Deep autoencoder design for RF anomaly detection in 5G O-RAN near-RT RIC via xApps

Abstract

The agile and rapid management of the operations that take place in 5G radio access networks (RAN) including the monitoring of RF fluctuations, throughput suffering, and handover issues due to mobility on the user equipment (UE) side is becoming critical. Therefore, it needs to be managed by a near real-time RAN intelligent controller (RIC) in the context of O-RAN concept where O-RAN aims at the democratization of 5G RAN components to provide flexibility and compatibility to the vendors in the 5G market. Accordingly, in this paper, we present a deep learning (DL)-based autoencoder design for detecting the RF anomalies at the UE side through the extended applications (xApps) running on 5G near real-time RIC, thus, providing better and seamless service continuity. Simulation results demonstrate that the proposed autoencoder is able to achieve better performance on RF anomaly detection compared to the existing models such as random forest and isolation forest. Compared to the isolation forest algorithm, the deep-autoencoder model gives 10% better results in terms of overall accuracy score.

Full text

Deep Autoencoder Design for RF Anomaly
Detection in 5G O-RAN Near-RT RIC via xApps
Osman Tugay Bas¸aran, Mehmet Bas¸aran, Derya Turan, Hamide G¨
ul Bayrak, Ya˘
gmur Sabucu Sandal
Technology & Innovation (TI) Team, Research in Next-Generation Communications (RNC) - 5G Research Group
Kartal R&D Center, Siemens San. Tic. A.S., Istanbul, Turkey
{tugay.basaran, mehmet.basaran, derya.turan, hamide.bayrak, yagmur.sabucu}@siemens.com
Abstract—The agile and rapid management of the operations
that take place in 5G radio access networks (RAN) including
the monitoring of RF fluctuations, throughput suffering, and
handover issues due to mobility on the user equipment (UE)
side is becoming critical. Therefore, it needs to be managed by
a near real-time RAN intelligent controller (RIC) in the context
of O-RAN concept where O-RAN aims at the democratization
of 5G RAN components to provide flexibility and compatibility
to the vendors in the 5G market. Accordingly, in this paper,
we present a deep learning (DL)-based autoencoder design for
detecting the RF anomalies at the UE side through the extended
applications (xApps) running on 5G near real-time RIC, thus,
providing better and seamless service continuity. Simulation
results demonstrate that the proposed autoencoder is able to
achieve better performance on RF anomaly detection compared
to the existing models such as random forest and isolation forest.
Compared to the isolation forest algorithm, the deep-autoencoder
model gives 10% better results in terms of overall accuracy score.
I. INTRODUCTION
The rapidly growing cellular mobile telecommunication
network is not convenient for supporting industrial commu-
nication requirements such as low latency and high reliability.
Along with 5G, new technologies such as enhanced mobile
broadband (eMBB), massive machine-type communication
(mMTC), and ultra-reliable and low-latency communication
(URLLC) were introduced to satisfy the requirements includ-
ing the growing number of mobile devices [1]. The ubiquitous
connectivity for private networks has encouraged the develop-
ment of the 5G network. The 5G technology proposes a mod-
ular structure as much as possible for not only new radio (NR)
but also core network [2]. With this new structure, it promises
faster speeds, more reliability, and energy optimization than
ever before [3]. Although this modular structure gives more
flexibility in terms of using NR for different use cases, it still
lacks in terms of fine-grained control mechanisms. To fill this
gap, an open architecture is created under the name of Open
Radio Access Network (O-RAN).
O-RAN aims at the transformation of the traditional RAN
structure into a fully open and vendor-independent architecture
that permits the compliance of all of the network components
including radio unit (RU), distributed unit (DU), and central
unit (CU) based on the Split 7.2 option in order to reduce the
operational and capital expenses of service and mobile network
equipment providers [4]. In addition to the main components
of RAN, O-RAN also includes an additional software-defined
component called RAN intelligent controller (RIC) that en-
ables control of RAN operations more intelligently. RIC is
broken down into near real-time (Near-RT) and non-real-time
(Non-RT) RIC according to the data size and control loop
duration they handle. Near-RT RIC is close to CU and operates
on small-sized data within 10 ms to 1 s, while Non-RT RIC
operates on large-sized data whose processing time requires
more than 1 s.
