Enhanced MIMO-OFDM Channel Estimation in

mmWave Systems: A Bi-LSTM and CNN Hybrid

Approach
Priyanshu Sandeep Kumar Singh Abhishek Dhiman
ECE Department ECE Department CSE Department
NIT HAMIRPUR (H.P)-177005 NIT HAMIRPUR (H.P)-177005 NIT HAMIRPUR (H.P)-177005
194524@nith.ac.in sksingh@nith.ac.in 195565@nith.ac.in

Abstract—Deep learning provides aapotent tool to counter • Spatial-Frequency-TemporalaCNNa(SFT-CNN): Fur-

the complexities of the channel estimation problemain hybrid- ther utilizes temporal correlations through Bi-LSTM lay-
precoded millimeter-wave (mmWave)amassive MIMO sys- ers, capturing dynamics in time-varying channels.
tems. In this paper, the synergistic integration of Convolu-
tionalaNeuralaNetworks (CNNs)aand BidirectionalaLong Short- Extensive experimental resultsademonstrateathat SF-CNN
TermaMemory (Bi-LSTM) [1] networksais used to propose three andaSFT-CNN have lower complexity than prior arts and
new channel estimation models. The SF-CNN concurrently uses show better performance than GMSE, but still inferior
spatial and frequency correlations by processing adjacent sub- to theaideal butaunpracticalaMMSEaestimator. Importantly,
carriers together. The Spatial-Frequency-Temporal CNN (SFT-
it can operate effectively in various scenario types. This
CNN), based on SF-CNN, adds the temporal dimension by
employing Bi-LSTM layers,aresulting in a substantial accuracy work introduces a novel approach forachannel estimationafor
increase. Extensive numerical experiments across a variety of sce- mmWaveamassiveaMIMO-OFDM systems [4], thus providing
narios demonstrate thataSF-CNN and SFT-CNN are superior to the foundation for the design of high-throughput and reliable
conventional estimators in terms of performance and complexity, network of the future.
approaching the optimal but impractical MMSE estimator. SFT-
CNN shows comparable accuracy to other models, demonstrating II. S YSTEM M ODEL
its potential forapractical use. These results provide a solid
foundation for the effectiveness of deep CNNs with Bi-LSTM
integration in fully exploiting channel correlations for robust
performance in mmWave massive MIMO channel estimation.
Index Terms—Bi-LSTM, SFT-CNN, SF-CNN

