[Paper Review] Multi-Dimensional Wireless Tomography with Tensor-Based Compressed Sensing

Takemoto Kazushi, Takahiro Matsuda+4 more|arXiv (Cornell University)|Jul 9, 2014

Sparse and Compressive Sensing Techniques18 references3 citations

TL;DR

This paper proposes a tensor-based compressed sensing framework for multi-dimensional wireless tomography to estimate shadowing loss distributions with higher accuracy than conventional vector-based compressed sensing. By exploiting the low-rank structure of the loss field tensor instead of sparsity in the frequency domain, the method achieves superior reconstruction performance, especially in low-noise environments, as validated through simulations using synthetic and real-world data models.

ABSTRACT

Wireless tomography is a technique for inferring a physical environment within a monitored region by analyzing RF signals traversed across the region. In this paper, we consider wireless tomography in a two and higher dimensionally structured monitored region, and propose a multi-dimensional wireless tomography scheme based on compressed sensing to estimate a spatial distribution of shadowing loss in the monitored region. In order to estimate the spatial distribution, we consider two compressed sensing frameworks: vector-based compressed sensing and tensor-based compressed sensing. When the shadowing loss has a high spatial correlation in the monitored region, the spatial distribution has a sparsity in its frequency domain. Existing wireless tomography schemes are based on the vector-based compressed sensing and estimates the distribution by utilizing the sparsity. On the other hand, the proposed scheme is based on the tensor-based compressed sensing, which estimates the distribution by utilizing its low-rank property. We reveal that the tensor-based compressed sensing has a potential for highly accurate estimation as compared with the vector-based compressed sensing.

Motivation & Objective

• To address the limitations of vector-based compressed sensing in multi-dimensional wireless tomography, which relies on sparsity in the frequency domain and may suffer from suboptimal reconstruction in correlated environments.
• To explore the potential of tensor-based compressed sensing—leveraging low-rank structure of the loss field tensor—for improved estimation accuracy in higher-dimensional monitored regions.
• To develop a novel multi-dimensional wireless tomography scheme that enables accurate localization of internal obstructions using fewer measurements.
• To validate the superiority of tensor-based recovery over vector-based recovery in terms of reconstruction error and robustness under varying noise conditions.

Proposed method

• Formulates the wireless tomography problem as a multi-dimensional inverse problem where the shadowing loss field is modeled as a tensor of order D ≥ 2.
• Applies tensor-based compressed sensing using Higher-Order Singular Value Decomposition (HOSVD) to exploit the low-rank property of the loss field tensor.
• Reconstructs the loss field tensor by retaining only the dominant HOSVD components, effectively approximating the true tensor with minimal reconstruction error.
• Compares the tensor recovery scheme with vector-based compressed sensing, where the loss field is vectorized and sparsity in the DFT domain is exploited.
• Uses a measurement model based on path loss between wireless nodes on the boundary, forming a linear system Y = A × X, where X is the loss field tensor.
• Employs a rank-constrained optimization approach to estimate the tensor by minimizing the Frobenius norm of the reconstruction error under low-rank constraints.

Experimental results

Research questions

• RQ1 Can tensor-based compressed sensing outperform vector-based compressed sensing in reconstructing multi-dimensional shadowing loss fields in wireless tomography?
• RQ2 How does the low-rank structure of the loss field tensor influence estimation accuracy compared to sparsity in the frequency domain?
• RQ3 What is the impact of noise level on the performance gap between tensor-based and vector-based recovery schemes?
• RQ4 How does the dimensionality of the monitored region affect the reconstruction accuracy of the proposed tensor-based method?

Key findings

• The tensor-based compressed sensing approach achieves lower reconstruction error than vector-based compressed sensing, particularly in low-noise environments, due to the more effective exploitation of the loss field's low-rank structure.
• For a two-dimensional loss field with rank 1, the tensor recovery scheme achieves a reconstruction error of approximately 0.32, compared to 2.92 for the vector-based method, demonstrating a significant improvement.
• The estimated singular values from the tensor recovery closely match the true values, with σ₁ ≈ 30.00 and σ̂₁ ≈ 29.68, indicating high fidelity in low-rank approximation.
• The tensor recovery method outperforms vector recovery in terms of Frobenius norm error, with κ(M) ≤ κ(V), confirming theoretical advantages of low-rank approximation over sparsity-based recovery.
• The simulation results show that the tensor-based method maintains superior accuracy even with reduced measurement counts, highlighting its efficiency in data-scarce scenarios.