DA-MUSIC

Citation for published version (APA):

Merkofer, J. P., Revach, G., Shlezinger, N., Routtenberg, T., & van Sloun, R. J. G. (2024). DA-MUSIC: Data-
Driven DoA Estimation via Deep Augmented MUSIC Algorithm. IEEE Transactions on Vehicular Technology,
73(2), 2771-2785. Article 10266765. https://doi.org/10.1109/TVT.2023.3320360

Document license:
TAVERNE

DOI:
10.1109/TVT.2023.3320360

Document status and date:

Published: 01/02/2024

Document Version:
Publisher’s PDF, also known as Version of Record (includes final page, issue and volume numbers)

Please check the document version of this publication:

• A submitted manuscript is the version of the article upon submission and before peer-review. There can be
important differences between the submitted version and the official published version of record. People
interested in the research are advised to contact the author for the final version of the publication, or visit the
DOI to the publisher's website.
• The final author version and the galley proof are versions of the publication after peer review.
• The final published version features the final layout of the paper including the volume, issue and page
numbers.
Link to publication

General rights
Copyright and moral rights for the publications made accessible in the public portal are retained by the authors and/or other copyright owners
and it is a condition of accessing publications that users recognise and abide by the legal requirements associated with these rights.

• Users may download and print one copy of any publication from the public portal for the purpose of private study or research.
• You may not further distribute the material or use it for any profit-making activity or commercial gain
• You may freely distribute the URL identifying the publication in the public portal.

If the publication is distributed under the terms of Article 25fa of the Dutch Copyright Act, indicated by the “Taverne” license above, please
follow below link for the End User Agreement:
www.tue.nl/taverne

Take down policy

If you believe that this document breaches copyright please contact us at:
[email protected]
providing details and we will investigate your claim.

Download date: 07. Sep. 2025

IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 73, NO. 2, FEBRUARY 2024 2771

DA-MUSIC: Data-Driven DoA Estimation via Deep

Augmented MUSIC Algorithm
Julian P. Merkofer , Student Member, IEEE, Guy Revach , Graduate Student Member, IEEE,
Nir Shlezinger , Senior Member, IEEE, Tirza Routtenberg , Senior Member, IEEE, and Ruud J. G. van
Sloun , Member, IEEE

Abstract—Direction of arrival (DoA) estimation of multiple arrival (DoA) estimation of multiple, broadband, and possibly
signals is pivotal in sensor array signal processing. A popular coherent signal sources plays a key role in a wide range of
multi-signal DoA estimation method is the multiple signal clas- applications, including radar, communications, image analysis,
sification (MUSIC) algorithm, which enables high-performance
super-resolution DoA recovery while being highly applicable in and speech enhancement [2], [3], [4]. Over the last decades,
practice. MUSIC is a model-based algorithm, relying on an ac- a multitude of different Direction of arrival (DoA) estimation
curate mathematical description of the relationship between the algorithms have been proposed, and the problem is an active
signals and the measurements and assumptions on the signals area of research [5], [6].
themselves (non-coherent, narrowband sources). As such, it is A leading scheme employed in many DoA applications
sensitive to model imperfections. In this work, we propose to
overcome these limitations of MUSIC by augmenting the algo- is the popular multiple signal classification (MUSIC) algo-
rithm with specifically designed neural architectures. Our pro- rithm [7], which can provide asymptotically unbiased estimates
posed deep augmented MUSIC (DA-MUSIC) algorithm is thus a of the number of incident wavefronts present, their approxi-
hybrid model-based/data-driven DoA estimator, which leverages mate frequencies, and their DoAs. multiple signal classification
data to improve performance and robustness while preserving the (MUSIC) and other classic approaches, e.g., conventional beam-
interpretable flow of the classic method. DA-MUSIC is shown to
learn to overcome limitations of the purely model-based method, forming [8] and MVDR beamforming [9], are based on knowl-
such as its inability to successfully localize coherent sources as edge of the underlying statistical model; this dependency in-
well as estimate the number of coherent signal sources present. duces several key limitations. Among the drawbacks of the
We further demonstrate the superior resolution of the DA-MUSIC model-based (MB) approaches is the inherent limitation of the
algorithm in synthetic narrowband and broadband scenarios as signal model from which they are derived, which considers
well as with real-world data of DoA estimation from seismic signals.
narrowband signals. This results, for example, in the inability to
Index Terms—DoA estimation, MUSIC, model-based deep consistently estimate the DoA of correlated (coherent) signals,
learning. as well as a failure to resolve closely spaced signals with an
insufficient number of samples or a low signal-to-noise ratio
I. INTRODUCTION (SNR) [5], [7].
OURCE separation, localization, and tracking are crucial To extend narrowband models to broadband DoA estima-
S tasks in sensor array processing. In particular, Direction of tion, various extensions and alternative approaches have been
explored [10], because when the impinging signals are broad-
band, multiple frequency ranges can carry different information
Manuscript received 11 January 2023; revised 14 August 2023; accepted 7
September 2023. Date of publication 28 September 2023; date of current version regarding the DoA angles at hand. Generally, these extensions
13 February 2024. This work was supported by the Spectralligence, EUREKA IA from narrowband to broadband can be subdivided into coher-
Call, ITEA4 under Project 20209. The work of Tirza Routtenberg was supported ent and incoherent processing methods. In coherent processing
by the PAZY Foundation, Israel. The review of this article was coordinated by
Prof. Zhiguo Shi. An earlier version of this work was presented at the 2022 methods, the covariance matrices of the observations in different
IEEE International Conference on Acoustics, Speech, and Signal Processing frequency bins are coherently combined using certain transfor-
(ICASSP) as the paper [DOI: 10.1109/ICASSP43922.2022.9746637]. (Corre- mation matrices. Most coherent techniques are based on the
sponding author: Julian P. Merkofer.)
Julian P. Merkofer was with the ETH Zürich, 8092 Zürich, Switzerland. He is coherent signal subspace (CSS) concept [11], which focuses
now with the Electrical Engineering Department, Eindhoven University of Tech- the transformations of the covariances at different frequencies
nology, 5612 AZ Eindhoven, The Netherlands (e-mail: [email protected]). into a surrogate narrowband model, yet utilizes different trans-
Guy Revach is with the D-ITET, ETH Zürich, 8092 Zürich, Switzerland
(e-mail: [email protected]). formation matrices. Incoherent broadband processing methods
Nir Shlezinger is with the School of ECE, Ben-Gurion University of the combine DoA estimates obtained separately for each frequency
Negev, Beer Sheva 8499000, Israel (e-mail: [email protected]). bin [12].
Tirza Routtenberg is with the School of ECE, Ben-Gurion University of
the Negev, Beer Sheva 8499000, Israel, and also with the ECE Department, The recent success of data-driven (DD) deep learning (DL)
Princeton University, Princeton, NJ 08540 USA (e-mail: [email protected]). across a wide range of disciplines gave rise to neural net-
Ruud J. G. van Sloun is with the Electrical Engineering Department, Eind- work (NN)-aided DoA estimators [13]. The works [14], [15],
hoven University of Technology, 5612 AZ Eindhoven, The Netherlands (e-mail:
[email protected]). [16], [17] implemented model-agnostic DoA estimation us-
Digital Object Identifier 10.1109/TVT.2023.3320360 ing a dense NN, a convolutional neural network (CNN), a

0018-9545 © 2023 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See https://www.ieee.org/publications/rights/index.html for more information.

Authorized licensed use limited to: Eindhoven University of Technology. Downloaded on February 14,2024 at 14:56:09 UTC from IEEE Xplore. Restrictions apply.
2772 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 73, NO. 2, FEBRUARY 2024