I. I NTRODUCTION
Millimeterawave (mmWave)acommunicationais capable of Fig. 1. System Model OF BiLSTM Integrated MIMO OFDM
delivering a large amount of information, but its foremost dis-
advantage is the high path loss. The dominating SIR becomes In this paper, we analyze the millimeter wave (mmWave)
the trend that causes severe challenges toward massive MIMO system for a massive MIMO-OFDM architectures [4]. The
implementation, as this technique does need bullvine space transmitteraandareceiver are connected, as illustrated in Fig
for dedicated radio frequency chains. Fewer channels which 1, by a large numberaofaantennas (NT at theatransmitter,
are commonly known as hybrid topologies, due to their better NR at theareceiver), which are only connected to a smaller
error margin, also come at the cost of accuracy in channel number of radio frequency chains (NTRF at the transmitter
estimation. and NRRF at the receiver). This mixed architecture provides
While the old technologies can get worse in case of complex RF chains at a low cost per antenna and therefore achieves
scenarios [2], deep learning makes up for it with a novel high channel gain, but introduces complexity into the channel
alternative. Thisapaper presents a new deepalearning solution estimation phase since the numberaof RFachains is lesser.We
for massive MIMO-OFDM channel estimation [3] using CNNs thus assume that NT >> NTRF andNR >> NRRF . Ac-
andaBi-LSTM networks in mmWave communication systems. cordingato [5], theaNR NT channelamatrix betweenathe re-
This combination allows for an accurate channel estimation ceiveraand theatransmitterain theadelay domainais givenaby
with a low computational cost. r L
NT NR X
We develop three innovative models: H(τ ) = αl δ(τ − τl )aR (φl )aH
T (ϕl ), (1)
L
• Spatial-FrequencyaCNNa(SF-CNN): Exploits spatial l=1
andafrequency correlations by processing adjacent sub- inawhichaL is aanumber ofamain paths,αl ∼
carriers together. CN (0, σα2 ) isathe propagationagain ofathe lthapath
withaσα2 abeingatheaaverageapower gain, τl isathe delayaof to make estimates. The channels that correspond to the same
thealth path, φl and ϕl ∈ [0, 2π] are, arespectively, time-frequencyapositions.Based in the estimatedachannels, in-
theaazimuth anglesaof arrival andadeparture (AoA/AoD)aat terpolationacan be used to get channels on theapilotless posts.
theatransmitter andathe receiver. Theaequivalent This is so evident. the accuracy ofainterpolation relies on the
responderavectors forauniform linearlyaarray (ULA) [4] accuracy ofainterpolation, expected channels and the rangeaof
mayabe writtenaas channels.Therefore, we will bring increased precision inathe
1 d d
pilot based channel. accuracy in this paper to achieve reference
