Machine learning aided near-field acoustic holography based on equivalent source method.

Area of Science:

• Acoustics
• Signal Processing
• Machine Learning

Background:

• Near-field acoustic holography (NAH) using equivalent source methods is common for sound source localization.
• Spherical harmonics are effective equivalent sources in non-reverberant conditions but struggle in reverberant environments.
• Data-driven approaches offer a potential solution for sound source characterization in complex acoustic spaces.

Purpose of the Study:

• To investigate the efficacy of machine learning (ML) models for sound source localization and characterization in reverberating environments.
• To compare the performance of different ML models against traditional optimization techniques.
• To evaluate the computational efficiency of the proposed ML methods.

Main Methods:

• Developed and tested ML models including linear regression (LR) with adaptive moment estimation (Adam), LR with limited-memory Broyden-Fletcher-Goldfarb-Shanno (L-BFGS), and multi-layer perceptrons (MLP).
• Simulated and experimentally validated methods using monopoles, vibrating plates, and loudspeakers in various acoustic conditions (reverberant rooms, free field).
• Compared ML model results against one-norm convex optimization (L1CVX) for performance evaluation.

Main Results:

• LR with L-BFGS demonstrated superior performance among the studied ML methods, outperforming L1CVX in environments with lower wall absorption coefficients for separable sources.
• The data-driven methods, particularly LR with L-BFGS, showed significant improvements in sound source localization accuracy.
• LR with L-BFGS achieved substantially faster inference times compared to other methods, including L1CVX.

Conclusions:

• Machine learning, specifically LR with L-BFGS, provides an effective and efficient approach for sound source localization and characterization in reverberating environments.
• The proposed data-driven methods offer a viable alternative to traditional techniques, especially when dealing with complex acoustic reflections.
• LR with L-BFGS presents a promising solution for real-time acoustic analysis due to its high accuracy and speed.