Journal of Physics: Conference Series

PAPER • OPEN ACCESS

A direction of arrival estimation method based on deep learning

To cite this article: Yuji Liu et al 2020 J. Phys.: Conf. Ser. 1550 032066

View the article online for updates and enhancements.

This content was downloaded from IP address 178.171.1.217 on 16/06/2020 at 13:38

IWAACE 2020 IOP Publishing
Journal of Physics: Conference Series 1550 (2020) 032066 doi:10.1088/1742-6596/1550/3/032066

A direction of arrival estimation method based on deep

learning

Yuji Liu1, Huixiu Chen2 and Biao Wang1*

1
School of electronic information, Jiangsu University of science and technology,
Zhenjiang, Jiangsu, 212003, China
2
School of electronic information, Jiangsu University of science and technology,
Zhenjiang, Jiangsu, 212003, China
*
Corresponding author’s e-mail: [email protected]

Abstract. Direction of arrival (DOA) estimation is a research subject of many experts and
scholars in information, control and communication, and it is also a key technology in smart
antenna and sonar array system. In this paper, the direction of arrival of the signal is estimated
by establishing a classification network model of deep learning. After training, the network
model can effectively identify the direction of arrival of the unknown signal. Under the same
signal-to-noise ratio (SNR) environment, it is verified by software simulation that the accuracy
is improved compared with the classical music algorithm.

1. Introduction
With the development of sonar array system, direction of arrival (DOA) estimation is an important
parameter in sonar array system. The accuracy of DOA estimation will directly affect the performance
of the whole sonar system. After decades of development, DOA estimation has formed a large number
of algorithms[1]. There are conventional beamforming (CBF) algorithm and adaptive minimum
variance distortion less response (MVDR) algorithm based on beamforming technology, and multiple
signal classification (Music) algorithm based on subspace decomposition[2].
This paper introduces a method of DOA estimation based on deep learning, which mainly identifies
the direction of the unknown signal by establishing a classification neural network model of deep
learning. Deep learning is a new direction and hot spot in the field of machine learning. Deep learning
studies the internal laws, representation levels and corresponding features of data. The information
obtained through learning is very meaningful for data interpretation. At present, deep learning has
achieved very good results in speech and image recognition[3].

2. model of array reception

This paper studies that the effect of the algorithm does not require the shape of the array, so this paper
uses a plane linear array as an example, the number of elements is 𝑀, and the spacing of elements is 𝑑.
This paper mainly studies the direction of arrival estimation in the far field, so let 𝑁 narrow-band signals
enter the ray array. Figure 1 shows the incident model of the signal.

Content from this work may be used under the terms of the Creative Commons Attribution 3.0 licence. Any further distribution
of this work must maintain attribution to the author(s) and the title of the work, journal citation and DOI.
Published under licence by IOP Publishing Ltd 1
IWAACE 2020 IOP Publishing
Journal of Physics: Conference Series 1550 (2020) 032066 doi:10.1088/1742-6596/1550/3/032066

𝑠𝑛 (𝑡)

𝜃𝑖

1 𝑑 2 3 4 𝑚 𝑀

Figure 1. Model of signal incidence.

Since the far-field signal is received, the wave front at the receiving array can be approximately a
plane, as shown in Figure 1. There is a wave path difference 𝐷𝑚 when the signal arrives at different
array elements, which can be obtained from the trigonometric function:
𝐷𝑚 = (𝑚 − 1)𝑑𝑠𝑖𝑛𝜃𝑛 (1)
The time difference can be derived from formula (1):
𝐷𝑚
𝜏𝑚 = (2)
𝑣
In formula (2), 𝑣 is the speed of signal propagation in water. Then the phase difference of the
received signal of the array element is:
𝛽 = 𝑒 −𝑗∗2𝜋𝑓𝜏𝑚 (3)
Substituting formula (1-2) into formula (3),
(𝑚−1)𝑑𝑠𝑖𝑛𝜃𝑛
𝛽 = 𝑒 −𝑗∗2𝜋𝑓 𝑣 (4)
𝑣
Since 𝑓 = 𝜆, formula (4) can be reduced to:
(𝑚−1)𝑑𝑠𝑖𝑛𝜃𝑛
𝛽 = 𝑒 −𝑗∗2𝜋 𝜆 (5)
The actual signal is a snapshot signal, so it is a discrete signal. Then the signal 𝑥𝑚 (𝑘) collected by
the uniform linear array:
𝑁