sparse-connected CNN, and a U-Net architecture, respectively. data to enable MUSIC to successfully operate in such scenarios,
Other methods [18], [19] utilize the knowledge of the spatial which is the focus of this work.
covariance matrices, where [18] trained a CNN based on a We introduced deep augmented multiple signal classifica-
classification task (allowing it to also operate with unknown tion (DA-MUSIC) in [1], a hybrid MB/DD implementation of
number of sources) and [19] analyzed purely data-driven (DD) MUSIC, which exploits the structure of the classic subspace
estimators and DD methods in combination with maximum like- algorithm while augmenting it with NNs to learn to enhance its
lihood estimation (MLE). The works [20], [21] also combined performance. The proposed architecture overcomes the funda-
DL with MLE by using multiple dense NNs and a ResNet, mental limitations of MUSIC, enabling it to accurately detect
respectively, to estimate a subset of candidate angles. [22] ex- the locations of coherent sources. Our design builds upon the
tends the architecture of [21] to estimate the number of sources insight that the sensitivity to model mismatch and inability to
with the ResNet. While such black-box NNs enable handling handle coherent and broadband sources of MUSIC lie in its
array imperfections due to their model-agnostic nature, they estimation of the noise and signal subspaces through an EVD
involve highly parameterized models that may be computation- of the empirical covariance matrix. Accordingly, the proposed
ally intensive and lack the interpretability of model-based (MB) DA-MUSIC improves this crucial step by obtaining a surrogate,
methods. pseudo covariance matrix through a recurrent neural network
Alternatively, NNs were used to robustify the MB MUSIC as a (RNN) from the measurements, which is learned along with a
form of hybrid MB/DD system [23], [24], [25]. Specifically, the NN that acts as a peak finder. This work presents crucial exten-
work [26] and [27] proposed to estimate the discretized MUSIC sions to DA-MUSIC allowing it to operate with an unknown and
spectrum from the (spatially smoothed) covariance matrix of the varying number of signal sources by successfully estimating the
measurements through the utilization of multiple convolutional number of sources for non-coherent and coherent signals. An
NNs. While these methods are more robust to model inaccura- additional neural augmentation allows the categorization into
cies compared with the original MUSIC algorithm, utilizing the noise and signal eigenvectors to be obtained in a learned fashion
MUSIC spectrum as a label for training causes them to experi- by training a classifier. Numerical results show an estimation
ence the same drawbacks as their MB counterpart. Another DD accuracy of 98% in comparison to MUSIC with 89% accuracy
approach proposed in [28] considered systems with subarray in the same non-coherent scenario. We further demonstrate DA-
sampling and uses NNs to obtain a single estimated covariance MUSIC’s capabilities in various SNR domains, in three different
matrix from incoherent subarray measurements. This NN-aided broadband scenarios, as well as with real-world seismic signals.
estimate is utilized for DoA recovery via the subspace-based DA-MUSIC not only manages to focus frequency components
MUSIC algorithm. The method addresses the fundamental de- but is also able to concurrently focus an interdependent elevation
pendency of MUSIC on the estimated covariance matrix, thereby angle to receive a stable azimuth estimation.
robustifying the MUSIC algorithm, yet it does not fully exploit The rest of the article is organized as follows: Section II
the NNs’ ability to improve MUSIC, as the NN is trained using describes the assumed system model and surveys some of the
the true covariance matrix as a label, without considering its related work of DoA estimation, as well as the technical de-
downstream task. Furthermore, the work [29] proposed a hybrid tails of the MB MUSIC algorithm; Section III introduces the
MB/DD approach that includes an eigenvalue decomposition DA-MUSIC algorithm; Section IV presents the results of the
(EVD) of full-row Toeplitz matrices reconstruction (FTMR) simulations; and Section V provides concluding remarks.
matrices [30] and classification deep NNs for estimation of Throughout the article, we use boldface lower-case letters
a MUSIC-like spectrum and detection of number of sources for vectors; e.g., x, and boldface uppercase letters for matrices,
present via eigenvectors and eigenvalues e.g., for X. The ith entry of a vector x is denoted by xi and
The works [31], [32] applied CNNs for broadband DoA entries separated by commas represent column vectors. We use
estimation of a single source. The two approaches, however, calligraphic letters to denote the discrete-time Fourier transform,
utilize very different preprocessing methods and are trained e.g., X (ω) is the frequency representation of a signal x(t). We
as classification and regression tasks, respectively. While the use (·) and (·)H to denote the transpose and Hermitian operators
classification approach limits itself to a fixed resolution, the respectively, and  ·  and  · F to denote the 2 and Forbenius
regressor of [32] outputs the DoA angles directly. Therefore, it norms respectively. Further, U and CN represent the uniform
can achieve an arbitrary resolution, yet it is limited to the spatial and the complex normal distributions. The symbol I denotes the
covariance matrix as input. Another approach [33] first decom- identity matrix.
posed the broadband signals into non-overlapping narrowband
components and used support vector regression to obtain the
DoA. II. SYSTEM MODEL AND PRELIMINARIES
Accordingly, there is a concrete need for DoA estimation In this section, we detail the system model for which we
algorithms that can resolve multiple signals that are possibly develop the proposed DA-MUSIC algorithm in Section III. To
broadband and coherent in a low complexity and interpretable that aim, we first discuss the signal model in Section II-A. Then,
manner. The aforementioned individual limitations of both ex- we formulate the DoA estimation problem in Section II-B, and
isting MB and DD DoA estimation algorithms thus motivate the survey some of the related literature in Section II-C. Since our
derivation of a NN-aided DoA estimator capable of leveraging proposed solution builds upon the MUSIC algorithm, we recall

Authorized licensed use limited to: Eindhoven University of Technology. Downloaded on February 14,2024 at 14:56:09 UTC from IEEE Xplore. Restrictions apply.
MERKOFER et al.: DA-MUSIC: DATA-DRIVEN DOA ESTIMATION VIA DEEP AUGMENTED MUSIC ALGORITHM 2773

MUSIC in Section II-D and briefly discuss the CSS method for
broadband extension in Section II-E.

A. Signal Model
We distinguish the considered signal model into two cases: the
conventional narrowband setting and its more general broadband
setup. For more details regarding the two models, we refer the
reader to the textbooks [10], [34], and [35].
1) Narrowband: The most common setup in the array signal
processing literature concerning DoA estimation is the narrow- Fig. 1. DoA estimation illustration.
band setting. Here, the time it takes for the waves to propagate
the array is assumed to be negligible such that the occurring which can, in turn, be written in vector-matrix form analogous
delays are sufficiently small. Therefore, the signal measurement to the narrowband system model (1) as follows:
model for an arbitrary array structure consisting of M sensor
elements measuring D impinging narrowband signals takes the X (ω) = A(ω, θ) S(ω) + V(ω). (7)
following form:
For a ULA, the broadband steering vectors are defined for
x(t) = A(θ) s(t) + v(t). (1) frequency ω and angle of interest ψ as
 
In (1), the measurements x(t) ∈ CM at time instance t depend on Δm Δm
a(ω, ψ) = 1, e−jω c sin ψ , . . . , e−jω(M −1) c sin ψ , (8)
the signals s(t) ∈ CD , which originate from the unknown angles
θ = [θ1 , . . . , θD ], while v(t) ∈ CM is additive white Gaussian where c is the propagation velocity.
noise. The matrix A(θ) ∈ CM ×D contains the steering vectors
{a(θd )}, i.e., B. Problem Formulation
 
A(θ) = a(θ1 ) . . . a(θD ) , (2) DoA estimation is concerned with localizing signal sources
by determining their angles of incidence utilizing the measure-
where {θd } are the source directions. For example, the steering ments of an array aperture [34]. Formally, this corresponds to
vector a(ψ) for a uniform linear array (ULA) with an element estimating the angles θ = [θ1 , . . . , θD ] from the measurements
spacing of Δm = /2, where  is the wavelength of the signals, x(t) = [x1 (t), . . . , xM (t)] measured at multiple time instances
is defined for direction ψ as t ∈ {1, . . ., T }. An illustration of such a system is depicted
  in Fig. 1. The following additional assumptions are imposed
a(ψ) = 1, e−jπ sin ψ , . . . , e−jπ(M −1) sin ψ . (3) upon the observation model as well as the derived synthetic
data. The signals are uncorrelated to the noise (though signals
Consequently, the steering vectors (and, in turn, the matrix may possibly be correlated with each other) and generated in
A(θ)) specify the underlying array geometry. The collection the far-field region. And there is uniform propagation in all
of the measurements at the array elements over multiple time directions in an isotropic and non-dispersive medium. We further
instances is defined as assume that we have knowledge (though possibly mismatched)
 
X = x(1) . . . x(T ) , (4) of the underlying array geometry, implying that we can compute
an approximation of a(ψ) or a(ω, ψ). We consider a data-driven
where T is referred to as the number of snapshots. setting where we have access to training data. This data is a set
2) Broadband: In practice, many signals are broadband, and of U pairs, {(Xu , θ u )}U u=1 , each comprising the observations
the delay caused by propagating the array aperture needs to be and DoA angles from where the signals originated. In many
incorporated into the signal model. The following notation mod- scenarios; e.g., in wireless communications, a specific training
els the measurements received at array element m ∈ {1, . . ., M } set can be developed before deployment. If the ground-truth
DoAs are not obtainable at all, the data-driven DoA estimator
must be trained by utilizing synthetic data that closely describes

D
xm (t) = sd (t − τmd ) + vm (t), (5) the real signals. Our goal is thus to leverage the available domain
d=1 knowledge and data to design a system for recovering the DoAs
θ from a corresponding observation matrix X whose columns
where τmd represents the time delay of the mth array element are T snapshots of the measured waveforms at the M sensors.
measuring source d ∈ {1, . . .., D} compared to the measure-
ment at the reference element m = 1. Transformed to the fre-
C. Related Literature
quency domain, the relationship in (5) at frequency ω becomes
We next provide an overview of the relevant DoA esti-
mation methods for the presented numerical evaluations of

D
Xm (ω) = e−jωτmd Sd (ω) + Vm (ω), (6) Section IV, followed by reviewing relevant extensions. An ex-
d=1 tensive overview of various MB DoA estimators can be found

Authorized licensed use limited to: Eindhoven University of Technology. Downloaded on February 14,2024 at 14:56:09 UTC from IEEE Xplore. Restrictions apply.
2774 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 73, NO. 2, FEBRUARY 2024

Fig. 2. Block diagram of the MUSIC algorithm.

