Introduction
Blind calibration is a critical task in sensor array processing, involving the estimation of unknown sensor gains and phase offsets without a known reference signal. This challenge is particularly acute for sparse arrays, such as nested arrays, which are favored for their ability to detect $O(N^2)$ sources using only $O(N)$ physical sensors. Despite their advantages, nested arrays have historically been difficult to calibrate blindly. Conventional wisdom suggested that Least-Squares (LS) methods were inapplicable due to insufficient unique equations. This paper introduces a novel, non-iterative approach that challenges this assumption, providing a highly efficient and accurate calibration framework.
Methodology: ML-based OWLS
The proposed method, termed Maximum Likelihood-based Optimally Weighted Least Squares (ML-based OWLS), leverages the "Partially-Toeplitz" structure of the signal covariance matrix. The authors identify that by slightly modifying the traditional nested array geometry—specifically setting the inter-element spacing factor $L=N_1$ instead of $L=N_1+1$—they can generate enough linear dependencies to solve for all unknown calibration parameters.
The algorithm follows three primary steps:
- Structural Exploitation: Utilizing the modified nested array design to ensure the system of equations is solvable.
- Log-Linear Transformation: Applying an element-wise logarithm to the sample covariance matrix to linearize the relationship between sensor offsets and received signals.
- Asymptotically Optimal Weighting: Deriving a weighting matrix ($\Lambda_\xi$) based on the statistical cross-correlations of estimation errors. This allows the estimator to prioritize more reliable equations, achieving the Best Linear Unbiased Estimate (BLUE).
Unlike iterative optimization techniques like Semidefinite Programming (SDP), ML-based OWLS provides a closed-form solution, ensuring deterministic and rapid computation.
Experimental Evaluation
The performance of the ML-based OWLS method was evaluated using synthetic simulations of an 8-sensor, 2-level nested array. The calibration phase involved 15 narrowband sources from random angles. The method was benchmarked against the state-of-the-art SDP-based approach.
Key Results
The experimental results demonstrate that ML-based OWLS significantly outperforms the SDP method in both accuracy and computational efficiency.
Accuracy
As shown in the MSE plots, the proposed method consistently achieves lower error rates for both gain and phase estimation across a wide range of Signal-to-Noise Ratios (SNR) and sample sizes ($T$). This superior performance is attributed to the optimal weighting scheme that accounts for the statistical coupling between gain and phase estimation errors.
Computational Efficiency
The most striking advantage of the ML-based OWLS method is its speed. Because it relies on a closed-form matrix calculation rather than iterative optimization, it is orders of magnitude faster than the SDP approach.
In a test with 8 sensors, the proposed method achieved a runtime of approximately 0.0032 seconds, compared to 1.616 seconds for the SDP method—a speedup of approximately 500x.
Conclusion
This research provides a breakthrough in the blind calibration of nested arrays. By debunking the myth that LS-based methods are unsuitable for these structures and introducing an asymptotically optimal weighting scheme, the authors have developed a calibration tool that is both more accurate and significantly faster than existing iterative solutions. Future work will focus on extending this framework to wideband signals and $K$-level nested array structures.
Summary updated