Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Mismatch-Robust Underwater Acoustic Localization Using A Differentiable Modular Forward Model

Dariush Kari, Yongjie Zhuang, Andrew C. Singer

TL;DR

This work tackles underwater acoustic localization under environmental mismatch by fusing a differentiable, physics-inspired forward model with test-time adaptation. It introduces a modular forward architecture that learns multipath path lengths implicitly via a Path Length Network, while enabling end-to-end optimization without labeled paths. The authors formulate a gradient-based ML estimator and extend it with Domain Adaptive GBL (DA-GBL), which jointly optimizes the source location and forward-model weights at inference time, providing a bounded-error guarantee under small environmental perturbations. Empirical results show substantial robustness to small mismatches and illustrate the method’s limitations under larger mismatches, highlighting the approach as a data-efficient, adaptable solution for UWA localization in variable environments.

Abstract

In this paper, we study the underwater acoustic localization in the presence of environmental mismatch. Especially, we exploit a pre-trained neural network for the acoustic wave propagation in a gradient-based optimization framework to estimate the source location. To alleviate the effect of mismatch between the training data and the test data, we simultaneously optimize over the network weights at the inference time, and provide conditions under which this method is effective. Moreover, we introduce a physics-inspired modularity in the forward model that enables us to learn the path lengths of the multipath structure in an end-to-end training manner without access to the specific path labels. We investigate the validity of the assumptions in a simple yet illustrative environment model.

Mismatch-Robust Underwater Acoustic Localization Using A Differentiable Modular Forward Model

TL;DR

This work tackles underwater acoustic localization under environmental mismatch by fusing a differentiable, physics-inspired forward model with test-time adaptation. It introduces a modular forward architecture that learns multipath path lengths implicitly via a Path Length Network, while enabling end-to-end optimization without labeled paths. The authors formulate a gradient-based ML estimator and extend it with Domain Adaptive GBL (DA-GBL), which jointly optimizes the source location and forward-model weights at inference time, providing a bounded-error guarantee under small environmental perturbations. Empirical results show substantial robustness to small mismatches and illustrate the method’s limitations under larger mismatches, highlighting the approach as a data-efficient, adaptable solution for UWA localization in variable environments.

Abstract

In this paper, we study the underwater acoustic localization in the presence of environmental mismatch. Especially, we exploit a pre-trained neural network for the acoustic wave propagation in a gradient-based optimization framework to estimate the source location. To alleviate the effect of mismatch between the training data and the test data, we simultaneously optimize over the network weights at the inference time, and provide conditions under which this method is effective. Moreover, we introduce a physics-inspired modularity in the forward model that enables us to learn the path lengths of the multipath structure in an end-to-end training manner without access to the specific path labels. We investigate the validity of the assumptions in a simple yet illustrative environment model.

Paper Structure

This paper contains 6 sections, 2 theorems, 18 equations, 4 figures.

Table of Contents

  1. Introduction
  2. Gradient-Based Localization (GBL)
  3. Modular Network
  4. Adapting To The Environment
  5. Empirical Evaluation
  6. Conclusion

Key Result

Theorem 1

Suppose that $\mathbf{v}_0 = \Gamma(\mathbf{w}_{\mathsf{tr}}, \gamma, \hat{\mathbf{p}}_0, \mathcal{A}(\mathbf{e}_{\mathsf{tr}}, \mathbf{p}_s, .))$. If the Assumptions assm:represent, assm:Lipschitz, assm:convexity, and assm:continuity hold, for a given environmental parameter $\mathbf{e}_{\mathsf{tr

Figures (4)

  • Figure 1: The module for path length computation is a fully connected network. Each path $i$ is identified by its number of surface and bottom reflections, $N_s$ and $N_b$. Also, $\rho_i$ indicates the reflection coefficient experienced during the $i$-th path. Here, $\rho_i = (-1)^{N_s(i)}$
  • Figure 2: Block diagram of forward domain adaptation. During the adaptation, the loss minimization is performed with respect to both $\mathbf{p}$ and $\mathbf{w}$.
  • Figure 3: RMSE performance under different SNR values.
  • Figure 4: RMSE performance under different mismatch values. Training depth is $200$ m and SNR $= 20$ dB.

Theorems & Definitions (4)

  • Definition 1
  • Theorem 1: Perturbation in The Solution $\mathbf{v}_0$
  • Definition 2: Positive or Negative Vector
  • Corollary 1