in [5], [6], and a recent literature review of DD DoA estimation These methods, however, experience significant bias with larger
approaches can be found in [36]. angular sectors of interest and can impose additional model
1) MB Narrowband Estimators: The conventional beam- assumptions such as noise statistics.
former (i.e. maximizing the steered response) is a basic ap- 3) DD Estimators: Inspired by the dramatic success of DL in
proach to DoA estimation [37]. An alternative family of DoA computer vision and natural language processing, recent years
estimators is based on subspace methods, which aim at recov- have witnessed a growing interest in the application of NNs for
ering the DoAs by identifying the noise and signal subspaces. DD DoA estimation. A recent review of DL DoA estimation
The MUSIC algorithm [7] is a highly popular subspace-based approaches can be found in [36] and essential methods were
method and has been researched extensively. Being the focus of covered in Section I. DL architectures presented in the numer-
this article, more information and a detailed description of the ical evaluations or otherwise closely related to this work are
algorithm can be found in Section II-D. Extensions of MUSIC summarized and discussed below.
include Root-MUSIC [38], a polynomial-rooting version, as The work [18] implemented a CNN architecture for DD DoA
well as spatially smoothed MUSIC, which removes the cor- estimation based on a multi-label classification task. Particularly,
relation between the incident signals by dividing the receiver the method takes real, imaginary, and phase information of
array into overlapping subarrays [39]. In practice, however, the the sample covariance as a three-channel input and predicts a
number of coherent sources is mostly unknown, and therefore probability grid of directions. Given the binary cross-entropy
the decorrelation effect of spatial smoothing is not obvious [40]. loss for training, the method is capable of operating with a
Another popular subspace-based method for DoA estimation is potentially unknown and varying number of sources. However,
estimation of signal parameters via rotational invariance tech- with a growing number of sources, the number of classes grows
niques (ESPRIT) [41] and its variations [42]. These methods rely exponentially.
heavily on the accuracy of the underlying model assumptions, Similarly to the above, [26] proposed CNNs for DoA esti-
are generally sensitive to array aperture perturbations, and are mation by taking the same channels of the sample covariance
inherently derived for narrowband signals. matrix as input. However, the CNNs are trained as a regression
2) MB Broadband Estimators: Broadband DoA estimation task estimating a discretized segment of the MUSIC spectrum.
algorithms can be categorized into two groups: incoherent meth- Thereby, they not only inherit the drawbacks and imitations of
ods and coherent methods. Generally, incoherent methods use MUSIC but are, besides increased robustness, dependant on the
independent frequency bins (IFBs) to process the DoA informa- underlying model assumptions.
tion at every frequency separately and then combine the results
of these narrowband DoA estimations. There are many different
variations in the implementation of this approach [43], [44], [45],
which typically vary in the computation of the covariance matri- D. MUSIC Algorithm
ces. Unfortunately, the computational complexity increases with As our proposed DA-MUSIC algorithm originates from the
each frequency bin. MUSIC algorithm, we next present this method in detail. MU-
Conversely, coherent methods combine the covariance matri- SIC, originally proposed by Schmidt in [7], considers the nar-
ces estimated for IFBs in order to apply narrowband techniques rowband signal model with incoherent sources (1), where the
directly over a single, so-called focused covariance matrix. An signals in s(t) are mutually independent. Fig. 2 visualizes a
overview of coherent techniques can be found in [10], [46]. simple outline of the MUSIC structure as a block diagram. The
A leading approach for coherent broadband DoA estimation approach takes the empirical covariance matrix of the received
is based on the CSS method [11], with example estimators measurements X, then conducts an EVD, followed by catego-
given in [47], [48], [49]. Many variations of the CSS method rizing the eigenvectors into signal and noise subspaces. The
are concerned with the crucial aspect of the focusing strategy, orthogonality between the two subspaces allows the formulation
proposing different techniques to combine the covariances at of a spatial spectrum, which contains peaks at DoA angles.
each IFB to obtain a faithful narrowband formulation. See, 1) Formal Derivation: With the assumption of the signal and
for example, [50], [51] and the more recent summary [52]. the noise being uncorrelated, the covariance matrix of x(t) is

Authorized licensed use limited to: Eindhoven University of Technology. Downloaded on February 14,2024 at 14:56:09 UTC from IEEE Xplore. Restrictions apply.
MERKOFER et al.: DA-MUSIC: DATA-DRIVEN DOA ESTIMATION VIA DEEP AUGMENTED MUSIC ALGORITHM 2775

Fig. 3. Block diagram of the DA-MUSIC algorithm.

given by covariance matrix and can, therefore, become indistinguishable

H from noise [54]. Furthermore, the MUSIC algorithm is inher-
KX = A(θ)KS A (θ) + λK0X , (9) ently a narrowband approach due to the assumptions imposed on
with KS being the covariance of the incident signals s(t). The the system model. Nonetheless, it can be extended to broadband,
matrix A(θ)KS AH (θ) is singular and has a rank of less than M i.e., signal models as in (7), and so we next describe the coherent
when the number of array elements M is strictly larger than the method to achieve this, which is also adopted in our proposed
number of signals D. Therefore, λ is an eigenvalue of KX (in DA-MUSIC.
the metric of K0X , which takes the form λK0X = σ 2 I for addi-
tive white Gaussian noise (AWGN) with variance σ 2 ). Further, E. Coherent Broadband
A(θ)KS AH (θ) has to be non-negative definite, because A(θ) The main concept behind coherent broadband DoA estimation
has full rank and KS is positive definite, and consequently, λ is to transform the different frequency covariance matrices into
in (9) is the minimal eigenvalue of KX , denoted λmin . The a single covariance matrix at a focusing frequency. Accord-
multiplicity of λmin corresponds to the number of incident ingly, coherent methods find appropriate transformations for
wavefronts, and equals N = M − D. each frequency, transform the covariances, and obtain a focused
MUSIC builds upon this representation of the covariance of covariance matrix by some form of averaging. In particular, the
the signals. The algorithm takes as input T snapshots of the CSS method [11] aims at combining the spatial signal subspaces
waveforms at M array elements, represented as X in (4), and to align the signal subspaces associated with the DoA along all
uses them to obtain an empirical estimate of KX via K̂X = frequency bins. To formulate this, let K(ω) be the covariance of
H
T XX . Then, the number of incident signals D is estimated
1
the frequency domain observations (7), and divide the spectrum
via into B IFBs with central frequencies {ωb }B b=1 . The focused
D̂ = M − N̂ , (10) covariance matrix, which is used as the input covariance for
narrowband DoA recovery, is estimated as
where N̂ is the estimated multiplicity of the minimal eigen-
value of K̂X . The eigenvectors corresponding to the N̂ smallest 
B
K= αb Tb K(ωb ) TH
b , (12)
eigenvalues form the noise subspace EN̂ , which is orthogonal
b=1
to the D dimensional signal subspace spanned by the incident
signal mode vectors. Consequently, MUSIC estimates the DoAs where Tb is the focusing matrix and frequency bins are pri-
by computing the spatial spectrum oritized by the weighting αb . The focusing matrices can be
determined by attempting to focus the spectral components at
1
P (ψ) = , (11) some frequency ωr and some focusing angles ψ by solving
aH (ψ)EN̂ EH
N̂
a(ψ)
Tb = arg min A(ωr , ψ) − T A(ωb , ψ)F . (13)
and the D̂ dominant peaks of P (ψ) are set as the estimated DoA T

angles θ̂. Coherent techniques have been shown to achieve a better

MUSIC is a popular and highly applicable subspace-based estimation accuracy and a smaller computational complexity
method that is reasonably efficient and statistically consis- as well as a lower resolution threshold than non-coherent meth-
tent [5], [53]. When the signal model is adequately accurate, ods [52]. However, the CSS methods require initial values for the
it can achieve super-resolution and deliver a highly accurate focusing matrix T, the reference frequency ωr , and the relevant
estimate of the number of signal sources present. Nevertheless, focusing angles ψ to find the focusing matrices with (13), and
the algorithm is sensitive towards the accuracy of the empirical are typically sensitive towards these initial values. Additionally,
estimate of KX , and cannot reliably estimate the DoA angles as it is not guaranteed that the alignment of the signal and noise
well as the number of sources of coherent signals. The reason for subspaces exists to form a viable general covariance matrix
this is that highly correlated signals cause zero entries within the without disarranging the noise subspace [51].

Authorized licensed use limited to: Eindhoven University of Technology. Downloaded on February 14,2024 at 14:56:09 UTC from IEEE Xplore. Restrictions apply.
2776 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 73, NO. 2, FEBRUARY 2024

Fig. 4. Detailed network structure of the DA-MUSIC algorithm.

III. DEEP AUGMENTED MUSIC

DA-MUSIC is a hybrid MB/DD DoA estimation algorithm Fig. 5. Detailed network structure of the DA-MUSIC algorithm with a sepa-
derived from the classic MUSIC algorithm by replacing crucial rately trained (internal) classifier.
and model mismatch sensitive elements of the model-based
structure with specific NNs. Fig. 3 outlines the resulting structure
of the DA-MUSIC architecture and highlights the remaining
similarities to the original MUSIC algorithm, depicted in Fig. 2.
Principally, DA-MUSIC builds upon the understanding that
the core challenges associated with the classic MUSIC algorithm
can be tackled by providing a surrogate covariance matrix. In
particular, as the CSS method provides a surrogate covariance K
that transforms a broadband signal model into a narrowband one,
a similar approach can be employed to handle coherent sources Fig. 6. Detailed outline of the DA-MUSIC subspace selection augmentation.
and array mismatches. Therefore, to improve the categorization
of the noise and signal subspaces, the correlation of the received
measurements is learned from temporal data by employing a of the spatial spectrum in (11), denoted P (ψ), identically to MB
dedicated RNN, which is augmented into the overall flow of MUSIC, by using the noise eigenvectors selected from K̃.
MUSIC. We next elaborate on the architecture in Section III-A, Next, DA-MUSIC attains the DoAs from the spatial spectrum
after which we present the training method in Section III-B and P (ψ) using an additional NN, comprised of a multi-layer per-
provide a discussion in Section III-C. ceptron (MLP) of three fully connected dense layers followed by
a single dense layer with linear activation. The input to the MLP
are R samples of P (ψ) taken uniformly in [0, 2π). The output of
A. Architecture the MLP is the set of estimated DoA angles θ̂. Through the linear
The proposed DA-MUSIC algorithm preserves the structure output activation the network obtains a larger range of freedom
of the MB MUSIC while replacing certain critical aspects with while the periodicity of the angles can be used to map the output
NNs. Our neural augmentations aim to improve the crucial steps to the desired angle range of interest. The benefits of using
of estimating the noise and signal subspaces from the empirical a NN-based peak-finder compared to a model-based one are
covariance and the translation of the spatial spectrum into DoAs two-fold. First, learning the translation of the pseudo-spectrum
via peak finding. By doing so, DA-MUSIC is not constrained into DoAs from data enables achieving improved resolution
by the additional model assumptions imposed in the derivation compared to conventional peak finding, since θ̂d ∈ [0, 2π) in-
of MUSIC, and can, as we will show, e.g. learn to successfully stead of being dependent on the number of angles ψ used to
localize coherent signals. To present the architecture of DA- evaluate the spectrum. Furthermore, peak finding is generally
MUSIC, we commence with the simple case where the number non-differentiable; thus, replacing it with a NN facilitates train-
of sources D is known, and then show how its estimation is ing DA-MUSIC end-to-end. The resulting architecture enables
incorporated. Details of the NNs used in our experimental study the application of gradient-based optimization, by propagating
are reported in Section IV. through the NNs as well as the EVD operation, as done in [55].
1) Known Number of Sources: Fig. 4 depicts a detailed out- Doing so allows us to jointly tune the noise subspace recovery
line of the individual elements of the DA-MUSIC architec- along with the translation of the MUSIC spectrum into DoAs by
ture. The respective output dimensions of the corresponding comparing its estimated DoAs with the true DoAs, as we detail in
components are given in the bracket notation. First, the input Section III-B.
signal x(t) is transformed into the pseudo covariance matrix 2) Varying Number of Sources: Delivering unbiased esti-
K̃ using a RNN implemented through a gated recurrent unit mates of the number of signal sources as well as the ability
(GRU). The final state of the GRU is passed to a dense layer to successfully localize these sources makes MUSIC highly
enabling a reshaping to the desired dimension of the pseudo applicable. The above-discussed DA-MUSIC architecture can
covariance matrix K̃ as well as the subsequent transformation of be extended to operate with a potentially unknown and varying
the complex space. Then, through the continued use of the EVD, number of signal sources, despite the determinant nature of the
the algorithm categorizes the subspaces using the eigenvectors. NNs. This is achieved by addressing three key aspects of the
Inserting the steering vectors a(ψ) allows to compute an estimate algorithm: the estimation of the number of sources; the selection

