Learning the Wireless V2I Channels Using Deep

Neural Networks
Tian-Hao Li1 , Muhammad R. A. Khandaker1 , Faisal Tariq2 , Kai-Kit Wong3 and Risala T. Khan4
1
School of Engineering and Physical Sciences, Heriot-Watt University, Edinburgh, United Kingdom
2
James Watt School of Engineering, University of Glasgow, United Kingdom
3
Department of Electronic and Electrical Engineering, University College London, United Kingdom
4
Institute of Information Technology, Jahangirnagar University, Dhaka, Bangladesh
Corresponding e-mail: [email protected]
arXiv:1907.04831v1 [eess.SP] 10 Jul 2019

Abstract—For high data rate wireless communication systems, two-way: infrastructure like traffic lights can provide informa-
developing an efficient channel estimation approach is extremely tion to cars and vice-versa. This communication system can
vital for channel detection and signal recovery. With the trend provide quantitative and real-time information that can be used
of high-mobility wireless communications between vehicles and
vehicles-to-infrastructure (V2I), V2I communications pose addi- for safety, mobility, and environmental benefits. When V2I
tional challenges to obtaining real-time channel measurements. communications are widely used, revolution of roadways will
Deep learning (DL) techniques, in this context, offer learning occur all over the world. For example, self-driving vehicles
ability and optimization capability that can approximate many will become reality and it is vital to make successful and safe
kinds of functions. In this paper, we develop a DL-based channel autonomous cars. To make this huge change happen, the basic
prediction method to estimate channel responses for V2I com-
munications. We have demonstrated how fast neural networks theory and techniques must now be developed urgently.
can learn V2I channel properties and the changing trend. The To enhance the performance of communication systems and
network is trained with a series of channel responses and known to solve signal processing and communications problems, deep
pilots, which then speculates the next channel response based on learning (DL) has recently drawn utmost popularity [4]–[7].
the acquired knowledge. The predicted channel is then used to A deep neural network (DNN) is an algorithm with learning
evaluate the system performance.
Index Terms—Vehicle-to-infrastructure; V2X; Channel estima-
ability and optimization capability that can approximate many
tion; Machine learning; Deep neural network. kinds of functions. Particularly, problems without any precise
numerical model can now be solved using DL methods [8].
I. I NTRODUCTION In order to leverage the advantages of using a large group
of data for communication performance improvement, several
Since the first generation (1G) wireless communication machine learning methods including supervised, unsupervised
network entered the market in the 1980s, the world has been and reinforcement learning have been proposed based on the
dramatically changed by the development of mobile commu- traditional approaches. The machine learning can be useful in
nication technologies. It has gone through several evolutions analyzing communication environment variance, making de-
in the past few decades, from 1G to 5G, and due to the cisions autonomously, transmission routing, network security,
huge potential demand all over the world, the technology will and system resource management [6], [8].
continue to upgrade rapidly [1]. While 2G, 3G and 4G were Recently, some works have concentrated in the area of chan-
about connecting people and parts of things, 5G will connect nel estimation using DL methods [9]–[13]. The authors in [10],
everything and it can provide unlimited access to anywhere, [11] proposed the back propagation (BP) learning algorithm
anytime, anybody and anything [2]. to build a multilayered perceptron (MLP) neural network as
In recent years, orthogonal frequency division multiplexing an estimator for OFDM communication channels. In [11], a
(OFDM) has become a popular choice for fast-speed and high- method of implicit channel state information (CSI) estimation
quality communication systems. However, channel modelling and direct recovery of transmission symbols based on deep
and channel estimation are two major challenges affecting learning has been proposed. The authors in [12] proposed an
the performance of OFDM systems. In order to estimate the approximate information passing network for millimetre-wave
channel response, pilot-based channel estimation is commonly massive multiple-input and multiple-output (MIMO) systems
adopted, in which a training sequence composed of known data based on learned denoising which is a deep learning model
symbols (pilots) is transmitted, and the channel parameters that can analyze channel structure and estimate channel from
are initially estimated using the received pilot signals [3]. a big training database. In [13], the authors developed a CSI
Minimum mean-square error (MMSE) and least squares (LS) feedback model by spreading a new learned CSI perception
are two traditional estimation approaches. and restoration architecture. By learning the spatial features
Vehicle-to-Infrastructure (V2I) communication is about the directly and combining the correlation of samples in time-