aR (φl ) = √ [1, e−j2π λ sin(φl ) , . . . , e−j2π λ (NR −1) sin(φl ) ]T values that are more trusted. can be replaced for interpolation.
NR
(2)
1 d d III.
aR (ϕl ) = √ [1, e−j2π λ sin(ϕl ) , . . . , e−j2π λ (NT −1) sin(ϕl ) ]T SF-CNN-B I LSTM A BASED A CHANNEL A ESTIMATION
NT
(3) In thisasection weaelaborate the SF-CNN BiLSTM based
where daandaλ stand for the carrier wavelength and the channel estimation,theaofflineatraining of SF-CNN BiLSTM
distance between neighboring antennas,arespectively. The fre- and the online deploymentaof the model.Thenathe computa-
quency domain channel of the kthasubcarrierain OFDM [4] is tionalacomplexity for the onlineaestimation isaanalyzed.
provided by the channel model .
r L
A. ChannelEstimationaProcedure with Bi-LSTM-Integrated
NT NR X k H
SF-CNN
H(τ ) = αl e−j2τl fs π K aR (φl )aT (ϕl ) , (4)
L The channel estimateaprocessafor neighboring Q(= 2) sub-
l=1

where fs denotesathe samplingarate andaK is the num- carriers k0 and k0 +1 isashown in Figure 1. Assumptionsahold
beraof OFDMasubcarriers. In order calculate Hk,athe trans- withoutaloss √ of generality, including Wk = W, Fk =
mitterapertains toaonly one RFachains to transmitathe beam FaandaXk = P I, for k ∈ k0 , k0 + 1 andaP denotes trans-
over all the channelsausing only one beam while theareceiver mitapower.aThe pilotasignal matrix, Yk , isaformulated as:
combines theareceived pilotasignal by meansaof all RF √
Yk = P WH Hk F + N ek (6)
chainsaassociated withadifferent beams. At the elementary
level, transmitter sends beamforming vectors , fk,u ∈ where Hk adenotesathe true channel matrix for subcarrier
C NT X1 , pilots, xk,u , with MT beamforming vectors, u = k, F represents the precoding matrix, W isathe beamform-
1, . . . , MT . unlike the transmit beamforming,aeach beam- ingamatrix, and N e k is the additive whiteaGaussiananoise
formingavector, fk,u , the receiver uses MR avectors wk,v = (AWGN)aat the receiver.aSubsequently, Yk undergoes the
1, ..., MR , v = 1, ..., NR to them. Since the receiver’s NRRF (< following steps:
MR ) RFachains canaonly use NRRF combining vectors per 1.Preprocessing and Tentative Estimation:
channel use, this situation makes the receiver unsymmetrical. The tentative estimationa(TE) module utilizesamatrices GL
Thereafter, if, by using all the combinations of MR beams and GR to generate a coarse estimation of Hk [4]:
as processing vectors to derive pilot carrying beamform- √
ing vector, the channel uses will become [ NMRF R
] . There- Rk = GL Yk GR = P GL WH Hk FGR + GL N e k GR (7)
R
fore the total channel delay of all beamforming vectors is where the selection of GL and GR depends on the antenna
MT [ NMRF R
] . This is followed by the pilot signal matrix of the configurations (MR , NT ):
R
kthasubcarrieraat theabaseband receiver then the pilot signal
matrix can beawritten down [4]. GL = Wif MR < NR (8)