According to the reference architecture of O-RAN, Near-
RT RIC deals with applications that include but are not
limited to the management of radio connection, mobility,
quality of service (QoS), and interference. However, it permits
users/operators to design specialized applications based on the
3rd party perspectives through extended applications called
xApps. Utilizing the designed xApps, it is also possible to
apply artificial intelligence/machine learning (AI/ML)-based
approaches into O-RAN for automating processes through
remote control [5].
O-RAN Alliance has an AI/ML working group dedicated
to standardizing AI/ML workflows for enhanced efficiency in
RAN operations [6]. AI/ML implementation on RAN opera-
tions attracts great attention in both academia and the industry
due to its huge potential.
II. RE LATE D WORKS
AI-based wireless network architecture and its implemen-
tation to the use cases focusing on user equipment (UE)
trajectory prediction and energy saving at the base station
(BS) are considered in [7] where only potential solutions
are suggested without demonstrating any key performance
indicators (KPIs) belonging to the aforementioned use cases.
In [8], architectural aspects and current status of disaggregated
O-RAN deployments are presented in detail by providing
AI/ML implementations and possible use case scenarios where
cell load prediction and energy efficiency issues are addressed
by long-short term memory (LSTM) and deep reinforcement
learning (DRL) techniques, respectively. In [9], the effect of
AI/ML usage on the control loop response time in RAN is
examined for resource adaptation where a drift-based solution
based on the retraining of the model or switching the best suit-
able trained model is proposed in order to avoid performance
degradation in prediction accuracy. In addition, RL is widely
applied to various problems such as intelligent user access
control schemes with the utilization of deep Q-network [10]

and the joint selection of better training model and resource
allocation for 5G network slicing services [11], and the total
cell throughput maximization by tuning parameters in the DU
side [12] etc.
O-RAN Near-RT RIC allows using xApps designed to serve
the dedicated tasks for network efficiency. In [13], a graph
neural network (GNN) dependent RL-based xApp is designed
as a combinatorial graph optimization problem for the man-
agement of user cell connectivity where the performance of the
DRL-based approach is superior to existing greedy solutions
and user cell association performance is demonstrated by three
different KPIs as sum user throughput, cell coverage, and
load balancing regarding the handover event. In [14], cell
throughput is predicted for intelligently managing the cellular
user mobility where neural network (NN) is designed by using
reference signal received power (RSRP) and reference signal
received quality (RSRQ) values as input nodes and download
success status as output node with 4 hidden nodes. It is shown
that the ML-based vector autoregressive approach is superior
to the traditional handover approach. In [15], a novel xApp
targeting quality of experience (QoE) enhancement is proposed
in the case of high-resolution video streaming services. In [16],
RIC fault-tolerant xApps are designed through techniques that
use state partitioning, partial replication, and fast re-route in
order to decrease the overhead.
O-RAN Near-RT RIC enables advanced closed-loop con-
trol applications for self-configuration, self-optimization, and
self-healing use cases of the self-organizing network (SON)
concept defined by 3GPP. Although there are still challenges
to achieving full autonomous RAN control and optimization
as outlined in [17], AI/ML solutions for intelligent closed-
loop control of open RAN are discussed and it is illustrated
how a throughput maximization solution is embedded in xApp.
In [18], an automated root cause analysis model including
anomaly detection, root cause analysis, measurement deploy-
ment, and system monitoring is proposed with a graph-based
unsupervised approach for cellular networks without a specific
scenario. However, two applications are proposed using a
link failure estimation model for handover optimization and
trajectory modification of unmanned aerial vehicles (UAVs)
[19]. In another study, [20], small latency deviations in RAN
are detected and localized by combining statistical, probabilis-
tic, and clustering models in a single learner over manually
obtained system logs. By combining three different models,
better results were obtained in terms of not classifying a
normal behavior as an anomaly compared to applying separate
models. There are generally two approaches to estimating
RSRP in 5G networks. One is to make an estimation based on
the empirical model defined in the standard and the other one
is based on the field measurements. The RSRP estimation is
presented using historical real data and statistical geographic
data over a two-tier NN model in [21].
As can be seen from current literature studies, AI/ML
studies focused on RAN have been practiced recently. This
developmental aspect of the 5G NR RAN also motivates our
research group to work in this direction. Robust AI/ML models
should be used to improve the performance of xApp/rApp
applications running on RIC [22]. Otherwise, a degradation in
overall network performance will be inevitable. For this rea-
son, we test the performance of the isolation forest (IF) algo-
rithm used in O-RAN anomaly detection xApp, suggesting that
a more sensitive and selective model for classifying anomalies
would be much more successful. To be more precise, the
AI/ML algorithm that will run on the Near-RT RIC only
provides healthy inference output through a well-pre-trained
model. For this reason, O-RAN has stipulated that the models
that can be deployed on the network must be pre-trained.
