Sparse Bayesian learning with Bernoulli-Gaussian priors for off-grid matched field processing.

Area of Science:

• Signal Processing
• Array Processing
• Compressed Sensing

Background:

• Conventional matched field processing (MFP) suffers from basis mismatch errors due to discrete grid computations.
• Existing basis mismatch mitigation methods are computationally intensive or require specific function forms, limiting their application to MFP.

Purpose of the Study:

• To develop a grid-adaptive model for alleviating basis mismatch in MFP.
• To propose an off-grid sparse Bayesian learning algorithm for precise source localization.

Main Methods:

• Localized optimization of grid nodes within a grid-adaptive model.
• Development of an off-grid Bernoulli-Gaussian sparse Bayesian learning algorithm using variational expectation-maximization.
• Reformulation of grid adjustment as a boundary-constrained linear least squares optimization.

Main Results:

• The proposed method effectively overcomes grid constraints for off-grid source localization.
• Incorporation of Bernoulli-Gaussian priors enhances sparsity without prior information.
• Demonstrated superior performance in localization success rate and sidelobe suppression over conventional methods.

Conclusions:

• The grid-adaptive model and off-grid algorithm provide precise source localization by addressing basis mismatch.
• The method offers significant improvements compared to traditional Bartlett and sparse Bayesian learning processors.
• Validated through numerical simulations and SwellEX-96 experimental data.