𝑥𝑚 (𝑘) = ∑ 𝑠𝑛 (𝑘)𝛽 + 𝑛𝑚 (𝑘) (6)

𝑛=1
In formula (6), 𝑛𝑚 (𝑘) is the interference noise signal for the 𝑀𝑡ℎ receiving array element. Formula
(6) is written in matrix form as follows:
𝑿 = 𝑨𝑺 + 𝑵 (7)
In formula (7),
𝑿 = [𝑥1 (𝑘), 𝑥2 (𝑘), ⋯ , 𝑥𝑀 (𝑘)]𝑇 is the received signal matrix of the array
𝑺 = [𝑠1 (𝑘), 𝑠2 (𝑘), ⋯ , 𝑠𝑁 (𝑘)]𝑇 is the source signal matrix
𝑵 = [𝑛1 (𝑘), 𝑛2 (𝑘), ⋯ , 𝑛𝑀 (𝑘)]𝑇 is the additive Gaussian white noise matrix received by the array

1 1 ⋯ 1
𝑑𝑠𝑖𝑛𝜃1 𝑑𝑠𝑖𝑛𝜃2 𝑑𝑠𝑖𝑛𝜃𝑁
−𝑗∗2𝜋 −𝑗∗2𝜋 −𝑗∗2𝜋
𝑒 𝜆 𝑒 𝜆 ⋯ 𝑒 𝜆
𝑨= is the flow pattern matrix of the array
⋮ ⋮ ⋮ ⋮
(𝑀−1)𝑑𝑠𝑖𝑛𝜃1 (𝑀−1)𝑑𝑠𝑖𝑛𝜃2 (𝑀−1)𝑑𝑠𝑖𝑛𝜃𝑁
[𝑒 −𝑗∗2𝜋 𝜆 𝑒 −𝑗∗2𝜋 𝜆 ⋯ 𝑒 −𝑗∗2𝜋 𝜆 ]

Because 𝑨 is a Vandermonde matrix, when 𝜃1 ，𝜃2 ， ⋯ ，𝜃𝑁 are different from each other, 𝑨
matrix is full rank. The autocorrelation matrix 𝑹𝒙𝒙 of the signal 𝑿(𝑘) can be obtained as follows:
𝑹𝒙𝒙 = 𝐸{𝑿𝑿𝑯 } (8)

2
IWAACE 2020 IOP Publishing
Journal of Physics: Conference Series 1550 (2020) 032066 doi:10.1088/1742-6596/1550/3/032066