Yk = W H
k Hk Fk Xk + Nk (5)
(WWH )−1 Wif MR ≥ NR (9)
where Wk = [Wk,1 , 1, ..., wk , MR ]and Fk = [fk,1 , ..., fk , MT ]
are combining matrixaandabeamformingamatrix, respectively,
FH = Wif MT < NT (10)
Xk is an MT MT diagonal matrix, whoseauth element is
xk,u .aN̄k = WHk Nk means word coefficient after combination
and Prior to combination, Nk contains additive white Gaus- FH (FFH )−1 if MT ≥ NT (11)
siananoisea(AWGN) havingaCN(0, 1) components. Frequency
domain as well as time domain insertion of the pilot is 2.Bi-LSTM Integration:
taken into account here. In particular, adjacent subcarriers • [1]Concatenate Rk and R(k+1) (both preliminary esti-
Q > 2(Q ≥ 2)are separated. start any persuasive paper mations):
with the same length pilot. period to construct a pilot sub- Xk = [Rk, Rk + 1] (12)
carrier block and the rest of the time slots. there is data
transmission in each coherence interval. The first and the • Pass Xk toaa Bi-LSTM layerawith sharedaparameters:
second pilot subcarrier blocks are spaced from each other by
Qd = (Q−d) ≥ 0 subcarriers that carry data. Pilots are trained Zk = Bi-LSTM(Xk , Φb ) (13)
 where Zk denotesatheaoutput of the Bi-LSTM layer, and • Process the four matrices through multiple convolutional
Φb represents thealearnable parameters of the Bi-LSTM. layers with ReLU activation and zero padding.
3. SF-CNN Processing: • Employ batch normalization for stability and gradient
flow.
• The SF-CNN utilizesZk as input
4. Output Layer:
Hk,
b Hk b + 1 = SF-CNN(Zk , Φc ) (14)
• Use convolutional filters to process the final feature maps.
where H b k and Hb k+1 are the estimated channel matrices • Apply hyperbolic tangent activation to map the output to
for subcarriers k and k+1, respectively, and Φc represents the the interval [-1, 1].
learnable parameters of the SF-CNN. 5. Post-processing:
IMPROVEMENTS: • Scale and combine real and imaginary parts to obtain
′ ′
This integration of Bi-LSTM with SF-CNN utilizes both estimated channel matrices H b (k ,k +1)
i
spatial and temporal correlations, which leads to the possibility Loss Function and Optimization:
to significantly outperform the results with SF-CNN alone During training, the SF-CNN minimizes a loss function like
in particular where the temporal dynamics are strong. It MSE or NRMSE between the estimated and true channel
is achieved by catching the temporal correlations between matrices, using an optimizer like Adam or RMSprop to update
the adjacent subcarriers which were just the beginning of its parameters.
estimating the overall channel.
Q
Ntr X
1 X ′ ′ ′ ′
B. BiLSTM INTEGRATEDaSF-CNN OFFLINEaTRAINING MSE Loss = ||Hik ,k +q−1 − Ĥki ,k +q−1 ||2F
Ntr c2 i=1 q=1
1) Training Set Generation: The training set for the pro- (15)
posed Bi-LSTM-integrated SF-CNN consists of Ntr samples An optimizer like Adam or RMSprop can be used to update
generatedaaccording toaa specific channel model in a simu- the network parameters based on the calculated loss.
lationaenvironment. Each sample (i) comprises input data Ri Design Considerations:
and target data Hi : The SF-CNN architecture leverages principles from CNN-
• Ri :Input data, aathree-dimensional complex matrix in based image processing while adapting them to the specific
C (NR ∗NT ∗Q) ,where Q represents the number of adjacent channel estimation task. It acts as a channel denoiser, and
sub-carriers processed together. For each ′ subcarrier k’ therefore, the size of feature maps in each layer is set
(k ,k′ +1)
(where k’ ϵ 1, 2, ..., K-1), Ri includes Ri , which toNR ∗ NT .
are the tentatively estimated channel matrices obtained • Nine convolutional hidden layers: Aim to comprehen-
through Eq. (7). sively capture the channel’s inherent structure.
• Hi :Target data,aanother three-dimensional complex ma- • 3 ∗ 3 filters:Align with [6]’s findings of achieving good
trix in C (NR ∗NT ∗Q) . For
′ ′
each sub-carrier k’ (k’ ϵ 1, 2, ..., performance with low complexity.
K-1), Hi includes Hik ,k +1 , which are the corresponding • Limited layer/filter increase:Empirical observations
true channel matrices. suggest diminishing performance gains and increased
A scaling constant c> 0 is appliedato adjust the value complexity beyond the chosen setup.
range of the real and imaginary parts of Hi , ensuring By integrating the Bi-LSTM and carefully designing the SF-
compatibility with the activation function used in the SF- CNN, this training process exploits both spatial and tem-
CNN’s output layer. poral correlations, potentially leading to significant accuracy
2) Network Architecture: The overall training pro- improvements over using the SF-CNN alone, especially in
cessaincorporatesaa Bi-LSTM module beforeafeeding data dynamic channel scenarios.
into the SF-CNN:
1.Bi-LSTM Module: C. BiLSTM INTEGRATED SF-CNN ONLINE DEPLOYMENT
(k′ ,k′ +1) Next, the Bi-LSTM-integrated SF-CNN and the tentative
• Takes Ri (real and imaginary parts concatenated)
for each subcarrier k’ as input. estimation (TE) module will be employed at the receiver to
• Uses shared parameters across subcarriers to capture
process pilot matrices Yk0 , Yk0 +1, ..., Yk0 +Q−1 and generate
temporal dependencies between adjacent subcarriers. an estimate of channel matrices H bk , H
0
b k + 1, On the other
0