Authorized licensed use limited to: Eindhoven University of Technology. Downloaded on February 14,2024 at 14:56:09 UTC from IEEE Xplore. Restrictions apply.
MERKOFER et al.: DA-MUSIC: DATA-DRIVEN DOA ESTIMATION VIA DEEP AUGMENTED MUSIC ALGORITHM 2777

process of the noise subspace from the eigenvectors; and the B. Training Procedure
adaptation of the output strategy to overcome the varying number
DA-MUSIC is trained end-to-end in a supervised setting as a
of DoA angles. multiple regression problem. As detailed in Section II-B, the
Estimating number of sources: Our design allows DA-MUSIC
training set is comprised of U tuples of sequences of mea-
to learn abstract mappings as pseudo covariance features K̃,
surements and their corresponding DoAs; i.e., the uth tuples
which is geared towards end-to-end training and is not restricted includes the Tu measurements Xu and their corresponding Du
to a natural ordering of the model-based covariances. Conse-
DoA angles θ u . We first describe how this data is used for
quently, instead of estimating the number of sources by inspect-
training when D is known and fixed; i.e., Du ≡ D. This acts
ing the magnitude of its eigenvalues, we opt for a data-driven as a preliminary step for discussing how the training procedure
approach. We augment the process of estimating the number
is carried out in the general case where the model does not have
of sources with a classifier, implemented through a MLP, as knowledge of D.
a classification task. Since subspace methods with M inputs 1) Known Number of Sources: Given a sequence of measure-
can resolve at most M − 1 signals, the classifier has M − 1
ments X as input, the model predicts the estimated DoA angles θ̂
classes. Fig. 5 depicts the DA-MUSIC architecture with an added that are compared to the true DoA angles θ. Gradient-based opti-
classifier taking the eigenvalues as an input and outputting a
mization is possible because every element of the architecture is
probability vector for the estimated number of sources with a
differentiable, allowing backpropagation through the complete
softmax activation. structure. Derived from the root mean squared periodic error
Computing noise subspace: The noise subspace selector
(RMSPE) [56], [57] the following loss function additionally
needs to know the number of sources D ∈ {1, . . ., M − 1}
compares all permutations of the predicted angles with the true
to classify the eigenvectors into signal and noise subspaces; angles to capture all possible assignments of the estimated DoA
i.e., by choosing the eigenvectors corresponding to the N =
to the true DoA. Thereby the minimal permutation RMSPE in-
M − D smallest eigenvalues. When the number of sources
cludes the permutation invariance of the DoAs and is obtained as:
is not known, we propose to weight each eigenvector by an
estimate of it belonging to the noise subspace. We do this   2  21
1  
by introducing an additional neural augmentation, depicted in RMSPE(θ, θ̂) = min modβ (θ − Pθ̂) , (14)
P∈PD D
Fig. 6, which uses a MLP to cluster the eigenvalues in a learned
fashion. where PD is the set of all D × D permutations and modβ
In particular, the MLP maps the estimated M eigenvalues denotes the element-wise modulus operation regarding the
into a vector q = [q1 , . . ., qM ] whose entries hold the individ- angle range of interest, e.g. β = π for ψ ∈ [−π/2, π/2) or
ual probabilities of choosing the corresponding eigenvector as β = 2π for ψ ∈ [0, 2π).
a noise eigenvector, implemented using a sigmoid activation. 2) Unknown Number of Sources: As discussed in the previ-
The selection is performed by computing Ẽ = Vdiag(q) = ous subsection, DA-MUSIC is designed to resolve a varying and
[q1 v1 . . . qM vM ], allowing to learn a suitable noise subspace. unknown number of sources by introducing an additional NN
Note that this setting specializes in the conventional approach classifier for the number of sources. To formulate the training
of assigning based on the multiplicity of the minimal eigen- procedure of the overall system, we use wc to denote the
values; in the conventional approach, the entries of q are trainable parameters of the classifier, while wd represents the
either ones or zeros and Ẽ coincides with the correspond- parameters of the remaining trainable modules of DA-MUSIC
ing subspace. Consequently, the proposed approach provides (covariance estimator, peak finder, and subspace selector). The
additional flexibility in selecting the noise subspace and fa- loss used to train wd is the RMSPE loss of (14), which is altered
cilitates coping with settings where distinguishing between during training to account for varying sources by computing,
the eigenvectors is challenging due, for example, to low  
SNRs. LRMSPE θ u , θ̂(Xu ) = RMSPE θ1:Du , θ̂1:Du (Xu ) .
Outputting varying number of DoAs: The most common (15)
strategy to overcome the dynamic output dimensions of NNs is In (15), θ̂(Xu ) denotes the M − 1 outputs of DA-MUSIC ap-
to scale the output dimension to the maximum occurring value. plied to Xu , while θ1:Du = (θ1 , . . ., θDu ). This means that only
In our case, since MUSIC cannot resolve more than M − 1 the first Du angles of the estimated DoA θ̂ are compared with
sources, the dimension of the last dense layer of DA-MUSIC the true DoA θ while the remaining angles of θ̂ are completely
is set to M − 1. The only additional modification required with ignored.
this strategy compared with known D is the slight alteration to The separate classifier is trained using the categorical cross-
the loss function discussed in Section III-B below. An advantage entropy of the classes as a loss function ensuring optimal train-
of this strategy is that the approach allows to extract up to M − 1 ing. In particular, letting λu be the input to the MLP classifier
DoA, where θ1 is most likely a true DoA angle, θ2 has a slightly when applying DA-MUSIC to Xu and letting D̂(λu ) be the
lower probability to be a DoA angle, etc. The final estimation is softmax output of the MLP applied to λu (with D̂i (λu ) being
thus carried out by first taking D̂ from the module that estimates its ith entry), the loss used for training wc is given by
the number of sources, and then using the first D̂ outputs as the 
recovered DoAs. LCE Du , D̂(λu ) = − log D̂Du (λu ). (16)

Authorized licensed use limited to: Eindhoven University of Technology. Downloaded on February 14,2024 at 14:56:09 UTC from IEEE Xplore. Restrictions apply.
2778 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 73, NO. 2, FEBRUARY 2024

TABLE I
Algorithm 1: Training DA-MUSIC. DEFAULT SIMULATION PARAMETERS

Specifically, the RNN utilized by DA-MUSIC is able to learn

an appropriate focusing matrix while concurrently correlating
the measurements comparably to (12). The usage of a RNN
applied to the time-domain signal and only passing the last state
to the next layer allows DA-MUSIC to operate with different
signal durations and possibly cope in real-time with dynamic
variations in the DoAs, though investigation of the latter is left
for future work. Our experiments indicate that a deeper RNN
It is noted that since (16) is used for training wc , then one aids the correlation aspect while a wider RNN allows a more
should block the gradient during backpropagation from passing complicated mapping with this correlation.
from the classifier to the EVD, as indicated by the dashed Another important component of the architecture of DA-
connections in Fig. 5. This ensures that the MLP learns from the MUSIC is its incorporation of the EVD as a means for division
eigenvalues themselves and not by influencing and disrupting into subspaces. Though computationally expensive, the internal
the GRU. Further, the regressor is completely independent of EVD allows DA-MUSIC to not only classify signal and noise
the classifier, which allows DA-MUSIC to operate with different subspaces with the eigenvalues but also significantly simplifies
classifiers if needed or with any other desired scheme delivering the estimation of the number of signal sources present. To enable
an estimate of D. The resulting training procedure (employing training end-to-end from the errors in (14), a NN-based peak
mini-batch gradient descent with ADAM [58]) is summarized finder is used in the form of a MLP. Model-based peak-finding
as Algorithm 1. is generally non-differentiable; therefore, replacing it with a
NN enables gradient-based optimization through the entire DA-
C. Discussion MUSIC structure. Furthermore, improved resolution can be
achieved by extracting the DoA from the spatial spectrum in
The design of the architecture of DA-MUSIC is derived from a learned manner (i.e., θ̂d ∈ [0, 2π) and it is not dependent on
the model-based MUSIC structure. This allows for exploiting the the number of angles ψ used to evaluate the spectrum P (ψ)).
successful aspects of the algorithm while improving certain criti- Moreover, while DA-MUSIC is formulated for ULAs, it can be
cal elements and alleviating important drawbacks. Replacing the applied with alternative configurations for which one can com-
empirical covariance estimation with a RNN is the key neural pute the steering matrix, as we also exemplify in our numerical
augmentation of DA-MUSIC, enabling the system to learn the study in Section IV-C.
pseudo covariance from the measurements themselves such
that the resulting surrogate model facilitates subspace-based
DoA recovery. Thereby, the performance of DA-MUSIC, for IV. NUMERICAL EVALUATIONS
example, is not affected by coherent signals and other related In this section, we present our numerical evaluations of the
issues discussed in Section II-D. Furthermore, learning end-to- proposed DA-MUSIC algorithm1 . Our experimental study is
end allows DA-MUSIC to operate with broadband signals, as comprised of evaluations in a narrowband synthetic setting
it effectively learns to produce a focused pseudo covariance, (Section IV-A); a broad synthetic setup (Section IV-B); and
similarly to the CSS methods discussed in Section II-C. It is experiments with real-world data corresponding to azimuth
noted that our augmentation approach depends on the array estimation in seismic arrays (Section IV-C).
geometry, as the steering vectors a(·) are used for computing the
MUSIC spectrum. Nonetheless, as we numerically demonstrate
in Section IV, DA-MUSIC learns to overcome mismatches in the
array geometry from the data without any alterations or renewed 1 The source code used in our experiments can be found at https://github.com/
training. DA-MUSIC/TVT23.

