Professional Documents
Culture Documents
Abstract—Periodic measurement reporting based beam man- with increasing number of beams and/or users for beam-based
agement is not sufficiently agile for 5G New Radio (NR) and NR systems [4]. Without sufficient number of measurement
comes with significant overhead that scales with the number reports frequently being fed back by the UE, signal blocking
of beams and users. Furthermore, such an approach to beam
selection is unlikely to be sufficient to avoid signal blocking in may interrupt the service since the base station does not
real world scenarios. We propose a method to accurately predict know which beam of which TRP to switch transmission to
in advance the best serving beams and transmission points as as the signal begins to degrade due to blocking. Between
users move through the network and thereby eliminate the need two measurement reports, best serving beams could become
for frequent measurement reporting. Our prediction approach obsolete and cause a degradation in service quality, which may
applies deep learning techniques similar to that used in Natural
Language Processing (NLP) for translation/sentence completion not be remedied until a handover or beam switch following
tasks to the problem of predicting the best serving beams. The the next measurement report. Thus, even if a user is less likely
proposed solution enables the network to proactively switch users to switch beams, frequent measurement reports are still ex-
to new beams or cells to reduce blockage and handover related changed to monitor the best beam. Exchange of measurement
interruptions especially in high mobility scenarios. We evaluate reports cause both messaging overhead and wastage of the
our scheme in realistic scenarios using a new modeling technique
where computer vision is used to obtain mobility traces of users time-frequency resources, degrading user throughput.
from videos of live environments. We show significant benefits in Our solution to this problem is to enable the gNB to predict
terms of measurement report overhead reduction and signal-to- the next best serving beam for a UE and to proactively switch
noise ratio enhancement through blockage prevention in several beams without waiting for the next measurement report. Users
scenarios. in vehicles and pedestrians follow paths defined by roads and
walkways. They do not move randomly and typically follow
I. I NTRODUCTION
social customs, for example, they do not run into each other,
Next generation base stations (gNB) will be connected to they yield way to others, and often walk together [5]. In
one or more transmission reception points (TRP) that are computer vision, the task of trajectory prediction is viewed
capable of transmitting beamformed signals based on a pre- as a sequence generation task to predict future trajectories
determined set of directional beam patterns called the beamset. of objects based on their past positions. The authors of [5]
A TRP consists of a radio unit and an antenna panel capable propose a long term short term memory (LSTM) deep learning
of forming the beams in the beamset, and TRPs connected model, which can learn general human movement and predict
to a base station may be spatially distributed. Narrow beams their future trajectories. This level of social intelligence has
with significant beamforming gain are required to overcome been mostly ignored in wireless access systems, and the re-
path loss and extend the coverage especially in mm wave, source allocation algorithms are designed largely ignoring the
which implies a large number of beams in the beamset to spatio-temporal aspects of user distributions or their mobility
cover a whole cell serving area. As users move through the patterns. On the other hand, there is a significant amount
network, they frequently switch between beams, TRPs and of research on localization of individual users and use of
gNBs in order to receive the strongest possible signal [1], location and velocity information to provide better service
[2]. To facilitate beam switching, the base station has to [6]. In particular, in mmWave bands, because of blocking and
periodically transmit using each of the beams in the beamset reflections from structures even the UEs close to each other in
reference signals in the form of synchronization signaling terms of physical location may be served by different beams,
blocks (SSB) or channel state information reference signals TRPs or gNBs. Therefore, beam prediction should not rely
(CSI-RS) that a user equipment (UE) can use to measure entirely on explicit localization.
the received signal quality on the different beams and report The main contribution of this paper is a novel method for
the measurements to the base station for beam management. beam management that relies on prediction of best serving
Moreover, conveying information about the allocation of these beams using a LSTM network. Rather than relying on physical
resources requires extra UE-specific radio resource control location, channel state information or angular spread informa-
(RRC) and medium-access control (MAC) signaling [3]. The tion the prediction is performed by treating the sequences of
number of required time-frequency resources for reference beam indices as a sequence of letters or words in a language
signals and the amount of uplink resources required to send thereby enabling the application of NLP techniques. We also
measurement reports from the UE pose an overhead problem describe how in-field training can be performed to train the
Authorized licensed use limited to: Carleton University. Downloaded on June 04,2021 at 22:57:35 UTC from IEEE Xplore. Restrictions apply.