data transmission between vehicles and infrastructure on the varying MIMO channels, the system greatly improves the
roads. V2I communication system is normally wireless and quality of recovery and the trade-off between the compression
ratio (CR) and the recovery quality [13]. travels at a constant speed along a straight road. Hence, at the
All of the above works [9]–[13] have developed DL-based same position the channel is highly similar even at different
channel estimation techniques in traditional quasi-stationery time slots, and it can be considered as position related channel
wireless communication systems. However, with the trend of estimation which means that change of channel is related to
high-mobility wireless communications between vehicles and the position of vehicle. The channel power gain between the
vehicles-to-infrastructure (V2I), many rising problems appear vehicle at position n and the BS is defined as [15]
because of the high variances in communication environments,
h(n) = G(n)β(n)Ad(n)−γ , (1)
which are fundamentally different from traditional wireless
communication problems. For the communication with large where G(n) is assumed to be exponentially distributed fast
Doppler-shift and fast-varying channels, the above methods fading power gain, β(n) is the log normal shadow fading
may not work because in fast time-varying fading environment component, A is the constant pathloss, d(n) is the distance
channel response is difficult to obtain in real time and the between the vehicle and the BS, and γ is the pathloss expo-
outdated CSI exhibits a significant negative impact on the nent. For a frequency-selective multipath fading channel, the
performance. The critical point to fit high-mobility networks channel impulse response (CIR) is given by
is to develop future informatized and intelligent vehicles L−1
combined with machine learning to benefit the communication
X
h(n) = Cl ejφl (n) δ(n − τl ) (2)
networks [14]. l=0
In this paper, we address the adverse effects of imperfect = αl δ(n − τl ), (3)
CSI on V2I communication systems by developing a DL-based
channel prediction approach. Two major challenges facing where L isP the total number of multipath component and φl , τl ,
L−1
V2I channel estimation addressed in this paper are i) the and αl = l=0 Cl ejφl (n) are phase, time delay, and the time-
way to find CSI in real-time with the knowledge of vehicle dependent complex path gain of the lth multipath component.
information including position and velocity, ii) the technique For the V2I channel, the time-variant phase associated with
to estimate the fast time-variant channel properly and recover the lth path is defined as [16, Chapter 2]
the information with minimum error. The major contributions
φl (n) = φl − 2πcτl /λc + 2πfD,l n, (4)
of this paper are as follows:
• Firstly, an OFDM modulation based V2I channel esti- where c = fc λc is the speed of light, λc is the wavelength of
mation technique is introduced as a baseline to verify the arriving plane wave (carrier signal), τl = dl /c, dl is the
effectiveness of the DL-based approach. length of the lth path, and fD,l is the maximum Doppler shift
• Secondly, DL-based channel prediction method is pro- of the lth propagation path. For OFDM transmission, the mth
posed to deal with the channel estimation problem in V2I modulated time domain symbols can be expressed as
communication system in high-mobility environments. N −1
• Thirdly, the performance of the proposed DL-based 1 X
xm (n) = √ , Xm (l)ej2πnl/N , n = 0, · · · , N − 1,
approach is compared with other common algorithms N l=1
through extensive numerical simulations. (5)
where Xm (l) is the mth modulated symbol, n is the subcarrier
index and N is the total number of subcarriers. Cyclic prefix
is then appended to the symbols to prevent inter-symbol
interference (ISI). The signals are then transmitted through
the wireless channel and the signal received at the BS can be
expressed as
ym (n) = hm (n) ⊗ xm (n) + wm (n), (6)
where ⊗ indicates the convolution operation and wm (n) is the
additive white Gaussian noise for the mth subcarrier. At the
Figure 1. The proposed V2V and V2I communication system.
receiver, the cyclic prefix is removed first from the received
signal, followed by parallel conversion to frequency domain
II. S YSTEM M ODEL by applying fast Fourier transform:
We consider a high-mobility vehicular communication sys- N −1
1 X
tem (cf. Fig. 1) in which a vehicle intends to communicate with Ym (k) = ym (n)e−j2πkn/N , k = 0, · · · , N − 1. (7)
N n=1
a roadside infrastructure, which can be a traditional cellular
base station (BS), traffic lamppost, building or any other fixed Thus in the frequency domain, the input-output relationship
structure. At both BS and the vehicle, only one antenna for can be expressed as
transmitting and receiving is considered. For simplification,
only 2D horizontal plane is considered where the vehicle Y (k) = H(k)X(k) + W (k), k = 0, · · · , N − 1. (8)
Consequently, the system can be described as a set of N B. The MMSE Algorithm
independent parallel Gaussian channels: The MMSE channel estimator takes the effect of channel
noise into consideration. Meanwhile, the MMSE estimate of
yk = hk xk + wk , k = 0, · · · , N − 1. (9) g is given by
For convenience, we rewrite (10) using matrix notations as [3] ĝMMSE = Rgy R−1
yy y (15)