Therefore, the performance evaluation of pre-trained models
should be studied with precision. When necessary, the model
should be renewed and designed with better performance, and
accuracy should be adapted accordingly. In this respect, our
starting point is to design and use a state-of-the-art model that
covers these requirements underlined by O-RAN.
The main contributions of our work are as follows:
•To the best of our knowledge, we are the first to imple-
ment a deep autoencoder model with a semi-supervised
learning perspective on O-RAN anomaly detection xApp,
which can detect RF anomalies on the UE.
•With the Semi-Supervised Deep Autoencoder model,
much less labeled data is used. Thus, such an architecture
will provide a resource advantage in different scenarios
where high-dimensional datasets are used.
•Since the Deep Autoencoder model is dynamically de-
signed, it can be modified to work with data types specific
to different 5G use cases. In other words, the existing
model can be used on O-RAN compatible 5G networks
with vital performance requirements.
•In order to demonstrate the success of the designed
autoencoder model, we also implement the IF algorithm
used in O-RAN anomaly detection xApp and made a
performance comparison where we obtain a 10% better
overall accuracy score for the designed deep autoencoder
model.
The rest of the paper is organized as follows. Section III
explains O-RAN infrastructure and AI/ML design. Section IV
describes the deep autoencoder design for anomaly detection.
Section V provides the simulation results. Finally, Section VI
concludes this work.
III. O-RAN RIC INFRASTRUCTURE & AI/ML DESIGN
O-RAN RIC is a framework that provides intelligent control
capability via open interfaces on 5G architecture designed
according to 3GPP NR 7.2 Split [23]. Especially with the
help of Near-RT and Non-RT sub-controllers, orchestration is
performed with the close-loop control provided on the RAN.
In Fig. 1, it is represented that Near-RT RIC and Non-RT
RIC are connected to RAN components via open interfaces.
Interfaces such as O1 and E2 are vital for the creation of
datasets for AI/ML models used in xApps/rApps applications,
and for obtaining key performance values (throughput, delay,
power consumption, etc.) over RAN modules. On the other
hand, a specific standard flow has been determined by O-RAN

Fig. 1. Overall architecture of the O-RAN RIC.
for the AI/ML algorithms to be used to function correctly on
the 5G NR network [24]. According to these standards, the
design should be made by following the six-stage flow:
i) Data Collection and Processing
ii) Training
iii) Validation and Publishing
iv) Deployment
v) AI/ML Execution and Inference
vi) Continuous Operations
Each stage in the standard flow has its own unique functions
and requirements. Let’s consider a 5G NR network where
xApps are used to serve different use cases and entailments.
For instance, xApp A might be specialized for traffic forecast-
ing, while xApp B might be running on Anomaly Detection.
In this case, the Data Collection and Processing part in the first
stage will be different for xApp A and xApp B. Therefore, the
features of the dataset can be determined as throughput, RSRP,
RSRQ, channel quality information (CQI), latency, modulation
and coding scheme (MCS) depending on the use case. Since
this large feature pool will cause variation in the dimensions
of the space and parameters that the models will learn, the
data processing stage phase should be carefully determined.
Another critical point among these standards is that O-RAN
does not accept working with untrained models as it will
cause inaccurate outcomes for the 5G network stack. The
AI/ML model to be used must be validated through the offline
training phase. In this way, it is aimed to predetermine biased,
inefficient, and problematic models.
A. Near Real-Time RAN Intelligent Controller (RIC)
Connections of Near-RT RIC to CU and DU via E2 interface
ensure the effective execution of RAN orchestration task.
Near-RT RIC handles this orchestration task with its structure
consisting of many sub-components. Its main responsibili-
ties include internal messaging infrastructure, conflict miti-
gation, subscription manager, security sub-system, database,
and shared data layer. These components work together to
ensure that the microservices called xApp are included in the
Anomaly
Detection
Influx DB
Traffic Steering
Quality of
Experience (QoE)
Predictor
1. RMR
Anomaly(UEID)
2. RMR (Prediction
Request(UEID))
3. RMR (Prediction
Response(UEID,cell1,tp1,
cell2,tp2))
E2 Sim
4. REST (Handover (UEID,cell1) )
KPIMON E2 – KPM 2.0M
Fig. 2. Representation of O-RAN anomaly detection xApp use case.
operation correctly, synchronously, and without causing any
conflict. Some of the sub-components have already been de-
veloped and functional, while others continue to be developed
and standardized by the O-RAN Working Groups. Such an
intense and modular structure has been preferred in order to
accurately transport the responses to and from xApp via the
A1, E2 and O1 interfaces.