,(((:LUHOHVV&RPPXQLFDWLRQVDQG1HWZRUNLQJ&RQIHUHQFH :&1&
predictor. The second major contribution of this work is to Although their approach is beneficial for UE discovery, it
show how the prediction, even when is not highly accurate, would not work to mitigate blockages for specific UEs along
can be combined with measurement reports to operate the a path. The authors in [14] propose a deep learning approach
system efficiently by using the prediction reliability metrics for coordinated beamforming using deep learning. The deep
obtained from the predictor. Hence our proposed method can learning model learns to predict the beamforming vectors at
be applied to all scenarios independent of how predictable cooperating base stations based on the pilot sequences. Our
the mobility patterns are. We evaluate our scheme in realistic approach is applicable for non-cooperating sites as well.
scenarios using a new modeling technique where computer The paper is organized as follows. In Section II we describe
vision is used to obtain mobility traces of users from videos the beam management problem and provide an overview of our
of live environments and video frame by video frame ray prediction-based solution. In Section III the neural network
tracing is used to model the radio environment. We show architecture, training and inference procedures are described
significant benefits of our approach in terms of measurement in detail. Section IV showcases the usage and benefits of our
report overhead reduction and blockage prevention in multiple framework through simulations on two different scenarios. We
use cases. conclude with a summary in Section V.
We rely on a LSTM neural network approach to predict
the next best beams or sites to serve the users [5]. LSTM II. P REDICTIVE BEAM MANAGEMENT
is a recurrent neural network where the network is capable To enable beam management for a large number of beams,
of learning sequences of arbitrary length while limiting the 5G radio access networks support hierarchical beam refine-
error propagation using gates. Sequence-to-sequence learning ment. First Level is SSB beams limited up to 64 within a
is about training models to convert sequences from one cell. These are transmitted periodically and used for initial
domain to another using an encoder and decoder [7]. The access by UE to establish connection to the base station.
encoder-decoder architecture is referred as autoencoder. The At the second level, narrower refinement beams with SSB
encoder encodes and compresses sequence data step by step. beams as root beams are used to improve system performance.
The output of the encoder, which is also an input to the For the purposes of beam refinement and tracking, the gNB
decoder, represents the input sequence as a fixed length vector. can configure CSI-RS resources, where each CSI-RS may
The decoder interprets it for each step of the autoencoder. be sent on a different beam and direct the UE to provide
LSTM auto-encoder architecture could be regarded as a data relevant measurement reports. With multiple TRPs in a rapidly
compression algorithm with loss where the compression and changing environment such as in a city with cars and people
decompression functions are learned from examples rather moving, frequent reports from a large number of beams will
than being engineered with specifically designed algorithms. be required to overcome blocking.
The data driven encoding and decoding functions are designed During periodic reporting of the best beam index, each
using neural networks. UE sends at specified time intervals measurement reports that
There are many publications in the area of beam manage- include reference signal received power (RSRP) and identifiers
ment focusing on tracking beams. The reference [8] provides of the measured beams. As illustrated in Figure 1 the reported
an extensive overview of the recently proposed measurement best beam is used by the gNB until the next measurement
techniques for beam and mobility management in mmWave report. The gNB typically picks the strongest beam reported by
cellular networks and their impact on standards. The authors the UE as the serving beam. That limits the agility of the gNB
in [9] propose an angle domain hybrid precoding and channel to react to changes in signal quality especially, if a beam pair
tracking method for mmWave massive MIMO system where is rapidly blocked and there is no sufficient time for regular
they utilize an ekinematic equation characterize the movement beam adjustment. Each time the measured RSRP is below a
of users and track user’s direction of arrival. In our work, configured value a beam failure instance is declared by the UE.