y = XFg + w, (10) where

Rgy = E gyH = Rgg FH XH ,

(16)
where X is a N × N diagonal matrix containing x ,
Ryy = E yyH = XFRgg FH XH + σw
2

T IN (17)
[xm (0) xm (1) · · · xm (N − 1)] as the main diagonal, g ,
T T (18)
[gm (0) gm (1) · · · gm (N − 1)] , w , [w0 w1 · · · wN −1 ] and
2

F0,0 ··· F0,N −1
 n o covariance matrices and σw is the noise
are the corresponding
2
.. .. .. variance E |wk | . Since the columns of F in (11) are
F, , (11)
 
. . . orthonormal, the frequency-domain MMSE estimate of h is
FN −1,0 ··· FN −1,N −1 given by [3]
i−1
is the DFT matrix with Fn,k = √1N e−j2πnk/N .
−1 2
h
ĥMMSE = FĝMMSE = FRgg FH XH XF σw + Rgg
The objective of this study is to estimate the CIR Hm (k)
−1 H H
from the observation of Ym (k) with known pilot signal × FH XH XF F X y. (19)
Xm (k). In the following, we will first discuss the conventional
IV. P ROPOSED M ACHINE L EARNING A PPROACH
channel estimation approaches, and then develop a deep learn-
ing based channel estimation algorithm for the V2I system (7). It has been shown in [11] that a multi-layer neural network
(MLNN) can provide a very good approximation of channels.
III. C ONVENTIONAL C HANNEL E STIMATION T ECHNIQUE Hence in the following, we propose a deep learning based
V2I channel estimation algorithm using back propagation
In conventional communication systems, both blind and technique.
non-blind channel estimation techniques have been considered
for estimating the CIR. Popular channel estimation algorithms
include maximum likelihood (ML), least mean square (LMS),
minimum mean square error (MMSE) and least square (LS)
methods. These methods have been studied thoroughly to
estimate CSI within a certain time and frequency range.
Although the LS algorithm has a worse performance in time-
varying environments compared to the other approaches, its
implementation is very simple. On the other hand, the MMSE
algorithm can be well-behaved in all general fading channels
with both frequency and time selectivity.