According to O-RAN, xApp is an independent software
plug-in used on the Near-RT RIC framework. On the other
hand, it is a software image consisting of a descriptor. By using
this microservice infrastructure, many different applications
are enabled to run on the network. In particular, it is aimed to
work with AI/ML Workflow. The sub-components included
in Near-RT RIC pave the way for xApps to be properly
designed and deployed to the network. The deployment that
is meant here can also be achieved from the point of view of
dockerization perspective used in Kubernetes infrastructure.
B. Anomaly Detection xApp
Anomaly characterizes the abnormal situation that has a
direct impact on performance, and it is being studied in many
different research areas so that it can be caught beforehand.
Wide-range algorithms have been developed to detect and
track anomalies in many areas such as finance, biomedical
systems, and data science.
Anomaly detection studies on communication systems
(5G/6G, wireless technologies) provide added value in terms
of network monitoring and security. Especially the capabilities
of AI/ML models used in innovative communication technolo-
gies make it easy to identify anomalies on 5G RAN. In this
sense, O-RAN has completed and released the xApp design
that will determine the anomalies on the UE side. Fig. 2
shows the use case scenario where Anomaly Detection xApp is
used. Each component in the use case is designed by different
developers of O-RAN Society. Besides Anomaly Detection
xApp, Traffic Steering (TS) and Quality-of-Experience (QoE)
Predictor (QP) xApps are designed, as well. The realization
scenario of the use case is to send the abnormal UEs de-
termined by Anomaly Detection xApp to TS xApp with a
RIC message router (RMR) message, where it is aimed to
eliminate this anomaly with QP xApp. The QP will classify
the problem caused by the throughput on the abnormal UE

Fig. 3. Visualization of the dataset via t-SNE
and take action accordingly to resolve it. In the first place, the
correct operation of the anomaly detection algorithm is the
central point of the realization of this whole flow. Therefore,
we start by testing the IF algorithm released by O-RAN for
Anomaly Detection xApp (Release 0.0.2, 2021).
1) Data Description: The IF algorithm running in Anomaly
Detection xApp has been trained with the dataset obtained
from the UEs. The dataset was collected from different types
of UE sources such as car, train passengers, pedestrians, and
waiting passengers. The dataset consists of 10000 samples
belonging to valuable features such as RSRP, RSRQ, RSSINR,
Physical Resource Block (PRB) usage, and Throughput. It also
contains object class features such as timestamp, NRCellIden-
tity, DU-id, and UE-id. Finally, it contains binary-classified (0
or 1) label information indicating whether there is an anomaly
or not. The dataset contains approximately 25% samples of
the anomaly class, suggesting that we are working with an
unbalanced dataset. In Fig. 3, classes of the dataset is visu-
alized using the t-Distributed Stochastic Neighbor Embedding
(t-SNE) nonlinear statistical method [25]. Since our dataset is
high-dimensional, it has been mapped into two-dimensional
using a data visualization technique. According to Fig. 3, it is
very clear how the samples of the anomaly class are sparsely
and haphazardly dispersed in space. Therefore, it is useful to
use a strong model to classify this minority class.
2) Isolation Forest Algorithm: Random Forest Classifier
was first used in Anomaly Detection xApp Release 0.0.1
(2020) version. Later, IF algorithm, which is an unsupervised
learning technique, was released. The algorithm is in a design
that produces tree-based solutions as can be seen from Fig.
4. Since it is an unsupervised learning technique, it would be
much more appropriate to use it in datasets where labels are
not available. It works to detect distinct and rare data samples
in the feature space. O-RAN developed this algorithm using
the scikit-learn library [26]. The algorithm flow is as follows:
i) Randomly select the features in the dataset and initiate
the split operation for each feature.
ii) Each data sample will have a root node and a leaf node,
calculate this average distance.
iii) For the tree representation, apply steps 1 and 2 to each
tree.