3. The basic principles of MUSIC algorithm

MUSIC algorithm decomposes the eigenvalues of the autocorrelation matrix of array signals to obtain
the signal subspace corresponding to the signal component and the noise subspace orthogonal to the
signal component, and then estimates the incident direction of the signal using the orthogonality of the
two subspaces.
The autocorrelation matrix 𝑹𝒙𝒙 of formula (8) is transformed as follows:
𝑹𝒙𝒙 = 𝐸[(𝑨𝑺 + 𝑵)(𝑨𝑺 + 𝑵)𝑯 ] (9)
𝑯 ]𝑨𝑯 𝑯 (10)
𝑹𝒙𝒙 = 𝑨𝑬[𝑺𝑺 + 𝑬[𝑵𝑵 ]
In equation (10), 𝑬[𝑺𝑺𝑯 ] is the autocorrelation matrix of the source signal, which is recorded as 𝑹𝑺𝑺 .
𝑬[𝑵𝑵𝑯 ]is the autocorrelation matrix of the noise signal, which is recorded as 𝑹𝑵𝑵 , which can be
rewritten as follows:
𝑹𝑵𝑵 = 𝜎 2 𝑰 (11)
2
In formula (11), 𝜎 is the noise power and 𝑰 is the unit matrix of order 𝑀. Therefore, formula (10)
can be written as follows:
𝑹𝒙𝒙 = 𝑨𝑹𝑺𝑺 𝑨𝑯 + 𝜎 2 𝑰 (12)
The white Gaussian noise is not related to the signal. Since 𝜎 2 > 0 and 𝑹𝒙𝒙 has full rank, the
eigenvalues and eigenvectors of 𝑹𝒙𝒙 can be obtained:
𝝀 = [𝜆1 , 𝜆2 , ⋯ , 𝜆𝑀 ] (13)
𝝂 = [𝜈1 , 𝜈2 , ⋯ , 𝜈𝑀 ] (14)
In formula (13-14), 𝑁 eigenvalues and eigenvectors are related to signals, and the rest 𝑀 − 𝑁
eigenvalues and eigenvectors are related to noise. Therefore, the eigenvalues are arranged in descending
order. The first 𝑁 are the eigenvalues of the signal, and the corresponding 𝑁 eigenvectors of these
eigenvalues are the eigenvectors of the signal, and the rest are the eigenvalues and eigenvectors of the
noise. The eigenvectors of 𝑁 signals form the signal subspace:
𝐸𝑠 = [𝜈1 , 𝜈2 , ⋯ , 𝜈𝑁 ] (15)
The eigenvectors of 𝑀 − 𝑁 noises are used to form the noise subspace:
𝐸𝑛 = [𝜈𝑁+1 , 𝜈𝑁+2 , ⋯ , 𝜈𝑀 ] (16)
Because signal subspace and noise subspace are orthogonal to each other. When 𝜃 = 𝜃𝑖 , 𝑖 ∈ [1, 𝑁]
𝑎 𝐻 (𝜃)𝐸𝑛 = 0 (17)
According to the characteristics of formula (17), a function similar to power spectrum is defined:
1
𝑃(𝜃) = 𝐻 (18)
𝑎 (𝜃)𝐸𝑛 𝐸𝑛𝐻 𝑎(𝜃)
In formula (18), the peak value of 𝑃 (𝜃) corresponds to 𝜃, which is the wave direction of the signal
[4].

4. Deep learning classification model

In order to solve the classification problem, deep neural network needs to firstly use the training data to
learn the weight parameters. When reasoning, using the weight just learned to can classify the output
data [5].
Since the correlation matrix 𝑹𝒙𝒙 in formula (12) is Hermitian matrix, the elements on the main
diagonal are all real numbers instead of containing the orientation information. So in this paper, only
the elements of the upper triangle region are selected to form the reference vector 𝑳 [6]:
𝑳 = [𝑹𝟏𝟐 , 𝑹𝟏𝟑 , ⋯ , 𝑹𝟏𝑴 , 𝑹𝟐𝟑 , ⋯ 𝑹𝟐𝑴 , ⋯ 𝑹(𝑴−𝟏)𝑴 ] (19)
In this paper, the reference vector 𝑳 is used as the input of the whole classification model, and [-90°,
+90°] is divided into 181 categories, and the probability vector 𝒁 of 181 categories is used as the output
of the model. In this paper, a neural network with three layers and 128 neurons in each layer is used to
predict the direction of arrival of unknown signals. The known training is set as
{(𝐿1 , 𝑍1 ), (𝐿2 , 𝑍2 ), ⋯ , (𝐿𝑖 , 𝑍𝑖 )}. 𝐿𝑖 is the input data, 𝑍𝑖 is the output data. The network is trained through
the known training data. In order to train samples better, the network must know the gap between the
predicted value and the real value of the model, so the importance of loss function is self-evident. The