Authorized licensed use limited to: Eindhoven University of Technology. Downloaded on February 14,2024 at 14:56:09 UTC from IEEE Xplore. Restrictions apply.
MERKOFER et al.: DA-MUSIC: DATA-DRIVEN DOA ESTIMATION VIA DEEP AUGMENTED MUSIC ALGORITHM 2779

TABLE II
RMSPE OF DIFFERENT DOA ESTIMATION ALGORITHMS WITH CONSTANT AND
KNOWN D FOR T = 200 SNAPSHOTS

Fig. 7. RMSPE of DoA estimation of D = 2 closely spaced sources to analyze

the resolution of the methods.

A. Synthetic Narrowband
The numerical evaluations of synthetic data presented be-
low are obtained by simulating the measurements x(t) ac-
cording to the narrowband system model (1). In particular,
we simulate a ULA with M = 8 array elements that measure
impinging waveforms originating from the DoA angles θ =
[θ1 , . . . , θD ], which are separately drawn from the uniform dis-
tribution U (−π/2, π/2). The signals s(t) = [s1 (t) . . . sD (t)]
are each drawn randomly from the complex Gaussian distri-
bution CN (0, 1) for all t modeling random amplitudes and
phases. The noises measured at the M array elements v(t) =
[v1 (t) . . . vM (t)] are also drawn from CN (0, 1) for all t, fol- Fig. 8. RMSPE of DoA estimation with various levels of mismatch in the
array geometry for D = 5 sources.
lowed by appropriate scaling to meet the constant SNR. In the
coherent cases, all signals have identical amplitudes and phases,
and when not stated otherwise, the simulation parameters are set
according to Table I. located close together at a Δθ distance from each other. The
1) Known Number of Sources: We first evaluate the RMSPE MB MUSIC algorithm is shown to collapse when the angular
in [rad] achieved when the number of sources D is known in difference approaches Δθ ≈ 0.1 radians, while DA-MUSIC
the synthetic narrowband scenario described above. The results, demonstrates a constant low error for all Δθ, indicating its
reported in Table II, compare the performance of the following improved resolution.
DoA estimators for a different number of sources D: Next, we evaluate DA-MUSIC in the presence of a mismatch
r The DA-MUSIC architecture is implemented according to in the array geometry. Fig. 8 depicts the RMSPE achieved
Fig. 3 and trained separately for each case D. when each element of the steering vector a(·) is corrupted
r The classic MB MUSIC algorithm, implemented as de- with zero-mean Gaussian noise, leading to a mismatch from
scribed in Section II-D, utilizes the external knowledge of the values used to compute the spatial spectrum. Indicating
D. improved robustness, DA-MUSIC is shown to overcome such
r The CNN architecture of [18], with a reduced stride in the mismatches in the array geometry from the data. The CNN,
first layer to account for M = 8, and an increased grid-size DeepMUSIC, and the Random algorithm remain unaffected as
for the last layer to achieve R = 360. they are independent of a(·).
r The DD DeepMUSIC proposed in [26], while incorporat- Fig. 9 depicts the performances differences when localizing
ing minor alterations that were necessary to accommodate D = 5 non-coherent signals with T = 200 snapshots available
for the difference in the setup, includes tuning of individual for different SNR levels in the range of [−20, 20] dB. DA-
hyperparameters to assure successful training of the CNNs. MUSIC shows a constant low SNR for positive dB settings and
The above algorithms are compared with random guessing of without any fluctuations which slowly decreases with increasing
the DoA angles. The results in Table II show that the proposed SNR.
DA-MUSIC notably outperforms all considered benchmarks, Fig. 10 depicts the performance degradation of the estimators
notably surpassing the MB MUSIC algorithm not only for with fewer snapshots available. The DD estimators are only
coherent sources but also for non-coherent ones, which is the trained for the case T = 200 as indicated by the circle around the
scenario for which the MB algorithm is designed. T = 200 marker, yet manage to operate with shorter sequences
To compare the resolution of the algorithms, Fig. 7 shows the during inference due to the recurrent unit or by taking the
RMSPE for localizing D = 2 non-coherent signals, which are covariance matrix as input.

Authorized licensed use limited to: Eindhoven University of Technology. Downloaded on February 14,2024 at 14:56:09 UTC from IEEE Xplore. Restrictions apply.
2780 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 73, NO. 2, FEBRUARY 2024

TABLE III
RMSPE OF DIFFERENT DOA ESTIMATION ALGORITHMS WITH VARYING AND
UNKNOWN D FOR T = 200 SNAPSHOTS

Fig. 9. RMSPE of DoA estimation of D = 5 signals with different SNRs.

2) Unknown and Varying Number of Sources: Table III

shows the results obtained in the exact same narrowband sce-
nario introduced in Section IV-A above but with an unknown and
varying number of sources D ∈ {2, . . ., 5}. Here, during infer-
ence, the DoA estimation algorithms do not have any knowledge
of the varying number of sources present. To compute the
RMSPE in such settings (i.e., if the DoA estimators output the
wrong number of sources D̂), the DoAs are either truncated
(least dominant peaks for the MB algorithms and the CNN
Fig. 10. RMSPE of DoA estimation with D = 5 signals with different number and highest indexed DoA angles for DA-MUSIC) or padded
of snapshots T . with random DoA angels until |θ̂| = |θ| = D. We compare the
following DoA estimators:
r The DA-MUSIC architecture is implemented according to
Fig. 5 with the classifier which predicts the number of
sources being trained along with the overall DoA estima-
tion method.
r The DA-MUSIC (RTC) variation is also implemented ac-
cording to Fig. 5, yet with a retroactively trained classifier
(RTC), i.e., we first train DA-MUSIC for a specific sce-
nario, and then we fix the DoA estimator modules and train
only the classification network for alternate scenarios.
r The Classic MUSIC algorithm is implemented as before
but determines the number of sources by estimating the
multiplicity of the smallest eigenvalue utilizing a pre-
determined threshold.
Fig. 11. RMSE obtained for multi-source DoA estimation (in the angle range r The CNN of [18] as introduced above.
[−π/3, π/3) and a source spacing of at least one degree).
r A conventional beamformer [8], utilizing a peak-finder to
estimate the number of sources by determining the number
of dominant peaks.
We conclude the evaluation of DA-MUSIC with a known r The Random algorithm corresponds to the base perfor-
number of sources by comparing the performances of the estima- mance when choosing DoA angles at random.
tors with the common RMSE measure in Fig. 11. Additionally, Figs. 12 and 13 show the accuracy in identifying the number of
we provide a comparison with the Ziv-Zakai bound for multi- sources versus the number of snapshots T obtained by the men-
source DoA estimation [59], [60]. The plot shows performances tioned algorithms during the estimation of D̂ for non-coherent
for different number of sources in a scenario with T = 200 and coherent signals, respectively. Again, the DD estimators are
snapshots, SNR = 10 dB, DoA angle range in [−π/3, π/3), only trained for the case T = 200 as indicated by the circle.
and a source spacing of at least one degree. These results demon- Unfortunately, the performance of the internal classifier of DA-
strate that among all considered DoA estimators, DA-MUSIC MUSIC is dependent on the number of snapshots, and to be
is consistently closest to the Ziv-Zakai bound (denoted ZZB in able to maintain a more constant accuracy it must be trained
Fig. 11), though all estimators are within some gap of this lower for each case. Consequently, DA-MUSIC (RTC), which trains
bound. its classifier retroactively, requires separate training for each

Authorized licensed use limited to: Eindhoven University of Technology. Downloaded on February 14,2024 at 14:56:09 UTC from IEEE Xplore. Restrictions apply.
MERKOFER et al.: DA-MUSIC: DATA-DRIVEN DOA ESTIMATION VIA DEEP AUGMENTED MUSIC ALGORITHM 2781

Fig. 12. Accuracy of estimating D̂ for various T with D ∈ {2, . . ., 5} non- Fig. 14. RMSPE for various T with varying and unknown number of
coherent narrowband signals. D ∈ {2, . . ., 5} non-coherent narrowband signals.

