Deep Learning-Driven Non-Contact Sound Source Localization via Multi-Axis Analysis with Laser Doppler Vibrometry.

Area of Science:

• Acoustics and Signal Processing
• Biomedical Engineering
• Machine Learning

Background:

• Accurate sound source localization is crucial for clinical diagnostics and understanding physiological insights from acoustic signals.
• Traditional diagnostic tools face limitations in complex sound propagation scenarios.
• Non-invasive methods are needed to enhance diagnostic accuracy and develop advanced assistive technologies.

Purpose of the Study:

• To develop and validate a non-invasive sound source localization methodology using multi-axis vibration analysis and deep learning.
• To estimate the direction of arrival (DoA) of sound sources with high precision.
• To explore the potential applications in clinical diagnostics, acoustic engineering, and assistive hearing.

Main Methods:

• Utilized Laser Doppler Vibrometry (LDV) to measure sound-induced surface vibrations.
• Extracted directional information from Log Power Spectra (LPS) and applied theoretical modeling of multi-axis vibrations.
• Integrated deep learning techniques, including convolutional operations and Bayesian inference, for DoA estimation.

Main Results:

• Achieved an average classification accuracy exceeding 97% for sound source localization across a wide angular range (-90° to 90°).
• Demonstrated consistent performance across experiments with two distinct materials and varying frequencies.
• Validated the effectiveness of the proposed framework in precise DoA estimation.

Conclusions:

• The multi-axis vibration analysis combined with deep learning offers a highly accurate and non-invasive approach to sound source localization.
• This methodology shows significant promise for enhancing diagnostic capabilities in cardiology, pulmonology, and other medical fields.
• The findings pave the way for advancements in acoustic engineering and the development of next-generation assistive hearing technologies.