[Paper Review] MULAN: A Blind and Off-Grid Method for Multichannel Echo Retrieval

Helena Peić Tukuljac, Antoine Deleforge+1 more|arXiv (Cornell University)|Oct 31, 2018

Blind Source Separation Techniques3 citations

TL;DR

This paper proposes MULAN, a novel blind and off-grid method for multichannel echo retrieval that directly estimates echo locations and weights in continuous time, bypassing discrete-time filter estimation. By leveraging finite-rate-of-innovation (FRI) sampling, MULAN achieves near-exact recovery of echo parameters with precision several orders of magnitude higher than on-grid methods, especially in realistic off-grid scenarios where conventional methods fail.

ABSTRACT

This paper addresses the general problem of blind echo retrieval, i.e., given M sensors measuring in the discrete-time domain M mixtures of K delayed and attenuated copies of an unknown source signal, can the echo locations and weights be recovered? This problem has broad applications in fields such as sonars, seismol-ogy, ultrasounds or room acoustics. It belongs to the broader class of blind channel identification problems, which have been intensively studied in signal processing. Existing methods in the literature proceed in two steps: (i) blind estimation of sparse discrete-time filters and (ii) echo information retrieval by peak-picking on filters. The precision of these methods is fundamentally limited by the rate at which the signals are sampled: estimated echo locations are necessary on-grid, and since true locations never match the sampling grid, the weight estimation precision is impacted. This is the so-called basis-mismatch problem in compressed sensing. We propose a radically different approach to the problem, building on the framework of finite-rate-of-innovation sampling. The approach operates directly in the parameter-space of echo locations and weights, and enables near-exact blind and off-grid echo retrieval from discrete-time measurements. It is shown to outperform conventional methods by several orders of magnitude in precision.

Motivation & Objective

• To address the fundamental limitation of on-grid echo estimation in multichannel blind source identification, where sampling grid mismatch degrades precision.
• To enable blind recovery of echo locations and weights without assuming discrete-time filter models or fixed filter lengths.
• To develop a method that operates directly in the continuous-time parameter space of echo delays and amplitudes.
• To overcome the basis-mismatch problem in compressed sensing by avoiding discrete-time discretization of echo parameters.
• To demonstrate superior performance in off-grid echo retrieval compared to conventional on-grid methods like CR and LASSO.

Proposed method

• MULAN formulates the echo retrieval problem in the continuous-time domain using a finite-rate-of-innovation (FRI) sampling framework.
• It models the multichannel signals as convolutions of a band-limited source with sparse impulse train filters in continuous time.
• The method estimates echo parameters (delays and amplitudes) directly in the parameter space via a non-convex optimization problem based on frequency-domain measurements.
• It uses a multi-frequency sampling approach with an odd number of frequencies (e.g., 201 or 401) to ensure invertibility and stability of the reconstruction.
• The optimization is initialized with 20 random starts and solved using a nonlinear least-squares algorithm with convergence threshold of 0.1%.
• The method does not require prior knowledge of filter length and only assumes the number of echoes $K$.

Experimental results

Research questions

• RQ1 Can echo locations and weights be recovered with sub-sample accuracy from discrete-time multichannel measurements without assuming discrete-time filter models?
• RQ2 How does the performance of a blind, off-grid echo retrieval method compare to conventional on-grid approaches in realistic, off-grid scenarios?
• RQ3 What is the impact of the number of sensors $M$, number of echoes $K$, and number of frequency samples $F$ on recovery success rate?
• RQ4 Can the FRI sampling framework enable robust blind echo retrieval when the source signal is unknown and the filters are not constrained to finite length?
• RQ5 What is the achievable precision of echo parameter estimation when avoiding the basis-mismatch problem inherent in discrete-time compressed sensing?

Key findings

• In the on-grid scenario, MULAN achieved a 59% success rate in full echo location recovery, outperforming CR (92%) but matching its location accuracy, while reducing weight RMSE by two to three orders of magnitude.
• In the more realistic off-grid scenario, MULAN achieved 70% success in full echo location recovery, compared to only 1% for CR and 2% for LASSO, demonstrating a dramatic improvement in localization robustness.
• MULAN’s weight estimation error was 0.00048 in successful off-grid tests, which is approximately 70 times smaller than CR’s 0.0442 and 70 times smaller than LASSO’s 0.0346.
• The recovery rate for echo locations and weights increased with higher $F$ (number of frequency samples), while the number of sensors $M$ had negligible impact on performance.
• The method showed robustness to initialization, with success rates increasing when using more random initializations, though at the cost of higher computational load.
• The results confirm that MULAN enables near-exact blind echo retrieval in continuous time, avoiding the basis-mismatch problem and achieving sub-sample precision.