A. The LS Algorithm
Figure 2. The proposed deep neural network for estimating the V2I channel.
The LS algorithm tries to minimize the squared error (i.e.,
Euclidian distance) between the transmitted and the received
signals which is expressed as A. Defining the Neural Network
As shown in Fig. 2, the neural network contains two inputs
2 H and two outputs and nh hidden nodes. The inputs to the
ky − XFgk = (y − XFg) (y − XFg) . (12)
network are the received signals, and the estimated channel
Taking the derivative of (12) and equating the derivative to 0, response parameters are the outputs. The two inputs and
the LS estimate of the channel frequency response is given by outputs of the MLNN are connected to the real and imaginary
components of the corresponding complex numbers, since

−1
ĥLS = F FH XH XF FH XH y. (13) neural networks work only with real numbers.
For the machine learning based channel estimation scheme,
Since the optimal training sequence is orthonormal [17], [18], we consider a supervised learning approach which estimates
(13) eventually reduces to the CIRs using a fully connected neural network. Note that we
do not consider any bias inputs to the neurons at any layer.
ĥLS = X−1 y. (14) We apply sigmoid function for the activation of the hidden
neurons, while the activation function at the output layer is
linear. The output of each neuron is generated by computing
the weighted sum of the inputs coming into that node and then
applying the activation function.
(1)
Let us now assume that wi,j is the ith input to jth hidden
(2)
neuron weight and wj,k is the jth hidden node to the kth
output node weight. The hidden layer activation function is
the sigmoid function
1 ez
Φ(z) = = . (20)
1 + e−z ez + 1
Thus, the hidden layer output is defined as
1 Figure 3. Proposed DNN-based channel estimation method.
Ojh = Φ (inj ) = , j = 1, · · · , nh , (21)
1 + e−inj
Pnin in (1)
where inj = i=1 Xi wi,j , nin is the number of input layer Note that the gradient of output layer activation function is
nodes and Xiin is the ith input to the DNN. Similarly, each always unity. Similarly, the back propagated updates of the
output node computes its net output as input-to-hidden-layer weights can be expressed as
"n #
1 o
out ∂Ok
out ∂Ojh
Okout = Φ (ink ) =
X
, k = 1, 2, (22)


1 + e−ink ∇wjk = η wjk tk − Ok Xi . (27)
∂wjk ∂wij
k=1
Pnh (2)
where ink = j=1 Oj wj,k . Then, the sum of weighted inputs For training the DNN, we first generate a large set of pilot
to the nodes is applied to the output layer activation function symbols following a particular modulation scheme. We then
which is assumed to be a linear function. Thus the network use a series of historical channel responses between moving
outputs can be calculated as vehicle and BS for optimizing the weights of the neural net-
work. The training process continues until the target accuracy
 !
nh nin
(2) (1)
X X
Okout = Φ  wj,k Φ wi,j Xi  , k = 1, 2. (23) is reached. The overall training procedure is summarized in
j=1 i=1 Fig. 3.
B. Training the Neural Network V. N UMERICAL S IMULATIONS
In training process, the weights of all the layers are adjusted In this section, we perform numerical simulations to demon-
according to the mismatches between the outputs and the strate the effectiveness of the proposed machine learning based
targets. In each epoch, the direction of changes are that tending channel estimation method for wireless V2I communications.
to minimize the mean-squared error (MSE). Accordingly, we Throughout this section, we compare the performance of the
define the cost function as the MSE between DNN output and proposed approach against the conventional LS and MMSE
the target output as based estimation schemes in [3]. We first demonstrate the
o n estimation accuracy of the DNN method considering a low-
1X
2 mobility environment. We then test the trained DNN perfor-
ε= tk − Okout , (24)
2 mance for V2I channel estimation.
k=1
We assume that the multipath V2I channel has time delays
where tk is the kth desired output and no is the number
and data is transmitted using 4-QAM modulation. The neural
of outputs (2 in this case). The gradient descent based back
network is designed to have a single hidden layer with 10
propagation learning rule is exploited to optimize the weights
neurons and single output layer with two nodes. Sigmoid
for improving the DNN performance. The gradient descent
and linear activation functions has been utilized in the hidden
method minimizes the cost function ε by updating the weights
layer and the output layer, respectively. The DNN algorithm
in the opposite direction of the gradient of the objective
is implemented in two ways. The first one is built from sketch
function w.r.t. to the weights [19]. Thus the weight update
using the structure illustrated in Section IV (identified as
in each iteration is given by
‘build’ in the figures), and the second one is implemented
∂ε based on the ANN toolbox in MATLAB, whose performance
∇w = −η , (25)
∂w has also been compared in simulations where applicable.
where 0 < η < 1 is the learning rate that determines the size Fig. 4 shows the performance of the proposed DNN ap-
of the steps in each iteration. Thus the update of hidden-to- proach in the training process according to normalized MSE
output-layer weights can be expressed as at different SNR environment. We run for 20 epochs for each
training data set and It is clear that the error between network