3
IWAACE 2020 IOP Publishing
Journal of Physics: Conference Series 1550 (2020) 032066 doi:10.1088/1742-6596/1550/3/032066

classification model established in this paper uses sparse category cross entropy as the loss function,
whose function is almost the same as that of category cross entropy, but sparse category cross entropy
does not need to encode the target vector uniquely [7]. Because of the properties of sigmoid function,
the derivative of the commonly used mean square error loss function is very small in most of the time,
which leads to the slow updating of 𝑤 and 𝑏. However, the classification cross entropy can overcome
this deficiency. The classification cross entropy formula is as follows:
𝐸 = − ∑ ∑ 𝑦𝑛𝑘 log 𝑦̂𝑛𝑘 (20)
𝑛 𝑘
In formula (20), 𝑦𝑛𝑘 represents the expected value of the 𝑘𝑡ℎ element of the 𝑛𝑡ℎ data, and 𝑦̂𝑛𝑘 is the
actual output value of the neural network. The role of cross entropy is to measure the difference between
the current output probability distribution and the expected probability distribution of the network. The
larger the cross entropy value, the farther the current output situation is from the expected, the worse the
effect.
The activation function of the output layer of this paper uses 𝑠𝑜𝑓𝑡𝑚𝑎𝑥 . 𝑆𝑜𝑓𝑡𝑚𝑎𝑥 is the most
commonly used function to deal with multi-classification problems. It is a generalization of logistic
regression model on multi-classification problems [8]. The definition of 𝑠𝑜𝑓𝑡𝑚𝑎𝑥 function is expressed
as follows:
𝑒 𝑎𝑘
𝑦𝑘 = 𝑛 𝑎 (21)
∑𝑖=1 𝑒 𝑖
In formula (21), 𝑦𝑘 is the output of the 𝑘𝑡ℎ neuron, and N represents the number of neurons in the
output layer of the network. The formula indicates that the numerator of the softmax function is the
exponential function of the input signal 𝑎𝑘 , and the denominator is the sum of the exponential functions
of all input signals. Because not only the output of 𝑠𝑜𝑓𝑡𝑚𝑎𝑥 function is a real number between 0 and 1,
but also the sum of the output value of 𝑠𝑜𝑓𝑡𝑚𝑎𝑥 function is 1, the output of 𝑠𝑜𝑓𝑡𝑚𝑎𝑥 function can be
interpreted as "probability". Generally, neural network takes the corresponding category of the neuron
with the largest output value as the recognition result, which is the basis of the final judgment of the
direction angle of the target signal in this paper. Since [-90°, +90°] are divided into 181 categories in
this paper, the output layer neurons of this neural network are set to 181.

5. Computer simulation comparison

5.1. Simulation based on deep learning algorithm

In this paper, the array receiving model is built in MATLAB, using a single frequency target sound
source in ocean far field, whose frequency is 1kHz. The number of elements in linear array is 10, and
𝜆
the array spacing 𝑑 = 2 . The incidence angle traverses [-90°,+90°], each angle is tested and sampled
1000 times, and the correlation matrix is obtained to form the reference vector 𝑳 required by the test set
and data set. Therefore, the corresponding relationship between 181000 groups of reference vectors 𝑳
and the vector 𝒁 related to the probability of angle 𝜃 is obtained.
The neural network model is established under the 𝑡𝑒𝑛𝑠𝑜𝑟𝑓𝑙𝑜𝑤 learning framework based on python.
181000 groups of data are randomly divided into 101812 training sets, 33938 verification sets and 45250
test sets by 𝑡𝑟𝑎𝑖𝑛_𝑡𝑒𝑠𝑡_𝑠𝑝𝑙𝑖𝑡 function.
In this paper, simulation is carried out under different water environment conditions with signal-to-
noise ratio of 20dB, 10dB, 0dB and -10dB respectively. The results are shown in Figure (2-5).
It can be seen from Figure (2-4) that in the environment of positive signal-to-noise ratio, the training
effect of deep learning neural network model is very good, the accuracy is very high, and there is almost
no fitting phenomenon. At the 10th epoch, the accuracy is more than 90%, and the loss function has
reached a very low value, so the training speed is very fast.
It can be seen from Figure 5 that in the environment of poor signal-to-noise ratio, the training effect
of deep learning neural network model is not as good as before, the accuracy is about 75%, and the over
fitting phenomenon is worse than that in the case of positive signal-to-noise ratio. This is due to the fact

4
IWAACE 2020 IOP Publishing
Journal of Physics: Conference Series 1550 (2020) 032066 doi:10.1088/1742-6596/1550/3/032066

that the values of reference vectors in low SNR environment are very similar, the degree of
discrimination is not high, and the difficulty of feature extraction is increased.

