proposed method significantly improves transfer learning capability even with smaller training datasets and conditions different from the training conditions. Experimental validation further confirms the superiority of the method in semi-anechoic environments. The results provide a new paradigm for small-sample acoustic imaging while deepening the interpretable analysis of the acoustic signal propagation mechanism.