∂Okout h
∇wjk = η tk − Okout O . (26) outputs and the targets is being minimized as the training
∂wjk j progresses. It can also be found that with higher SNR, which
 Figure 4. MSE performance during the training process. Figure 6. MSE at different phases of the DNN appraoch.

Figure 5. BER performance during the training process. Figure 7. Training progression over different epochs.

means the noise power is low compared with signal power, achieved at the 49th epoch with the minimum error distribu-
the error of the estimated channel response becomes lower. tion. Fig. 8 shows the histogram of each subset. From Fig. 9,
Then, Fig. 5 compares the bit error rate (BER) performances it can also be seen that the outputs fit the targets very well.
of all the methods. It can be seen that the LS algorithm Finally, we plot the BER performance of the algorithms
results in the worst behaviour compared to the others and it in Fig. 10. It can be seen from the results that the BER
cannot correctly estimate the channel especially when SNR is performance of the traditional DNN methods is significantly
very low. From this figure it is clear that the performance of worse than the other methods. The reason for this is that the
proposed back propagation DNN approaches are much better channel estimation in V2I communication is influenced by
than the LS algorithm at low SNR values as well as high many factors, such as fast variation of channels, large Doppler-
SNR values. The results also show that both DNN approaches shift, high interference and noise. These results further justify
perform close to each other in a low-mobility environment, and the effectiveness of the proposed DL-based approach for V2I
are very close to the MMSE algorithm. Obviously, the MMSE channel estimation.
algorithm performs the best according to the simulation results.
Next, simulations of a V2I wireless communication system VI. C ONCLUSION
in a fast time-varying fading channel are carried out to We have introduced a deep learning based channel esti-
demonstrate the performances of different channel estimation mation approach for V2I wireless communication systems.
methods. Figs. 6 to 9 show the training process of DNN for We have demonstrated that the channels in high-mobility
V2I channel prediction. In these simulations, the entire data set environment can be estimated using DNN based prediction
is divided into three subsets namely: training (70%), validation methods with a group of historical CIR to solve the outdated
(15%) and testing (15%), respectively. During the training CSI problem. Extensive simulation results illustrate that the
process, the MSE reduces sharply and the best estimate is proposed channel prediction method is able to dramatically
 improve the performance of channel estimation in particular
in high-mobility environment. Considering the nonuniform
movement of vehicles, including variant position and changing
velocity in the training process of DNN can be an interesting
future work.