• Outputs a combined representation Zi

k′
incorporating hand, real channel information for fine-tuning can often be
spatial-temporal information difficult to come by. But it also means bigger online fine-
tuning overhead and bring the channel closer to the original
2. SF-CNN Input:
channel sequence. Interestingly, as revealed in the figures 6
k′
• Concatenate Zi from the Bi-LSTM for all sub-carriers and 7 under Section V, the Bi-LSTM-SF-CNN which was pre-
k’ to form a three-dimensional tensor. trained offline exhibited good tolerance to different unseen
• Separate real and imaginary parts, resulting in four real- channel statistic. This indicates that further online fine-tuning
valued matrices. may bring few performance gains and often unnecessary in
3. Convolution Layers: useful situations.
 IV. B I LSTM INTEGRATED SFT-CNN C HANNEL where σ and tanh are activation functions, xt is the input
E STIMATION vector, ht is the hidden state, and the Ws , Us , and bs are
This section proposes two channel estimation approaches weight matrices and bias vectors for each gate and cell
leveraging both spatial and temporal correlations: a Bi-LSTM- state update.
integrated SFT-CNN . Bi-LSTM-Integrated SFT-CNN: 2. Estimation Accuracy:
This approach exploits the inherent temporal correlations Normalized Mean Squared Error (NMSE) is used to evaluate
within time-varying channels by incorporating a Bi-LSTM performance:
module before feeding data into the SFT-CNN.
Time-varying Channel Model: E||H − Ĥ||2F
NMSE = (23)
In time-varying scenarios, the frequency-domain channel at ||H||2F
the kth subcarrier becomes:
r where H and H b represent the true and estimated channels,
NT NR X πf k
respectively.
Hk(t) = l = 1L αl e−j2π( K −νl t) aR (ϕl )aH
T (ϕl ),
L 3. Discussion:
(16)
While LSTM-CNN incorporates dedicated temporal modeling,
where Vl denotes the Doppler shift of the lth path and αl is
its complex architecture leads to significantly higher training
the corresponding path gain [7].
and testing times compared to both SF-CNN and the proposed
Temporal Correlation Modeling:
Bi-LSTM-integrated SF-CNN.
Following [6] and [7], we model the temporal correlation
Furthermore, studies in Section VI demonstrate that:
between channels in successive coherence intervals as:
p • Bi-LSTM-integrated SF-CNN effectively leverages tem-
Hk [n] = ρHk [n − 1] + 1 − ρ2 Θ[n], (17) poral correlations for improved accuracy while maintain-
ing comparable complexity to SF-CNN, surpassing both
where Hk [n] = Hk (nT ) is the discrete-time version of Hk (t),
SF-CNN and LSTM-CNN in performance.
Equation (17)highlights the temporal dependence between
• Its simple architecture and low complexity make it well-
channels in consecutive intervals, suggesting the potential for
suited for practical channel estimation tasks, especially
improved estimation by exploiting this correlation.
in dynamic mmWave channels.
V. C OMPARING C HANNEL E STIMATION A PPROACHES :
VI. NUMERICAL RESULTS
SF-CNN, LSTM-CNN, AND B I -LSTM- INTEGRATED
SF-CNN We investigate the CNN-based channel estimation perfor-
This section compares three channel estimation approaches: mance using a high-performance system that is fitted with
SF-CNN, LSTM-CNN, and the proposed Bi-LSTM-integrated state-of-the-art NVIDIA GPUs and optimized frameworks
SF-CNN, focusing on estimation accuracy, complexity, and such as TensorFlow/PyTorch. With a 100,000 sample dataset
their suitability for exploiting temporal correlations. 1. and customized training settings, we assessed our methods on
Architecture and Equations: different channel conditions with different metrics (NMSE,
PSNR, SDR, BER). To demonstrate our methods’ efficacy,
we also evaluated computing efficiency and contrasted them
• SF-CNN: Employs multiple convolutional layers to di-
with the most advanced approaches.
rectly process channel representations (Rk in Eq. (7)),
but lacks dedicated temporal modeling. A. Bi-LSTM Integrated SF-CNN Channel Estimation: Supe-
• LSTM-CNN: Combines Long Short-Term Memory rior Performance and Robustness
(LSTM) layers for temporal modeling with convolutional
This section explores the performance and robustness of
layers for spatial processing. LSTM equations involve
the proposed Bi-LSTM-integrated SF-CNN for channel esti-
gates and cell state updates [8], making them more
mation, showcasing its advantages over standard methods.
complex than regular convolutional layers.
Impressive Accuracy and Complexity Benefits:
• Bi-LSTM-integrated SF-CNN (proposed): Incorporates
a Bi-LSTM module before feeding features into the SF- • Figure 2 demonstrates that the Bi-LSTM-integrated SF-

CNN. The Bi-LSTM captures temporal dependencies CNN significantly outperforms non-ideal MMSE in the
using equations similar to standard LSTMs [1]: urban micro (UMi) street non-line-of-sight (NLOS) [6]
scenario for two adjacent subcarriers.
it = σ(Wi xt + Ui ht−1 + bi ) (18) • By leveraging frequency correlation, it even approaches
the performance of ideal MMSE with true covariance,
ft = σ(Wf xt + Uf ht−1 + bf ) (19)
especially at low and medium SNRs.
ot = σ(Wo ht + bo ) (20) Exceptional Robustness to Channel Variations:
• Figure 3 highlights the remarkable robustness of the
ct = ft ⊙ ct−1 + it ⊙ tanh(Wc xt + Uc ht−1 + bc ) (21)
Bi-LSTM-integrated SF-CNN against varying channel
ht = ot ⊙ tanh(ct ) (22) statistics.
 Fig. 4. Robustness of SF-CNN based approaches to different numbers of
Fig. 2. Robustness of MMSE and SF-CNN based approaches to different main paths
scenarios

