A multilinear direction finding (DF) approach for a sensor-array with multiple scales of invariance

Sébastian Miron, Yang Song, David Brie, Kainam Thomas WongHAL 2013

TL;DR

This paper introduces a novel direction-finding (DF) algorithm for sensor arrays with multiple spatial invariances, leveraging a multi-dimensional Candecomp/Parafac (CP) model. The proposed two-stage method outperforms existing ESPRIT-based approaches, especially in scenarios with limited snapshots or correlated sources, and can even handle single-snapshot cases.

What Problem Does This Paper Address?

Accurately determining the direction of arrival (DOA) of signals is crucial in many applications, such as radar, sonar, and wireless communications. Traditional high-resolution DF techniques like MUSIC and ESPRIT have been successful, but they often rely on specific array geometries and can struggle with correlated sources or limited data. This paper addresses the challenge of DOA estimation in sensor arrays that possess multiple levels of spatial invariance, meaning the array structure repeats itself at different scales. This multi-scale invariance offers a rich structure that can be exploited for improved DOA estimation, but existing methods are not well-suited to fully leverage this property. The authors aim to develop a DF algorithm that can effectively utilize this multi-scale invariance, generalize existing CP-based methods, and offer superior performance, particularly in challenging scenarios.

Methods

The core of the proposed method lies in modeling the data from a multi-scale sensor array as a multidimensional Candecomp/Parafac (CP) model. This model captures the inherent structure arising from the multiple scales of invariance. The authors then propose a two-stage algorithm for DOA estimation:

  1. CP Decomposition: The first stage involves fitting the CP model to the collected data. This is achieved using an Alternating Least Squares (ALS) procedure, or more specifically, the COMFAC approach, which recursively estimates the components of the CP model. This stage effectively extracts the steering vectors for each source at each scale level.
  2. Sequential DOA Estimation: The second stage takes the estimated steering vectors from the first stage and uses them to estimate the source direction-cosines. This is done by minimizing a cost function that aggregates the errors across all scale levels. The minimization is performed sequentially, starting from the lowest scale level (which provides unambiguous but high-variance estimates) and progressively moving to higher levels. This graduated approach helps to overcome the non-convexity of the cost function and find the global minimum, leading to more accurate and unambiguous DOA estimates.

Datasets

Not discussed in the paper. The paper focuses on a theoretical framework and simulation-based evaluation.

Key Insights and Ideas

  • CP Model for Multi-Scale Arrays: The paper demonstrates that data from multi-scale sensor arrays can be elegantly represented by a multidimensional CP model, effectively capturing the spatial invariance properties.
  • Two-Stage Estimation for Robustness: The two-stage approach (CP decomposition followed by sequential DOA estimation) is key. The first stage leverages the CP structure to extract source information, while the second stage uses a clever sequential minimization to achieve unambiguous and accurate DOA estimates, even in the presence of ambiguities at higher scales.
  • Handling Correlated Sources: Unlike some ESPRIT-based methods that require uncorrelated sources, the CP model naturally accommodates correlated sources, leading to better performance in such scenarios.
  • Single-Snapshot Capability: The method's ability to handle single-snapshot data is a significant advantage, making it applicable in scenarios where data acquisition is limited. This is achieved by the inherent structure of the CP model and the sequential estimation process.
  • Generalization of Existing Work: The proposed approach extends previous CP-based DOA estimation techniques to arrays with an arbitrary number of spatial invariances, offering a more general framework.

Key Results

  • Outperformance of ESPRIT: Monte Carlo simulations show that the proposed multilinear DF approach consistently outperforms the ESPRIT-based method from Wong and Zoltowski [2] in terms of root mean-square error (RMSE).
  • Improved Performance with Limited Snapshots: The proposed algorithm shows a significant advantage over ESPRIT when the number of snapshots is small (e.g., K < 11).
  • Robustness to Correlated Sources: The method maintains superior performance even when sources are correlated, a scenario where ESPRIT struggles due to its underlying assumptions.
  • Single-Snapshot Success: In the single-snapshot scenario, the proposed method yields fair results, while ESPRIT becomes unusable.
  • Identifiability Conditions: The paper analyzes the identifiability conditions for the CP model, showing that unique estimation of DOAs is achievable under mild conditions, often met in practical applications.

Limitations and Future Work

  • Computational Complexity: While powerful CP fitting algorithms have improved convergence speed, the computational burden of the proposed method can still be higher than ESPRIT in some cases. However, the paper notes that closed-form solutions for CP decomposition exist for specific data structures, which can match ESPRIT's complexity.
  • Number of Estimable Sources: The proposed method, while more accurate, can estimate a smaller number of sources compared to ESPRIT for certain array configurations.
  • Future Work: The paper suggests that further research could explore closed-form solutions for CP decomposition in the context of Vandermonde structured data, which is common in array processing. Additionally, investigating the specific identifiability conditions for fully coherent or collocated sources could be a valuable direction. The paper also hints at the potential for extending this approach to non-rectangular array configurations.

Summary updated

Abstract

In this paper, we introduce a novel direction finding algorithm for a multi-scale sensor-array, that is, an array presenting multiple scales of invariance. We show that the collected data can be represented as a Candecomp/Parafac (CP) model, for which we analyze the identifiability properties. A two-stage algorithm for direction-of-arrival (DOA) estimation with such an array is also proposed. This approach generalizes the results given in [Sidiropoulos et al., 2000] to an array that presents an arbitrary number of spatial invariances. We illustrate, on a particular array geometry, that our method outperforms the ESPRIT-based approach introduced in [Wong and Zoltowski, 1998]. Moreover, we show that the single-snapshot case can be handled by our method, provided that the array includes at least three scale-levels.
BibTeX
@inproceedings{miron2013multilinear,
  author = {Sébastian Miron and Yang Song and David Brie and Kainam Thomas Wong},
  title = {{A multilinear direction finding (DF) approach for a sensor-array with multiple scales of invariance}},
  booktitle = {HAL},
  year = {2013},
}

Loading related content...

Figure