Time-Domain Off-Grid Sparse Bayesian Learning for High-Resolution Time-Delay Estimation

When estimating the acoustic channel impulse response (CIR) at low frequencies, matched filter (MF)-based methods suffer from limited resolution due to the narrow source bandwidth. To address this limitation, we propose a time-domain sparse Bayesian learning (SBL) framework that enables CIR estimation independently of the source frequency band. Unlike existing frequency-domain SBL approaches, the proposed method formulates the estimation problem as a time-domain linear system and incorporates off-grid modeling along with multiple measurements to achieve high-resolution results. The performance of the method is evaluated through simulations across varying signal-to-noise ratio (SNR) and number of measurements, and compared against conventional MF and on-grid SBL techniques. Its effectiveness is further validated using real-world experimental data collected in a shallow water environment. The results demonstrate that the proposed framework significantly improves temporal resolution, enhances denoising performance, and successfully resolves closely spaced arrivals that are otherwise indistinguishable.