Fig. 15. RMSPE for various T with varying and unknown number of
D ∈ {2, . . ., 5} coherent narrowband signals.
Fig. 13. Accuracy of estimating D̂ for various T with D ∈ {2, . . ., 5} coherent
narrowband signals.

different broadband scenarios are specified in Table IV if not

specified otherwise. We simulate the following DoA estimators:
number of snapshots, yet manages to achieve the most accurate r The DA-MUSIC architecture is implemented according
predictions. to Fig. 3, but the GRU parameters are scaled (having
The corresponding performances in localizing the varying 10 times more parameters available) to enable optimal
number of sources are depicted in Figs. 14 and 15 for non- learning despite the more complex broadband scenarios.
coherent and coherent signals, respectively. The shown RMSPE r The classic MUSIC algorithm is implemented in its nar-
is an average over all considered D ∈ {2, . . ., 5}. The fundamen- rowband format as described in Section II-D and utilizes
tal limitation of the MB MUSIC structure to estimate the number steering vectors calibrated to the exact array element spac-
of signal sources for coherent signals also severely impacts the ing using min /2.
localization abilities of the algorithm. r Broadband MUSIC corresponds to an incoherent broad-
band extension of MUSIC [46] and is implemented us-
B. Synthetic Broadband ing 10 [Hz] per IFB; i.e., |ωb − ωb−1 | = 10 [Hz] for all
b ∈ {1, . . ., B}.
We proceed to evaluate DA-MUSIC in a broadband set- r The DoA estimators are again compared to choosing DoA
ting. The previously introduced synthetic environment requires angles at random.
certain alterations and assumptions to account for broadband sig- We consider the following three different signal models of the
nals. The sensor elements of the receiving array are adequately broadband signals sd (t) for d ∈ {1, . . ., D}:
spaced, and the element spacing is therefore assumed to be 1) Broadband Scenario 1: A broadband signal is obtained as
min 1 c narrowband signals modulated on different carrier frequencies,
Δm = = , (17) i.e.,
2 2 fmax
sd (t) = s̄d exp (2πjfc,d t), (18)
where min is the minimal wavelength corresponding to the
maximal occurring frequency fmax and the frequency spectrum where for each d ∈ {1, . . ., D}, both s̄d and fc,d are randomly
of interest is considered to be within [fmin , fmax ]. The measure- drawn from CN (0, 1) and U (fmin , fmax ) respectively.
ments are simulated utilizing the broadband system model (7), 2) Broadband Scenario 2: Broadband signals are obtained
where the elements of S(ω) and W(ω) are the Nf -point FFTs via orthogonal frequency division multiplexing (OFDM) [61],
of the elements of s(t) and w(t). The parameter values of the which are modulated on the same carrier frequency. The signals

Authorized licensed use limited to: Eindhoven University of Technology. Downloaded on February 14,2024 at 14:56:09 UTC from IEEE Xplore. Restrictions apply.
2782 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 73, NO. 2, FEBRUARY 2024

TABLE IV
SIMULATION PARAMETERS OF THE BROADBAND SCENARIOS

are considered in baseband and take the following form

1 
K−1
Fig. 16. RMSPE for various T with known number of D = 2 signals from
sd,OFDM = s̄k exp (2πjkΔfd t/K), (19) Broadband Scenario 1.
K
k=0

where s̄k ∼ CN (0, 1) is randomly drawn for each of the K

subcarriers and the bandwidth is Δfd = fmax − fmin .
3) Broadband Scenario 3: A combination of the two pre-
vious scenarios and consists of OFDM signals modulated on
different carrier frequencies
sd (t) = exp (2πjfc,d t) sd,OFDM , (20)
where fc,d is drawn randomly from U (fmin , fmax − Δfd ) to
account for the signal bandwidth Δfd .
Results: Figs. 16, 17, and 18 present the RMSPE obtained
when localizing D = 2 broadband signals from Broadband
Scenario 1, 2, and 3 respectively. The number of snapshots
Fig. 17. RMSPE for various T with known number of D = 2 signals from
goes as high as the sampling frequency fs = 2000 [Hz] and is Broadband Scenario 2.
given logarithmically. This high number is suitable for the MB
broadband MUSIC algorithm to achieve reliable transformation
from the time domain to the frequency domain. DA-MUSIC is
again only trained for the case T = 200, as indicated by a circle
around the marker, yet manages to perform similarly well with
a higher number of snapshots. As expected, the classic narrow-
band MUSIC algorithm completely fails to operate with these
broadband signals, while DA-MUSIC consistently achieves the
most accurate estimates, outperforming the MB broadband
MUSIC algorithm, except for Broadband Scenario 3 with a
very large number of snapshots T > 103 , where DA-MUSIC
trained with much shorter sequences is slightly outperformed by
the MB estimator. These results demonstrate the suitability of
DA-MUSIC for coping with broadband scenarios with a limited
Fig. 18. RMSPE for various T with known number of D = 2 signals from
number of observations. Broadband Scenario 3.
Figs. 19, 20, and 21 analyze the performances with differently
sized frequency ranges [fmin , fmax ]. It is noted that the archi-
tecture of DA-MUSIC is almost invariant towards such scalings
C. Non-Synthetic Data: Azimuth Estimation in Seismic Arrays
and manages to handle signals during inference with much larger
bandwidths or modulated with higher carrier frequencies than In this section, we demonstrate the feasibility and the perfor-
the signals of the training data. Specifically, DA-MUSIC is mance of DA-MUSIC in processing non-synthetic seismic data.
only trained for signals with carrier frequencies and bandwidths The seismic data were recorded by the German Experimental
within 0 to 1000 [Hz] as indicated by the circle around the marker. Seismic System (GERES) array located in the Bavarian Forest,
The results in Figs. 19–21 show that DA-MUSIC, whose com- Germany. GERES is part of the Comprehensive nuclear-Test
plexity is fixed, learns to achieve the most accurate estimates, Ban Treaty Organization (CTBTO) international monitoring
outperforming the MB broadband MUSIC; the latter requires an system, and is a well-maintained and calibrated station. Data
increase of the number of IFBs to overcome a larger frequency from GERES is continuously streamed to the International Data
range and has a constant 10 [Hz] per bin in the depicted results Centre (IDC) of the CTBTO, where it is analyzed. The array is
leading to a severe increase in computational complexity. composed of 25 vertical seismometers with a minimal distance

Authorized licensed use limited to: Eindhoven University of Technology. Downloaded on February 14,2024 at 14:56:09 UTC from IEEE Xplore. Restrictions apply.
MERKOFER et al.: DA-MUSIC: DATA-DRIVEN DOA ESTIMATION VIA DEEP AUGMENTED MUSIC ALGORITHM 2783

TABLE V
RMSPE IN [RAD] OF DIFFERENT DOA ESTIMATION ALGORITHMS
FOR SEISMIC DATA

r1 , . . ., rM . Using these parameters, the steering vectors take

the following form:
 
Fig. 19. RMSPE across different frequency ranges for the carrier frequencies a(f, u, ψ, α) = e−j2πf ur1 k(ψ,α) , . . . , e−j2πf urM k(ψ,α) ,
of D = 2 signals from Broadband Scenario 1.

for some frequency f and with the wave vector for certain
elevation α and azimuth ψ of interest,

k(ψ, α) = [sin α cos ψ, sin α sin ψ, cos α].

We compare DA-MUSIC with following DoA estimators for the

setting settings α = −π/4 and f = 1 [Hz]:
r Broadband MUSIC, corresponding to an incoherent broad-
band extension of MUSIC [46], and instead of using the
constant f = 1 [Hz] it utilizes 10 IFB with frequencies in
[0, 20] [Hz].
r The classic MUSIC algorithm, implemented in its narrow-
band format as described in Section II-D with additionally
Fig. 20. RMSPE across different frequency ranges for the bandwidth of D = 2
signals from Broadband Scenario 2. filtering the measurements with an experimentally cali-
brated low-pass filter, allowing only frequencies within [0,
10] [Hz] to pass.
r A conventional beamformer [8].
r Choosing a DoA angle at random.
The results, reported in Table V, show that DA-MUSIC man-
ages to outperform the MB estimators by not only focusing
the frequency component but also by concurrently focusing
the interdependent elevation angle to receive a stable azimuth
estimation. On the other hand, the MB algorithms require further
knowledge of the elevation and frequency at hand to operate
reliably. While the errors in Table V may appear to be relatively
large, it is noted that the average error achieved via expert analy-
sis reported by the IDC Reviewed Event Bulletin is 0.4243 [rad].
Fig. 21. RMSPE across different frequency ranges for the carrier frequencies This indicates the ability of DA-MUSIC to achieve comparable
of D = 2 signals from Broadband Scenario 3. results and to notably outperform the MB estimators while
operating with simplified and approximated model parameters.

between the sites of 124 [m] and an aperture of approximately

V. CONCLUSION
2.13 [km]. The seismic signal at each sensor is sampled at
40 [Hz]. Details about the GERES arrays and the exact array We presented a hybrid MB/DD implementation of the MUSIC
configuration can be found in [62]. algorithm for DoA estimation. The proposed DA-MUSIC was
We use data from October to December 2021, in which shown to mitigate some of the limitations and drawbacks of the
GERES detected arrivals for 2904 events of which 2816 were classic method. DA-MUSIC is operable with an unknown num-
used during this analysis, 2534 for training, and 282 for testing. ber of sources and with broadband signals while being adaptable
We employ sliding windows of length 100 seconds with a shift to various scenarios and robust towards severe mismatches in the
of 25 seconds around the signal arrival time as designated by array geometry. The proposed hybrid MB/DD approach provides
the IDC analysis. The following parameters for each event are a viable alternative in both low and high snapshot domains and
obtained from the IDC’s Reviewed Event Bulletin: the azimuth shows a remarkable resolution capability compared to both MB
DoA angle θ, the slowness value u, and the sensor positions and DD benchmarks in various settings.

