Advanced Three-Dimensional Acoustic Vector Sensor System: Physically-Informed Machine Learning-Driven Frequency Response Flattening and Multi-Axis Correlation Noise Optimization

In this paper, a digital three-dimensional acoustic vector sensor system is proposed to enhance the frequency response flattening and noise handling performance through advanced signal preprocessing algorithms. The system adopts physical information-based machine learning to optimize the parameters of multi-stage digital filters to accurately fit the frequency response curves of thermoacoustic particle velocity-sensitive units, and designs real-time zero-phase digital filters to achieve frequency flattening and broaden the operating bandwidth. Meanwhile, multi-channel correlation is used to optimize the noise processing, enhance the signal detection capability in low signal-to-noise ratio environment, and maintain the performance of spatial direction estimation. Anechoic chamber experiments verify that the system achieves a frequency flatness of less than $\pm 3\text{dB}$ in the 100 Hz-10 kHz band and improves the signal-to-noise ratio by about 10 dB while maintaining channel correlation. These results confirm the effectiveness and reliability of the system, laying the foundation for high-precision applications in sound source detection and noise source localization.