whereη is the minimum requirement for the system input SNR.
2.7
Summary
This chapter has firstly reviewed the background theories of sensor array processing and beamforming techniques. Then, introductions to the existing work in the literatures on the conventional adaptive beamforming algorithms and robust adaptive beamforming tech- niques have been presented. Lastly, the problem of distributed beamforming for wireless communication systems and the existing approaches and techniques have been discussed. This chapter is provided as a background support to the rest of the chapters where sig- nificant improvements and developments as well novel techniques are proposed. In the following chapters, we firstly introduce novel cost-efficient robust adaptive beamform- ing methods that based on recursive shrinkage methods, cross-correlation exploitations, subspace projections and low-rank techniques. Then, distributed beamforming and relay selection methods and robust distributed beamforming techniques are proposed.
Low-Complexity Shrinkage-Based
Mismatch Estimation (LOCSME)
Algorithms for Robust Adaptive
Beamforming
Contents
3.1 Introduction . . . 35 3.2 System Model and Problem Statement . . . 37 3.3 Batch LOCSME Algorithm . . . 38 3.4 Stochastic Gradient LOCSME Type Algorithm . . . 42 3.5 Conjugate Gradient LOCSME Type Algorithms . . . 45 3.6 Performance Analysis . . . 51 3.7 Simulations . . . 57 3.8 Summary . . . 62
3.1
Introduction
Sensor array signal processing techniques and their applications to wireless communica- tions, sensor networks and radar have been widely investigated in recent years. Adaptive beamforming is one of the most important topics in sensor array signal processing which has applications in many fields. However, adaptive beamformers may suffer performance degradation due to small sample data size or the presence of the desired signal in the train- ing data. In practical environments, desired signal steering vector mismatch problems like signal pointing errors [18], imprecise knowledge of the antenna array, look-direction mis- match or local scattering may even lead to more significant performance loss [7].
3.1.1
Prior and Related Work
In order to address these problems, robust adaptive beamforming (RAB) techniques have been developed in recent years. Popular approaches include worst-case optimization [7], diagonal loading [8, 9, 37], and eigen-decomposition [18, 19]. However, general RAB designs have some limitations such as their ad hoc nature, high probability of subspace swap at low SNR and high computational cost [11].
Further recent works have looked at approaches based on combined estimation proce- dures for both the steering vector mismatch and interference-plus-noise covariance (INC) matrix to improve RAB performance. The worst-case optimization methods in [7, 33–35] solve an online semi-definite programming (SDP) while using a matrix inversion to esti- mate the INC matrix. The method in [12] estimates the steering vector mismatch by solv- ing an online sequential quadratic program (SQP) [12], while estimating the INC matrix using a shrinkage method [12]. Another similar method which jointly estimates the steer- ing vector using SQP and the INC matrix using a covariance reconstruction method [15], presents outstanding performance compared to other RAB techniques. However, their main disadvantages include the high computational cost associated with online optimiza- tion programming, the matrix inversion or reconstruction process, and slow convergence.
3.1.2
Contributions
In this chapter, we develop an RAB algorithm with low complexity, which requires very little in terms of prior information, and has a superior performance to previously reported RAB algorithms. Our technique estimates the steering vector using a low-complexity shrinkage-based mismatch estimation (LOCSME) algorithm [31]. LOCSME estimates the covariance matrix of the input data and the INC matrix using the oracle approximat- ing shrinkage (OAS) method. The only prior knowledge that LOCSME requires is the angular sector in which the desired signal steering vector lies. Given the sector, the sub- space projection matrix of this sector can be computed in very simple steps [11–13, 15]. In the first step, an extension of the OAS method [16] is employed to perform shrinkage estimation for both the cross-correlation vector between the received data and the beam- former output and the received data covariance matrix. LOCSME is then used to estimate the mismatched steering vector and does not involve any optimization program, which results in a lower computational complexity. In a further step, we estimate the desired signal power using the desired signal steering vector and the received data. As the last step, a strategy which subtracts the covariance matrix of the desired signal from the data covariance matrix estimated by OAS is proposed to obtain the INC matrix. The advan- tage of this approach is that it circumvents the use of direction finding techniques for the interferers, which are required to obtain the INC matrix.
Then, we develop a stochastic gradient (SG) adaptive version of the LOCSME tech- nique [31], denoted LOCSME-SG, which does not require matrix inversions or costly recursions to update the beamforming weights adaptively. In particular, the SCM is estimated only once using a knowledge-aided (KA) shrinkage [20, 32] algorithm along with the computation of the beamforming weights based on the estimated steering vector through SG recursions. Moreover, we also develop an adaptive LOCSME technique based on the conjugate gradient (CG) adaptive algorithm, resulting in CG type algorithms, de- noted LOCSME-CCG and LOCSME-MCG. Different from LOCSME-SG, the CG type algorithms not only updates the beamforming weights, but can also estimate the mis- matched steering vector, which sequentially performs the estimation of the mismatched vector by LOCSME in every snapshot. An analysis shows that both LOCSME-SG and LOCSME-CG achieve one degree lower complexity than the original LOCSME. Simu- lations also show an excellent performance which benefits from the precise estimation
provided by the shrinkage approach. Our contributions are summarized as follows:
• The derivation of LOCSME batch algorithm.
• The development of LOCSME type SG and CG algorithms.
• An investigation of the effect of shrinkage on the estimation accuracy of the algo- rithms.
• A study of the performance and the complexity of the proposed and existing algo- rithms.
This chapter is organized as follows. The system model and problem statement are de- scribed in Section 3.2. The derivation of the LOCSME algorithm and steering vector mis- match estimation are provided in Section 3.3. Section 3.4 presents the proposed adaptive LOCSME-SG algorithm whereas Section 3.5 presents the proposed LOCSME-CCG and LOCSME-MCG algorithms. Section 3.6 provides the shrinkage and complexity analyses. Section 3.7 presents the simulation results. Section 3.8 gives the summary.