Authorized licensed use limited to: Eindhoven University of Technology. Downloaded on February 14,2024 at 14:56:09 UTC from IEEE Xplore. Restrictions apply.
2784 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 73, NO. 2, FEBRUARY 2024

ACKNOWLEDGMENT [24] N. Shlezinger, Y. C. Eldar, and S. P. Boyd, “Model-based deep learning: On

the intersection of deep learning and optimization,” IEEE Access, vol. 10,
We thank Dr. Yochai Ben Horin for constructive and valuable pp. 115384–115398, 2022.
joint discussions on the seismic data, and for providing the data. [25] N. Shlezinger, J. Whang, Y. C. Eldar, and A. G. Dimakis, “Model-based
deep learning,” Proc. IEEE, vol. 111, no. 5, pp. 465–499, May 2023,
doi: 10.1109/JPROC.2023.3247480.
REFERENCES [26] A. M. Elbir, “DeepMUSIC: Multiple signal classification via deep learn-
ing,” IEEE Sensors Lett., vol. 4, no. 4, pp. 1–4, Apr. 2020.
[1] J. P. Merkofer, G. Revach, N. Shlezinger, and R. J. G. van Sloun, “Deep [27] C. Liu, W. Feng, H. Li, and H. Zhu, “Single snapshot DOA estimation
augmented music algorithm for data-driven DoA estimation,” in Proc. based on spatial smoothing MUSIC and CNN,” in Proc. IEEE Int. Conf.
IEEE Int. Conf. Acoust., Speech, Signal Process., 2022, pp. 3598–3602. Signal Process., Commun. Comput, 2021, pp. 1–5.
[2] I. Bilik, O. Longman, S. Villeval, and J. Tabrikian, “The rise of radar for [28] A. Barthelme and W. Utschick, “DoA estimation using neural network-
autonomous vehicles: Signal processing solutions and future research di- based covariance matrix reconstruction,” IEEE Signal Process. Lett.,
rections,” IEEE Signal Process. Mag., vol. 36, no. 5, pp. 20–31, Sep. 2019. vol. 28, pp. 783–787, 2021.
[3] D. Rahamim, J. Tabrikian, and R. Shavit, “Source localization using vector [29] D. T. Hoang and K. Lee, “Deep learning-aided coherent direction-of-
sensor array in a multipath environment,” IEEE Trans. Signal Process., arrival estimation with the FTMR algorithm,” IEEE Trans. Signal Process.,
vol. 52, no. 11, pp. 3096–3103, Nov. 2004. vol. 70, pp. 1118–1130, 2022.
[4] J. Foutz, A. Spanias, and M. K. Banavar, “Narrowband direction of arrival [30] W. Zhang, Y. Han, M. Jin, and X. Qiao, “Multiple-toeplitz matrices
estimation for antenna arrays,” Synth. Lectures Antennas, vol. 3, no. 1, reconstruction algorithm for DoA estimation of coherent signals,” IEEE
pp. 1–76, 2008. Access, vol. 7, pp. 49504–49512, 2019.
[5] H. Krim and M. Viberg, “Two decades of array signal processing research: [31] S. Chakrabarty and E. A. Habets, “Broadband DoA estimation us-
The parametric approach,” IEEE Signal Process. Mag., vol. 13, no. 4, ing convolutional neural networks trained with noise signals,” in
pp. 67–94, Apr. 1996. Proc. IEEE Workshop Appl. Signal Process. Audio Acoust., 2017,
[6] Z. Ahmad and I. Ali, “Three decades of development in DoA esti- pp. 136–140.
mation technology,” Indonesian J. Elect. Eng. Comput. Sci., vol. 12, [32] W. Zhu and M. Zhang, “A deep learning architecture for broadband
pp. 6297–6312, 2014. DoA estimation,” in Proc. IEEE 19th Int. Conf. Commun. Technol., 2019,
[7] R. Schmidt, “Multiple emitter location and signal parameter estimation,” pp. 244–247.
IEEE Trans. Antennas Propag., vol. 34, no. 3, pp. 276–280, Mar. 1986. [33] L. li Wu and Z. T. Huang, “Coherent SVR learning for wideband
[8] M. S. Bartlett, “Smoothing periodograms from time-series with continuous direction-of-arrival estimation,” IEEE Signal Process. Lett., vol. 26, no. 4,
spectra,” Nature, vol. 161, no. 4096, pp. 686–687, May 1948. pp. 642–646, Apr. 2019.
[9] J. Capon, “High-resolution frequency-wavenumber spectrum analysis,” [34] B. Friedlander, “Wireless direction-finding fundamentals,” in Classi-
Proc. IEEE, vol. 57, no. 8, pp. 1408–1418, Aug. 1969. cal and Modern Direction-of-Arrival Estimation, T. E. Tuncer and B.
[10] T. E. Tuncer, T. K. Yasar, and B. Friedlander, “Narrowband and wideband Friedlander, Eds. Boston, MA, USA: Academic Press, 2009, ch. 1,
DoA estimation for uniform and nonuniform linear arrays,” in Classical pp. 1–51.
and Modern Direction-of-Arrival Estimation. Boston, MA, USA: Aca- [35] Z. Chen, G. Gokeda, and Y. Yu, Introduction to Direction-of-Arrival
demic Press, 2009, ch.4, pp. 125–160. Estimation. Norwood, MA, USA: Artech House, 2010.
[11] H. Wang and M. Kaveh, “Coherent signal-subspace processing for the de- [36] S. Ge, K. Li, and S. N. M. Rum, “Deep learning approach in DoA
tection and estimation of angles of arrival of multiple wide-band sources,” estimation: A systematic literature review,” Mob. Inf. Syst., vol. 2021,
IEEE Trans. Acoust., Speech, Signal Process., vol. 33, no. 4, pp. 823–831, pp. 6392875:1–6392875: 14, 2021 .
Aug. 1985. [37] B. D. Van Veen and K. M. Buckley, “Beamforming: A versatile approach
[12] G. Su and M. Morf, “The signal subspace approach for multiple wide-band to spatial filtering,” IEEE ASSP Mag., vol. 5, no. 2, pp. 4–24, Apr. 1988.
emitter location,” IEEE Trans. Acoust., Speech, Signal Process., vol. 31, [38] B. Friedlander, “The root-music algorithm for direction finding with
no. 6, pp. 1502–1522, Dec. 1983. interpolated arrays,” Signal Process., vol. 30, no. 1, pp. 15–29, 1993.
[13] P.-A. Grumiaux, S. Kitić, L. Girin, and A. Guérin, “A survey of sound [39] T.-Jun Shan, M. Wax, and T. Kailath, “On spatial smoothing for direction-
source localization with deep learning methods,” J. Acoustical Soc. Amer., of-arrival estimation of coherent signals,” IEEE Trans. Acoust., Speech,
vol. 152, no. 1, pp. 107–151, 2022. Signal Process., vol. 33, no. 4, pp. 806–811, Aug. 1985.
[14] S. Chakrabarty and E. A. P. Habets, “Multi-speaker DoA estimation using [40] Q. Chen and R. Liu, “On the explanation of spatial smoothing in music
deep convolutional networks trained with noise signals,” IEEE J. Sel. algorithm for coherent sources,” in Proc. IEEE Int. Conf. Inf. Sci. Technol.,
Topics Signal Process., vol. 13, no. 1, pp. 8–21, Mar. 2019. 2011, pp. 699–702.
[15] Z.-M. Liu, C. Zhang, and P. S. Yu, “Direction-of-arrival estimation based [41] A. Paulraj, R. Roy, and T. Kailath, “Estimation of signal parameters via
on deep neural networks with robustness to array imperfections,” IEEE rotational invariance techniques- ESPRIT,” in Proc. 19th Asilomar Conf.
Trans. Antennas Propag., vol. 66, no. 12, pp. 7315–7327, Dec. 2018. Circuits, Syst. Comput., 1985, pp. 83–89.
[16] L. li Wu, Z. M. Liu, and Z. T. Huang, “Deep convolution network for [42] V. Vasylyshyn, “Direction of arrival estimation using ESPRIT with sparse
direction of arrival estimation with sparse prior,” IEEE Signal Process. arrays,” in Proc. IEEE Eur. Radar Conf., 2009, pp. 246–249.
Lett., vol. 26, no. 11, pp. 1688–1692, Nov. 2019. [43] M. Wax, T.-J. Shan, and T. Kailath, “Spatio-temporal spectral analysis by
[17] H. Hammer, S. E. Chazan, J. Goldberger, and S. Gannot, “Dynami- eigenstructure methods,” IEEE Trans. Acoust., Speech, Signal Process.,
cally localizing multiple speakers based on the time-frequency domain,” vol. 32, no. 4, pp. 817–827, Aug. 1984.
EURASIP J. Audio, Speech, Music Process., vol. 2021, no. 1, pp. 1–10, [44] S. Chandran and M. K. Ibrahim, “DoA estimation of wide-band signals
2021. based on time-frequency analysis,” IEEE J. Ocean. Eng., vol. 24, no. 1,
[18] G. K. Papageorgiou, M. Sellathurai, and Y. C. Eldar, “Deep networks for pp. 116–121, Jan. 1999.
direction-of-arrival estimation in low SNR,” IEEE Trans. Signal Process., [45] S. Argentieri and P. Danes, “Broadband variations of the music high-
vol. 69, pp. 3714–3729, 2021. resolution method for sound source localization in robotics,” in Proc.
[19] A. Barthelme and W. Utschick, “A machine learning approach to DoA IEEE/RSJ Int. Conf. Intell. Robots Syst., 2007, pp. 2009–2014.
estimation and model order selection for antenna arrays with subarray [46] Y.-S. Yoon, L. M. Kaplan, and J. H. McClellan, “DoA estimation of
sampling,” IEEE Trans. Signal Process., vol. 69, pp. 3075–3087, 2021. wideband signals,” in Proc. Adv. Direction-of-Arrival Estimation, Artech,
[20] D. Hu, Y. Zhang, L. He, and J. Wu, “Low-complexity deep-learning-based 2006, pp. 49–68.
DOA estimation for hybrid massive MIMO systems with uniform circular [47] J. Krolik and D. Swingler, “Focused wide-band array processing by spatial
arrays,” IEEE Wireless Commun. Lett., vol. 9, no. 1, pp. 83–86, Jan. 2020. resampling,” IEEE Trans. Acoust., Speech, Signal Process., vol. 38, no. 2,
[21] M. L. L. de Oliveira and M. Bekooij, “Deep-MLE: Fusion between a neural pp. 356–360, Feb. 1990.
network and MLE for a single snapshot DOA estimation,” in Proc. IEEE [48] T.-S. Lee, “Efficient wideband source localization using beamforming
Int. Conf. Acoust., Speech Signal Process., 2022, pp. 3673–3677. invariance technique,” IEEE Trans. Signal Process., vol. 42, no. 6,
[22] M. L. L. de Oliveira and M. Bekooij, “ResNet applied for a single-snapshot pp. 1376–1387, Jun. 1994.
DOA estimation,” in Proc. IEEE Radar Conf., 2022, pp. 1–6. [49] B. Friedlander and A. Weiss, “Direction finding for wideband signals
[23] N. Shlezinger, J. Whang, Y. C. Eldar, and A. G. Dimakis, “Model-based using an interpolated array,” in Proc. 25th Asilomar Conf. Signals, Syst.,
deep learning,” Proc. IEEE, vol. 111, no. 5, pp. 465–499, May 2023. Comput., 1991, pp. 583–587.