R EFERENCES
[1] F. Tariq, M. R. A. Khandaker, K.-K. Wong, M. Imran,
M. Bennis, and M. rouane Debbah, “A speculative study on
6G,” IEEE Commun. Magazine, June 2019 (submitted). Available:
https://arxiv.org/pdf/1902.06700.pdf.
[2] J. G. A. et al., “What will 5G be?” IEEE J. Sel. Areas Commun., vol. 32,
pp. 1065–1082, June 2014.
[3] J. J. van de Beek, O. Edfors, M. Sandell, S. K. Wilson, and P. O.
Borjesson, “On channel estimation in OFDM systems,” in Proc. IEEE
45th Veh. Technol. Conf., vol. 2, Chicago, IL, USA, 1995, pp. 815–819.
Figure 8. Histogram of errors at different processes. [4] C. Jiang, H. Zhang, Y. Ren, Z. Han, K.-C. Chen, , and L. Hanzo,
“Machine learning paradigms for next-generation wireless networks,”
IEEE Wireless Commun., vol. 24, pp. 98–105, 2016.
[5] C. Zhang, P. Patras, and H. Haddadi, “Deep learning in mobile and
wireless networking: A survey,” IEEE Commun. Surveys & Tutorials,
2019.
[6] J. Gao, M. R. A. Khandaker, F. Tariq, K.-K. Wong, and R. T. Khan,
“Deep neural network based resource allocation for V2X communi-
cations,” in IEEE 90th Vehicular Technology Conference: VTC2019-
Fall, 22-25 Sep. 2019, Honolulu, Hawaii, USA (submitted). Preprint
arXiv:1906.10194.
[7] H. Ye and G. Y. Li, “Deep reinforcement learning for resource allocation
in V2V communications,” in Proc. IEEE Int. Conf. Commun. (ICC),
Kansas City, MO, 2018.
[8] O. Simeone, “A very brief introduction to machine learning
with applications to communication systems,” Available online:
https://arxiv.org/abs/1808.02342, 2018.
[9] M. Mehrabi, M. Mohammadkarimi, M. Ardakani, and Y. Jing,
“Decision directed channel estimation based on deep neural net-
work k-step predictor for MIMO communications in 5G,” Available:
https://arxiv.org/abs/1901.03435, Jan. 2019.
[10] N. Taspinar and M. N. Seyman, “Back propagation neural network
approach forchannel estimation in OFDM system,” in Proc. Wireless
Commun., Netw. Inf. Security (WCNIS), June 2010, pp. 265–268.
[11] H. Ye, G. Y. Li, and B.-H. Juang, “Power of deep learning for channel
estimation and signal detection in OFDM systems,” IEEE Wireless
Figure 9. Training regression.
Commun. Lett., vol. 7, pp. 114–117, Feb. 2018.
[12] H. He, C. Wen, S. Jin, and G. Y. Li, “Deep learning-based channel
estimation for beamspace mmWave massive MIMO systems,” IEEE
Wireless Commun. Lett., vol. 7, pp. 852–855, Oct. 2018.
[13] T. Wang, C. Wen, S. Jin, and G. Y. Li, “Deep learning-based CSI
feedback approach for time-varying massive MIMO channels,” IEEE
Wireless Commun. Lett., vol. 8, pp. 416–419, Apr. 2019.
[14] L. Liang, H. Ye, and G. Y. Li, “Toward intelligent vehicular networks:
A machine learning framework,” IEEE IoT Journal, vol. 6, pp. 124–135,
Feb. 2019.
[15] L. Liang, G. Y. Li, and W. Xu, “Resource allocation for D2D-enabled
vehicular communications,” IEEE Trans. Commun., vol. 65, pp. 3186–
3197, July 2017.
[16] G. L. Stüber, Principles of Mobile Communication, 4th ed. Springer,
2017.
[17] Y. Rong, M. R. A. Khandaker, and Y. Xiang, “Channel estimation of
dual-hop MIMO relay systems using parallel factor analysis,” IEEE
Trans. Wireless Commun., vol. 11, pp. 2224–2233, Jun. 2012.
[18] Y. Rong and M. R. A. Khandaker, “Channel estimation of dual-hop
MIMO relay system via parallel factor analysis,” in Proc. 17th Asia-
Pacific Conf. Commun. (APCC’2011), Sabah, Malaysia, Oct. 2-5, 2011.
[19] S. Ruder, “An overview of gradient descent optimization algorithms,”
Figure 10. BER versus SNR for the testing data. arXiv preprint. Available: arXiv:1609.04747, June 2017.