Fig. 5. Robustness of MMSE and SF-CNN based approaches to different

Fig. 3. NMSE versus SNR for the SFT-CNN based Channel Estimation and scenarios
MMSE channel estimation

• Figure 5 showcases the significant impact of temporal

• Trained in UMi street NLOS, it maintains strong perfor- correlation on NMSE performance in the UMi street
mance when tested in both UMi street NLOS and urban NLOS scenario [6].
macro (UMa) NLOS scenarios, while MMSE approaches • Bi-LSTM SFT-CNN achieves superior accuracy com-
suffer significant degradation due to their dependence on pared to both SF-CNN and non-ideal MMSE by effec-
specific channel statistics. tively utilizing channel information from previous coher-
ence intervals and adjacent subcarriers.
Adaptability to Different Path Numbers:
• This improvement comes at minimal additional complex-
• Figure 4 evaluates the robustness to the number of main ity, demonstrating the efficient integration of temporal
paths (L). Despite being trained with L=3, the Bi-LSTM- features.
integrated SF-CNN effectively handles scenarios with • Notably, Bi-LSTM SFT-CNN even approaches the ideal
L=1, 2, and even 4. MMSE performance at low and medium SNRs, showcas-
• Notably, it surpasses least squares (LS) channel estima- ing its effectiveness in various settings.
tion when L=4 and SNR is less than 18 dB, demonstrating
Exceptional Robustness Against Channel Variations:
its flexibility.
• Figure 6 emphasizes the remarkable robustness of Bi-
LSTM SFT-CNN against mismatched channel statistics.
B. Bi-LSTM SFT-CNN: Leveraging Temporal Correlations for
• While both ideal and non-ideal MMSE suffer substantial
Superior Performance and Robustness
performance degradation when applying a UMi street
This section explores the Bi-LSTM SFT-CNN channel es- NLOS model to the UMa NLOS scenario, Bi-LSTM SFT-
timation approach, highlighting its ability to exploit temporal CNN maintains its strong performance.
correlations for improved performance and robustness. • This resilience stems from its ability to learn inherent
Enhanced Accuracy Through Temporal Modeling: channel structures, making it less sensitive to specific
 statistical variations. mismatched channel statistics, exceeding both the Bi-LSTM
SF-CNN and non-ideal MMSE. These results show the great
potential of deep learning for mmWave channel prediction,
opening the door to mmWave communication systems that
are more accurate and efficient.
Future Directions: Refining Robustness and Exploring
Deeper Insights Future work can explore broader channel
robustness through deeper analysis of complex mmWave
MIMO structures, guiding DNN design and tuning for even
more efficient and adaptable communication systems.