we use realistic traces extracted from videos for creating test If the number of such instances exceeds a configured value,
scenarios. The authors in [10] employ an analog architecture, the UE declares beams failure and initiates the beam recovery
i.e. the base station uses a small number of radio frequency procedure to find a new beam pair where the connectivity
chains to drive a large number of antenna elements and forms could be restored [15]. However, this process will result in
the tracking beams in the analog domain. [11] uses machine service interruption or increased latency for packet delivery
learning to predict channel states and to derive beamforming and lowered throughput. By predicting the next beams, TRPs
vectors. Our approach is more practical since we assume grid or gNBs to serve a UE we can be avoid beam and radio link
of beams as widely used by the current beam-based wireless failures and save on time and frequency resources needed for
standards. The authors in [12] propose extracting the user recovery.
locations in high speed trains for beam tracking purposes. The essence of our work is to use an LSTM autoencoder
We do not require explicit user localization for predicting network to learn the features of the beam space, the environ-
next likely beams. Our approach could easily be applied to ment and the user mobility pattern. The only data the gNB
the high-speed train use case as well. The authors in [13] needs to accumulate is the serving beam index reports which
propose an algorithm to prioritize the sectors where more users are already available at gNB for each UE being served. The
are expected to be found according to previous experience. beams across gNBs need to be labeled with non-overlapping
Authorized licensed use limited to: Carleton University. Downloaded on June 04,2021 at 22:57:35 UTC from IEEE Xplore. Restrictions apply.
,(((:LUHOHVV&RPPXQLFDWLRQVDQG1HWZRUNLQJ&RQIHUHQFH :&1&
Authorized licensed use limited to: Carleton University. Downloaded on June 04,2021 at 22:57:35 UTC from IEEE Xplore. Restrictions apply.
,(((:LUHOHVV&RPPXQLFDWLRQVDQG1HWZRUNLQJ&RQIHUHQFH :&1&
Fig. 5. Training
are captured and stored in gNB memory along with their
Fig. 4. The LSTM autoencoder consists of two LSTM cells and a dense time stamps. During the training phase there may be some
layer. Each LSTM layer has 16896 parameters. The dense layer has 3267 non-optimal sequences of beams, e.g., the best beams may
parameters. It takes 2 ms to train a batch of 128 samples with CPU resources
(10 x I9-7900X 3.30GHz@4.30 GHz 128 GB Memory) be blocked soon after switching to that beam and radio link
failures may occur. So, training data can be constructed by
indices starting from n+1 conditioned on indices up to current using the sequence of beam indexes with good signal quality
index n. and suppressing intermediate beam indices with very poor
Once the training is completed the autoencoder is run in signal strength and short dwell times. Therefore, the gNB
may select an actual beam sequence that served the UE for a
inference mode where the beam indices used previously are
given period without any outage or construct a training beam
fed into the autoencoder and next set of beams are predicted.
In the inference mode, the input sequence of size P is encoded sequence from measurement reports ignoring beam switching
triggered with a gain below a threshold or followed by another
into the state vector output of the encoder. The state vectors
beam switch within a very short dwell time.
and the target sequence are fed to the decoder. The decoder
The beam sequences are interpolated to have an index for
predicts the probabilities of each beam to be the next. The
equidistant time intervals e.g every 10 ms which we refer as a
predicted beam sequence is appended to the target sequence.
slot. If the data in the memory does not include a beam index
This process is repeated until F future beams are predicted.
for each 10 ms sampling interval, then the previous beam may
The LSTM autoencoder which is used for predictions con-
be repeated, for example, because the same beam may be used
sists of two LSTM cells and a dense layer. Figure 4 shows the
for a period of time over multiple 10 ms intervals. Through
specific structure and parameters of the LSTM autoencoder.
such interpolation and processing of real beam sequences we
Each input beam index integer value is encoded as a binary
construct for each UE a beam sequence sampled at constant
vector that is all zero values except the beam index. This step
time intervals for training the predictor.