Authorized licensed use limited to: Eindhoven University of Technology. Downloaded on February 14,2024 at 14:56:09 UTC from IEEE Xplore. Restrictions apply.
MERKOFER et al.: DA-MUSIC: DATA-DRIVEN DOA ESTIMATION VIA DEEP AUGMENTED MUSIC ALGORITHM 2785

[50] E. D. di Claudio and R. Parisi, “WAVES: Weighted average of signal Nir Shlezinger (Senior Member, IEEE) received the
subspaces for robust wideband direction finding,” IEEE Trans. Signal B.Sc., M.Sc., and Ph.D. degrees in electrical and com-
Process., vol. 49, no. 10, pp. 2179–2191, Oct. 2001. puter engineering from Ben-Gurion University, Beer-
[51] Y.-S. Yoon, L. M. Kaplan, and J. H. McClellan, “TOPS: New DOA sheba, Israel, in 2011, 2013, and 2017, respectively.
estimator for wideband signals,” IEEE Trans. Signal Process., vol. 54, He is currently an Assistant Professor wtih the School
no. 6, pp. 1977–1989, Jun. 2006. of Electrical and Computer Engineering, Ben-Gurion
[52] F. Ma and X. Zhang, “Wideband DOA estimation based on focusing signal University. From 2017 to 2019, he was a Postdoctoral
subspace,” Signal, Image Video Process., vol. 13, PP. 675–682, 2019. Researcher with the Technion, and from 2019 to 2020,
[53] S. M. Kay, Fundamentals of Statistical Signal Processing: Estimation he was a Postdoctoral Researcher with the Weizmann
Theory. Englewood Cliffs, NJ, USA: Prentice-Hall International Editions, Institute of Science, where he was awarded the FGS
1993, pp. 27–37. prize for outstanding research achievements. His re-
[54] H. Krim and J. Proakis, “Smoothed eigenspace-based parameter estima- search interests include communications, information theory, signal processing,
tion,” Automatica, vol. 30, no. 1, pp. 27–38, 1994. and machine learning.
[55] O. Solomon et al., “Deep unfolded robust PCA with application to clutter
suppression in ultrasound,” IEEE Trans. Med. Imag., vol. 39, no. 4,
pp. 1051–1063, Apr. 2020.
[56] T. Routtenberg and J. Tabrikian, “Bayesian parameter estimation using
periodic cost functions,” IEEE Trans. Signal Process., vol. 60, no. 3, Tirza Routtenberg (Senior Member, IEEE) received
pp. 1229–1240, Mar. 2012. the B.Sc. degree from the Technion Israel Institute of
[57] T. Routtenberg and J. Tabrikian, “Non-Bayesian periodic Cramér-Rao Technology, Haifa, Israel, in 2005, and the M.Sc. and
bound,” IEEE Trans. Signal Process., vol. 61, no. 4, pp. 1019–1032, Ph.D. degrees in electrical and computer engineer-
Apr. 2013. ing from Ben-Gurion University, Beersheba, Israel,
[58] D. P. Kingma and J. Ba, “Adam: A method for stochastic optimization,” in 2007 and 2012, respectively. She is currently an
2014, arXiv:1412.6980. Associate Professor with the School of Electrical
[59] Z. Zhang, Z. Shi, and Y. Gu, “Ziv-Zakai bound for DOAs estimation,” and Computer Engineering, Ben-Gurion University,
IEEE Trans. Signal Process., vol. 71, pp. 136–149, 2023. Beersheba, Israel. She was a Postdoctoral Fellow with
[60] T. Routtenberg and J. Tabrikian, “A general class of outage error probabil- the School of Electrical and Computer Engineering,
ity lower bounds in Bayesian parameter estimation,” IEEE Trans. Signal Cornell University, Ithaca, NY, USA, from 2012 to
Process., vol. 60, no. 5, pp. 2152–2166, May 2012. 2014. Since 2014, she has been a faculty Member with the School of Elec-
[61] S. B. Weinstein, “The history of orthogonal frequency-division multiplex- trical and Computer Engineering, and Ben-Gurion University of the Negev,
ing [history of communications],” IEEE Commun. Mag., vol. 47, no. 11, Beer-Sheva. During 2022–2023, she was a William R. Kenan, Jr. Visiting
pp. 26–35, Nov. 2009. Professor for Distinguished Teaching with Princeton University, Princeton, NJ,
[62] H.-P. Harjes, “Design and siting of a new regional array in central Europe,” USA. Her research interests include signal processing in smart grids, statistical
Bull. Seismological Soc. America, vol. 80, pp. 1801–1817, 1990. signal processing, estimation and detection theory, and signal processing on
graphs. She is an Associate Editor for IEEE TRANSACTIONS ON SIGNAL AND
INFORMATION PROCESSING OVER NETWORKSand IEEE SIGNAL PROCESSING
LETTERS, and a Member of the IEEE Signal Processing Theory and Methods
Julian P. Merkofer (Student Member, IEEE) re- Technical Committee. She was the recipient of the Best Student Paper Award
at the International Conference on Acoustics, Speech and Signal Processing
ceived the B.Sc. and M.Sc. degrees in electrical engi-
(ICASSP) 2011, IEEE International Workshop on Computational Advances
neering and information technology from the Federal
in Multi-Sensor Adaptive Processing (CAMSAP) 2013 (co-author), ICASSP
Institute of Technology in Zurich (ETHZ), Zurich,
2017 (co-author), and IEEE Workshop on Statistical Signal Processing (SSP)
Switzerland, in 2019, and 2021, respectively. He is
2018 (co-author). She was awarded the Negev scholarship in 2008, Lev-Zion
currently working toward the Ph.D. degree with the
scholarship in 2010, Marc Rich Foundation Prize in 2011, and Toronto Prize for
Biomedical Diagnostics Laboratory, Eindhoven Uni-
versity of Technology, Eindhoven, The Netherlands. Excellence in Research in 2021.
His research interests include machine learning for
signal processing, model-based deep learning, and
magnetic resonance spectroscopy. Ruud J. G. van Sloun (Member, IEEE) received
the M.Sc. and Ph.D. degree (cum laude) in electrical
engineering from the Eindhoven University of Tech-
Guy Revach (Graduate Student Member, IEEE) re- nology, Eindhoven, The Netherlands, in 2014, and
2018, respectively. He is currently an Associate Pro-
ceived the B.Sc. (cum laude) and M.Sc. degrees from
fessor with the Department of Electrical Engineering,
the Andrew and Erna Viterbi Department of Elec-
Eindhoven University of Technology, Eindhoven, the
trical and Computer Engineering, Technion – Israel
Institute of Technology, Haifa, Israel, in 2008 and Netherlands. From 2019 to 2020, he was a Visiting
Professor with the Department of Mathematics and
2017, respectively. Since 2019, he has been working
Computer Science, Weizmann Institute of Science,
toward the Ph.D. degree with the Institute for Sig-
Rehovot, Israel, and from 2020 to 2023, he was a
nal and Information Processing (ISI), ETH Zürich,
Zürich, Switzerland, supervised by Dr. Hans-Andrea kickstart AI Fellow with Philips Research. His current research interests include
closed-loop image formation, deep learning for signal processing and imaging,
Loeliger. He did his master’s thesis under the su-
active signal acquisition, model-based deep learning, compressed sensing, ul-
pervision of Prof. Nahum Shimkin on planning for
trasound imaging, and probabilistic signal and image reconstruction. He was the
cooperative multiagents. He is currently a Researcher with a proven industry
recipient of an ERC starting grant, NWO VIDI grant, NWO Rubicon grant, and
track record as an innovator and system engineer. His research interests include
Google Faculty Research Award.
the intersection of machine learning with signal processing, and more specifi-
cally combining sound theoretical principles from classical signal processing
with state-of-the-art machine learning algorithms for tracking and detection
problems. Before coming to ETH Zürich, he was with the Israeli wireless
communication industry for more than ten years, first as a real-time embedded
Software Engineer and then a Software Manager. He was the main innovator be-
hind state-of-the-art, software-defined radio (SDR) for wireless communication,
game-changing and groundbreaking in terms of size, weight, and power. As a
System Engineer, he defined major aspects of SDR requirements and architecture
for hardware, software, network, cyber defense, signal processing, data analysis,
and control algorithms.

Authorized licensed use limited to: Eindhoven University of Technology. Downloaded on February 14,2024 at 14:56:09 UTC from IEEE Xplore. Restrictions apply.