R EFERENCES
[1] Chengying Zhao, Xianzhen Huang, Yuxiong Li, and Muhammad
Yousaf Iqbal. A double-channelahybrid deep neural network based
on cnn andabilstm foraremaining useful life prediction. Sensors,
20(24):7109, 2020.
Fig. 6. Robustness of SF-CNN based approaches to different numbers of [2] Zhijin Qin, Hao Ye, Geoffrey Ye Li, and Biing-Hwang Fred Juang.
main paths Deepalearningain physicalalayer communications. IEEE Wireless Com-
munications, 26(2):93–99, 2019.
[3] Yucheng Wang, Wei Xu, Hua Zhang, and Xiaohu You. Wide-
Adaptability to Different Path Numbers: bandammwave channel estimation for hybrid massiveamimo with low-
precisionaadcs. IEEE Wireless Communications Letters, 8(1):285–288,
2018.
[4] Hua Dong, aPeihao andaZhang, Geoffrey Ye Li, aNavid NaderiAlizadeh,
and Ivan Simões Gaspar. Deepacnn for wideband mmwave massive
mimo channel estimationausing frequency correlation. In ICASSP 2019-
2019 IEEE International Conference on Acoustics, Speech and Signal
Processing (ICASSP), pages 4529–4533. IEEE, 2019.
[5] Zhen Gao, Chen Hu, Linglong Dai, and Zhaocheng Wang. Chan-
nelaestimation for millimeter-wave massiveamimo with hybrid precod-
ing overafrequency-selective fadingachannels. IEEE Communications
Letters, 20(6):1259–1262, 2016.
[6] 3GPP. Study on channelamodel for frequenciesafrom 0.5 to 100 ghz,
2018.
[7] Kaiming He and Jian Sun. Convolutionalaneural networks
ataconstrained time cost. In Proceedings of the IEEE conference on
computer vision and pattern recognition, pages 5353–5360, 2015.
[8] Xingjian Shi, Zhourong Chen, Hao Wang, Dit-Yan Yeung, Wai-Kin
Wong, and Wang-chun Woo. Convolutionalalstm network: A machine
learningaapproach for precipitation nowcasting. Advances in neural
information processing systems, 28, 2015.
Fig. 7. Robustness of SFT-CNN based approaches to different numbers of
[9] Ahmed Alkhateeb, Omar El Ayach, Geert Leus, and Robert W Heath.
main paths
Channelaestimation and hybrid precodingafor millimeter wave cellu-
larasystems. IEEE journal of selected topics in signal processing,
• Figure 7 evaluates the robustness of Bi-LSTM SFT-CNN 8(5):831–846, 2014.
[10] Hao Ye, Geoffrey Ye Li, and Biing-Hwang Juang. Power of
to varying numbers of main paths (L). deepalearning for channel estimationaand signal detection inaofdm sys-
• Despite being trained with L=3, it demonstrates an overall tems. IEEE Wireless Communications Letters, 7(1):114–117, 2017.
improvement in accuracy compared to the SF-CNN, even [11] Chao-Kai Wen, Wan-Ting Shih, and Shi Jin. Deepalearning foramassive
mimo csi feedback. IEEE Wireless Communications Letters, 7(5):748–
when tested with mismatched L values (1, 2, and 4). 751, 2018.
• This adaptability further strengthens the case for Bi- [12] Tianqi Wang, Chao-Kai Wen, Shi Jin, and Geoffrey Ye Li.
LSTM SFT-CNN as a practical solution across diverse Deepalearning-based csi feedback approachafor time-varying massive
mimo channels. IEEE Wireless Communications Letters, 8(2):416–419,
channel conditions. 2018.
[13] Peihao Dong, Hua Zhang, and Geoffrey Ye Li. Machinealearning
VII. C ONCLUSION : A DVANCING MM WAVE C HANNEL prediction based csiaacquisition for fdd massive mimoadownlink. In
E STIMATION WITH D EEP L EARNING 2018 IEEE Global Communications Conference (GLOBECOM), pages
1–6. IEEE, 2018.
Deep learning-based methods for mmWave large MIMO- [14] Yo-Seb Jeon, Song-Nam Hong, and Namyoon Lee. Blind detectionafor
OFDM systems were investigated in this research, leading mimo systems withalow-resolution adcs using supervised learning. In
2017 IEEE International Conference on Communications (ICC), pages
to notable gains in performance and increased flexibility. 1–6. IEEE, 2017.
By making good use of spatial and frequency correlations, [15] Vutha Va, Junil Choi, and Robert W Heath. The impact of beamwidth
the Bi-LSTM SF-CNN achieved almost perfect MMSE on temporal channel variation in vehicular channels and its implications.
IEEE Transactions on Vehicular Technology, 66(6):5014–5029, 2016.
performance at a reduced complexity. Taking this a step
further, the Bi-LSTM SFT-CNN added temporal correlations
and continued to perform well even in the presence of