TL;DR
This paper tackles the challenge of accurately estimating the direction of arrival (DOA) of signals while simultaneously identifying faulty sensors in an array. They propose a novel method that unifies these two tasks by recognizing the inherent connection between the low-rank structure of the signal and the row-sparse nature of sensor distortions, leading to improved performance over existing techniques.
What Problem Does This Paper Address?
Accurate direction-of-arrival (DOA) estimation is crucial in many applications like radar and wireless communications. However, real-world sensor arrays often suffer from distortions, such as gain and phase uncertainties, which can severely degrade the performance of traditional DOA estimation methods. While previous work has addressed this by treating DOA estimation (often modeled as a low-rank problem) and distorted sensor detection (often modeled as a row-sparse problem) separately, this approach can lead to performance losses. This paper aims to bridge this gap by jointly optimizing both aspects, recognizing that the low-rank signal component and the row-sparse sensor distortion component are not independent but rather intrinsically linked. They also incorporate prior knowledge about the maximum level of sensor distortion.
Methods
The core of the proposed method lies in an alternating optimization scheme that entangles the low-rank and row-sparse components. The problem is formulated as a joint optimization of a low-rank matrix (representing the clean signal) and a sparse vector (representing sensor distortions).
- Entangling Components: Instead of treating the low-rank and row-sparse components separately, the authors model the distorted observation as $Y = Z + \text{diag}\{\gamma\}Z + N$, where $Z$ is the low-rank clean signal matrix, $\gamma$ represents the sensor distortions, and $N$ is noise. This formulation explicitly links the two.
- Convex Relaxation: The original problem involves non-convex terms (rank and $l_0$ norm). These are replaced by their convex surrogates: the nuclear norm for rank and the $l_1$ norm for sparsity. A smoothing parameter $\mu$ is introduced to handle the non-smoothness of the nuclear norm.
- Alternating Optimization: The problem is then solved iteratively:
- Low-Rank Component Update (Z): With the distortion vector $\gamma$ fixed, the update for $Z$ has a closed-form solution derived from minimizing the Frobenius norm of the residual and the nuclear norm of $Z$.
- Sparse Component Update ($\gamma$): With the low-rank matrix $Z$ fixed, the update for $\gamma$ is formulated as a linearly constrained LASSO problem. This is then transformed into a quadratic programming (QP) problem, which can be efficiently solved using standard solvers.
- Distorted Sensor Detection: Once the distortion vector $\hat{\gamma}$ is estimated, the distorted sensors are identified by analyzing the magnitudes of its entries. A specific algorithm is proposed to determine the number and locations of these sensors based on the sorted magnitudes and a threshold.
Datasets
Not discussed in the paper. The evaluation is performed using simulated data.
Key Insights and Ideas
- Inherent Connection: The most significant insight is the explicit modeling of the inherent connection between the low-rank signal component and the row-sparse sensor distortion component. This is a departure from previous methods that treated them independently, leading to a more accurate representation of the system.
- Maximal Distortion Level: Incorporating prior knowledge about the maximal distortion level of sensors provides an important constraint that improves the robustness and accuracy of the detection process.
- Unified Framework: The paper demonstrates that DOA estimation and distorted sensor detection can be effectively solved within a single, unified optimization framework, rather than as separate, sequential problems.
- Closed-form for Z, QP for $\gamma$: The development of a closed-form solution for the low-rank component and a QP formulation for the sparse component makes the alternating optimization approach computationally tractable and efficient.
Key Results
- Improved Performance: The proposed method consistently outperforms existing state-of-the-art methods (including IRLS [16], ADMM [27], APG [26], and SVT [25]) in terms of Root Mean Squared Error (RMSE) and Resolution Probability (ResProb) for DOA estimation across various Signal-to-Noise Ratios (SNRs) and numbers of snapshots.
- Competitive Detection Rate: While ADMM [27] shows a slightly better success detection rate, the proposed method achieves competitive performance in identifying distorted sensors.
- Convergence: The algorithm converges, although the paper notes that the IRLS algorithm converges faster in terms of the number of iterations, but the proposed method achieves better overall performance.
- Computational Complexity: The computational complexity is dominated by matrix inversion and QP solving, resulting in $O(N_{\text{iter}}M^3)$, where $N_{\text{iter}}$ is the number of iterations and $M$ is the number of sensors.
Limitations and Future Work
- Convergence Speed: The paper acknowledges that the proposed algorithm converges slower than the IRLS algorithm in terms of iterations. Future work could focus on accelerating the convergence.
- Hyperparameter Tuning: While methods for selecting hyperparameters ($\lambda_1, \lambda_2, \mu$) are discussed (e.g., CV, 2D search), their optimal selection can still be a challenge and might impact performance.
- Assumptions: The method assumes narrowband signals and a uniform linear array (ULA). Extending it to wideband signals or different array geometries could be a future direction.
- Number of Sources: The DOA estimation assumes the number of sources ($K$) is known. Developing a method that can also estimate $K$ would be beneficial.
Summary updated