is called one-hot-encoding. The input sequence is the sequence
After a beam sequence of an arbitrary length is constructed,
of one-hot-encoded vectors which are fed into the input layer
we create samples of length P + F consisting of a first portion
labeled by input1 . Since we have N = 99 possible beams,
of P beam indices, and a second portion of future F beam
each one-hot-encoded vector is of size 99. The auto-encoder
indexes as the correct output for training. Consider a past
consists of two LSTM cells of size 32 which are denoted in
sequence of 600 ms. At 10 ms sampling rate, the past sequence
Figure 4 as lstm1 and lstm2 . The output of the decoder dense
would consist of P = 60 slots. If we want to predict 40 ms
layer is a vector where each entry is the probability of each
ahead, the future samples will be of size F = 4 slots. Figure 5
beam being the next determined using a SoftMax function. The
illustrates the neural network training steps. The past samples
index corresponding to the highest probability is deemed as the
collected from all UEs are fed in batches to the neural network.
most likely beam, the index of the second highest probability
The prediction of the neural network is compared with the
is deemed as the second most likely beam and so on.
correct future samples. The parameters of the neural network
B. Generating Beam Sequences are adjusted to minimize the error between predictions and
ground truth. We use categorical cross entropy function as
When gNBs are first deployed, they are operated in training defined in (1) where each unique beam index is treated as a
mode where beam refinement procedures are implemented category. In (1) F is the number of samples to be predicted, C
with traditional CSI-reporting configuration. During this time, is the number of unique beam index categories , 1yi∈Cc is the
each user device is asked to report substantial number of indicator function of the ith prediction belonging to the cth
measurement reports for estimating the quality of different category. pmodel [yi ∈ Cc ] is the probability predicted by the
beams. Training mode could also be triggered if substantial model for the ith observation to belong to the cth category.
changes are made to the antenna orientation or to the infras-
1
F C
tructure such as major new building or road construction in
− 1γ ∈C log Pmodel [γi ∈ Cc ] (1)
the neighborhood. F i=1 c=1 i c
In training mode, gNB configures UEs to measure specific,
likely beams or all available beams. These reports are used The process of tuning the neural network parameters based
directly to determine the best beam to serve the user during on a set of training samples may be repeated for many training
the training phase. The sequence of beams used by the UEs samples for many UEs, in order to train the beam sequence
Authorized licensed use limited to: Carleton University. Downloaded on June 04,2021 at 22:57:35 UTC from IEEE Xplore. Restrictions apply.
,(((:LUHOHVV&RPPXQLFDWLRQVDQG1HWZRUNLQJ&RQIHUHQFH :&1&
Fig. 6. Inference with autoencoder Fig. 7. People detected with computer vision algorithms on a single frame
of a video (left) and entire trajectories extracted from a video overlayed on
top of each other (right).
model to provide an accurate prediction. Training with the
entire user population’s past experiences improve prediction
accuracy for current individual users.
Authorized licensed use limited to: Carleton University. Downloaded on June 04,2021 at 22:57:35 UTC from IEEE Xplore. Restrictions apply.
,(((:LUHOHVV&RPPXQLFDWLRQVDQG1HWZRUNLQJ&RQIHUHQFH :&1&
Fig. 9. Top K Prediction Accuracy: Proportion of time slots in which one out Fig. 10. Top-2 Reliability: Percentage of all the samples where one of the
of the top K likely predicted beams for a slot includes the best beam/TRP to Top-2 beams is a correct prediction versus best beam prediction probability
serve the user in that slot. With Top-2 prediction, we can achieve 96% beam
prediction accuracy and 99.5% TRP prediction accuracy.
Authorized licensed use limited to: Carleton University. Downloaded on June 04,2021 at 22:57:35 UTC from IEEE Xplore. Restrictions apply.
,(((:LUHOHVV&RPPXQLFDWLRQVDQG1HWZRUNLQJ&RQIHUHQFH :&1&
Authorized licensed use limited to: Carleton University. Downloaded on June 04,2021 at 22:57:35 UTC from IEEE Xplore. Restrictions apply.