Figure 2. The evaluation model in snr = Figure 3. The evaluation model in snr =
20dB. 10dB.

Figure 4. The evaluation model in snr = Figure 5. The evaluation model in snr = -
0dB. 10dB.

5.2. Comparison of simulation results with music algorithm

In the MATLAB environment, the experimental conditions and the depth learning algorithm are the
same, respectively take 20dB, 10dB, 0dB and -10dB SNR, incidence angle traversal [-90°,+90°], and
the same angle test 1000 times. The MUSIC algorithm is used to search the spectrum peak, and the
predicted angle value is output to compare with the incident angle. The accuracy under four different
SNR environments is calculated respectively, and compared with the accuracy of the depth learning
algorithm, as shown in Table 1.
Table 1 Comparison of DOA estimation between deep learning and MUSIC algorithm
Accuracy
SNR/dB
Deep learning algorithm MUSIC algorithm
20 98.35% 98.65%
10 97.85% 96.58%
0 94.22% 89.83%
-10 74.87% 53.63%
It can be seen from the table that in the environment of high SNR, the accuracy of deep learning
algorithm is equivalent to music algorithm, which can identify the far-field sound source direction well.
However, when the SNR is reduced, the deep learning algorithm has better anti-jamming ability and
higher accuracy than music algorithm. When the SNR is -10dB, the accuracy of deep learning is 50%

5
IWAACE 2020 IOP Publishing
Journal of Physics: Conference Series 1550 (2020) 032066 doi:10.1088/1742-6596/1550/3/032066

higher than music algorithm, which proves that deep learning has more advantages in low SNR
environment.

6. Conclusion
This paper discusses a method of DOA estimation, which is based on the depth neural network. It can
learn the autocorrelation matrix information of array signals at different incident angles, and build the
classification model. Through learning a large number of data, this method can accurately predict the
incident angle in the environment of high SNR and adaptively learn and adjust the model during the
reduction of low SNR, which is better than MUSIC algorithm. The method is proved to be feasible by
computer simulation.

References
[1] T. GUO, Y. WANG, L. ZHANG. A direction of arrival estimation method based on spectral
separation of the sample covariance matrix.[J]. JOURNAL 0F VIBRATION AND
SHOCK. 2018,37(12):23-28,47.
[2] T. GUO, Y. WANG, L. ZHANG, ZHANG Lichen. Underwater Acoustic Array SMI-MVDR
Spatial Spectral Estimation Based on Random Matrix Theory.[J]. JOURNAL 0F
VIBRATION AND SHOCK,2017.
[3] Y. Lecun, L. Bottou, Y. Bengio and P. Haffner, “Gradient-based learning applied to document
recognition,” in Proceedings of the IEEE, vol. 86, no. 11, pp. 2278-2324, Nov. 1998.
[4] M. M. SHI. Simulation Implementation of Direction of Arrival Estimation based on MUSIC
Algorithm.[J]. Communications Technology.2018,51(11):2570-2574.
[5] Duan, X.;Wang, Y.;Pedrycz, W.;Liu, X.;Wang, C.;Li, Z..AFSNN: A Classification Algorithm
Using Axiomatic Fuzzy Sets and Neural Networks(Article)[J].IEEE Transactions on Fuzzy
Systems.2018,Vol.26(No.5)：3151-3163.
[6] Q. N. ZHANG，H. L. ZHAO. Direction of Arrival Estimation Based on a Support Vector
Machines.[J]. Science Technology and Engineefing.2012,(6):1409-1411.
[7] Durand, Thibaut, Nazanin Mehrasa and Greg Mori. “Learning a Deep ConvNet for Multi-label
Classification with Partial Labels.” CoRR abs/1902.09720 (2019): n. pag.
[8] Reverdy, Paul;Leonard, Naomi Ehrich.Parameter Estimation in Softmax Decision-Making
Models With Linear Objective Functions.[J].IEEE Transactions on Automation Science &
Engineering.2016,Vol.13(No.1)：54-67.