Off-grid fast relevance vector machine
algorithm for direction of arrival estimation
ISSN 1751-8784
Received on 16th January 2015
Revised on 6th September 2015
Accepted on 29th September 2015
doi: 10.1049/iet-rsn.2015.0304
www.ietdl.org
Jincheng Lin
✉
, Xiaochuan Ma, Shefeng Yan, Chengpeng Hao, Geping Lin
Key Laboratory of Information Technology for Autonomous Underwater Vehicles, Institute of Acoustics, Chinese Academy of Sciences,
Beijing 100190, People’s Republic of China
✉ E-mail: ljcmym@163.com
Abstract: Direction of arrival (DOA) estimation is a basic and important problem in signal processing and has been widely
applied. Its research has been advanced by the recently developed methods based on Bayesian comp ressiv e sensing
(BCS). Among these methods, the ones combined with an off-grid (OG) model have been proved to be more accurate
than the on-grid ones. However, the conventional BCS-based methods have a disadvantage of the slow speed. In this
study, a high-efficiency iterative algorithm, based on the fast relevance vector machine and the OG model, is
developed. This new approach applies to both the single- and multiple-snapshot cases. Numerical simulations show
that the proposed method estimates DOAs more accurately than the ℓ
1
-penalisation method and compute s more
efficiently than the conventional BCS-based methods. Finally, comparisons with state-of-the-art methods and Cramer–
Rao bound are also reported.
1 Introduction
Direction of arrival (DOA) estimation using sensor arrays is
important in radar, sonar, seismic systems, acoustic source
localisation, mobile communication etc. [1 , 2]. The narrowband
far-field source case is assumed in this paper, and the direction
information is to be estimated. There are a lot of high-resolution
DOA estimation algorithms such as MUSIC [3, 4] and ESPRIT
[5]. However, most of these methods need a covariance matrix
estimate, and rely on different prerequisites, for example, high
signal-to-noise ratio (SNR) level, large number of snapshots and
estimate of source number. In many practical applications, only a
very small number of snapshots, even a single snapshot, are
available for DOA estimation. Therefore, it is necessary to design
estimators for the scenario of few snapshots (even one).
Fortunately, a number of effective methods have been proposed
for DOA estimation with low number of snapshots, for example,
iterative adaptive approach (IAA)-based algorithms [6–10] and
compressive sensing (CS)-based algorithms [10–12].
In [6], an IAA for amplitude and phase estimation (IAA-APES) has
been proposed. The IAA-APES algorithm is an iterative and
non-parametric algorithm that provides an accurate and
high-efficiency estimate under severe snapshot limitations. The
Bayesian information criterion [13] is used in conjunction with
IAA-APES to give sparse results, which are usually assumed in
DOA estimation. Furthermore, a parametric relaxation-based cyclic
approach (RELAX) [14, 15] has also been used in
IAA-APES&RELAX [6] to further improve the estimation accuracy.
Actually, IAA-APES&RELAX is a high-efficiency off-grid (OG)
spectral estimation algorithm. The comparisons between different
OG algorithms will be demonstrated in our numerical simulations.
Recently, DOA estimation techniques have been advanced by the
CS [16] methods. In the case of a single snapshot for DOA
estimation, ℓ
1
-penalisation is a favourable approach to the sparse
signal recovery, because it does not depend on the sample
covariance matrix and sources’ correlativity, and can accurately
estimate multiple DOAs only in a single snapshot [11].
In the CS field, sparse Bayesian learning (SBL)/inference (SBI)
with the relevance vector machine (RVM) [17] has been another
popular method for the sparse signal recovery. The concept of
Bayesian CS (BCS) has been proposed in [18]. In BCS, the CS
inversion problem is formulated from an SBL perspective. The
analysis of RVM [19, 20] has proved that the RVM provides a
tighter approximation to the ℓ
0
-norm sparsity than the ℓ
1
-norm.
The approach using BCS in DOA estimation has been presented in
[21], including single-snapshot BCS [18] and multiple-snapshot BCS
(MT-BCS) [22]. Some RVM-based DOA estimation methods are
also researched in [23–30]. In the case of multiple snapshots, the
SBL approach for the ℓ
1
-SVD model [11] has been proposed in
[31]. These methods all show the accurate and sparse results of the
DOA estimation problem. Nevertheless, only the conventional
RVM has been used in these articles, which is involved in
inverting a large matrix and consumes a large amount of
computation. Compared with the conventional RVM algorithm, the
fast RVM algorithm developed in [32, 33] can compute more
efficiently. Through adding, deleting and re-estimation candidate
basis functions in each iteration, the fast RVM can monotonically
maximise the marginal likelihood and choose basis functions smarter.
Although the existing CS or BCS-based methods have shown their
outstanding performance in DOA estimation, there are still more or
less deviations when the actual DOAs are not on the sampling grid.
Not to be constrained on the sampling grid, OG methods for DOA
estimation are proposed in [34]. There are also some further
discussion about the OG method in [35–38]. In [31 ], an SBL
method based on the OG model for DOA estimation has been
proposed. This method is referred as OG-SBI. It can obtain a more
accurate estimation not restricted by the fixed sampling grid
(on-grid model) and an excellent sparsity as BCS.
Nevertheless, the OG-SBI algorithm is realised by the
conventional RVM, there is no advantage in computational
efficiency. In this paper, we propose a fast RVM algorithm using
the OG adjustment for both single- and MT DOA estimation. The
proposed algorithm has the advantages of both the accurate
estimation by OG model and the high-efficiency computation by
fast RVM. We refer to the proposed algorithm in this paper as OG
fast RVM (OG-FastRVM). Some state-of-the-art single-snapshot
algorithms and subspace-based algorithms are taken as
comparisons in numerical simulations. Through numerical
simulations, we show that OG-FastRVM has a smaller root mean
square error (RMSE) in comparison with the ℓ
1
-penalisation on a
fixed grid. The simulations also demonstrate the advantages of
OG-FastRVM compared with the state-of-the-art OG algorithms,
IET Radar, Sonar & Navigation
Research Article
IET Radar Sonar Navig., 2016, Vol. 10, Iss. 4, pp. 718–725
718
&
The Institution of Engineering and Technology 2016