Fig. 4. Representation of isolation forest algorithm.
iv) Find the average distance between all trees.
v) Since the average distance between normal data samples
is higher compared to anomalies, anomalies are classified.
Finally, the IF algorithm consists of important parameters
such as n estimators,max samples and contamination. In
their implementation, O-RAN used RandomizedSearchCV
hyperparameter tuning estimator of scikit-learn library for
the parameters including contamination. The contamination
parameter is the rate of anomalies in the dataset. Normally,
while the IF algorithm will take the contamination parameter
statically, three different candidate contamination values are
tried with this hyperparameter tuning step.
IV. AUTO EN CO DER DESI GN F OR AN OM ALY DETE CT IO N
In 1986, the first version of the autoencoder studies was
published [27]. When autoencoder models (which are basically
unsupervised learning techniques) are designed with a semi-
supervised learning perspective, very prospering results are
obtained recently. When the O-RAN dataset is evaluated, we
decide that it will be beneficial to work with such a modernist
and state-of-the-art model.
A. Model Design & Hyperparameter Tuning
While designing the model, it is realized that an autoencoder
will be used with a semi-supervised learning approach. In this
context, a very low percentage of samples belonging to the
non-anomaly class (n were fed to the autoencoder, allowing the
model to learn the hidden representation of the non-anomaly
class. At this stage, while the non-anomaly class is learned by
the model, the learning performance of the model is tested by
controlling the training and validation loss per epoch. By using
the obtained hidden representation layer weights, the samples
belonging to the anomaly class are classified by the model.
The working mechanism of the designed autoencoder model
is as follows:
i) Create autoencoder network.
ii) Standardize and transform feature values.
iii) Show a small amount of non-anomaly class samples to
autoencoder.
iv) Create latent representation layers of non-anomaly class
samples.
v) Predict anomaly class samples by using Step (iii).
vi) Train and validate the final classifier after the hyperpa-
rameter tuning stage.

Encoder
Latent
Representation
Input
Decoder
Output
Input
Fig. 5. Representation of the designed autoencoder model.
Considering the unbalanced dataset, we decide to proceed
with the advanced deep autoencoder model during model
design. However, experiments are started by using a simple
deep neural network (DNN) model and then designing a
low-layer autoencoder model. Thus, the advantage of the
deep autoencoder model has been expressed more clearly.
When we start with DNN experiments, it is seen that the
predictive performance of the anomaly class drops drastically.
Even if the varying number of layers and neuron units,
different optimizers, and activation functions are used during
the experiments, there is no improvement in this performance.
In addition, since the DNN model works from a supervised
perspective, the advantage of using less labeled data provided
by the autoencoder model is no longer significant. In the
second stage, the simple autoencoder model is designed and
the performance of the model is examined. When MSE and
reconstruction loss are examined, it is observed that the model
cannot learn effectively.
Therefore, it is decided to design a more powerful deep
autoencoder model. In this direction, we design the model with
a structure consisting of shallow layers to be symmetrical in
the encoder and decoder parts. Experiments are carried out
while designing the bottleneck part of the autoencoder that
connects the encoder and decoder which has a direct impact
on the learning capacity of the model. In the experiments
on the bottleneck size, it is seen that the model copies
the input instead of reconstructing it by overfitting when it
is kept larger than the number of features. In such cases
where the bottleneck size is too small, it is observed that
the model cannot learn enough to reconstruct the input and
the mean squared error (MSE) increased. MSE is calculated
between the real input and the reconstructed input. Batch
normalization layers are used on both the encoder and decoder
sides to prevent the model (with 31297 parameters) from
overfitting and to keep the convergence time at a reasonable
period. While determining the optimizer, experiments are
TABLE I
DEE P AUTO EN COD ER MO DE L LAYER S AND PA RAM ET ER S
Autoencoder Layers Output Shape Parameters
input 1 (InputLayer) (None, 20) 0
dense (Dense) (None, 100) 2100
batch normalization (BatchNorm) (None, 100) 400
dense 1 (Dense) (None, 75) 7575
batch normalization 1 (BatchNorm) (None, 75) 300
dense 2 (Dense) (None, 50) 3800
batch normalization 2 (BatchNorm) (None, 50) 200
dense 3 (Dense) (None, 25) 1275
batch normalization 3 (BatchNorm) (None, 25) 100
dense 4 (Dense) (None, 3) 78
dense 5 (Dense) (None, 3) 12
batch normalization 4 (BatchNorm) (None, 3) 12
dense 6 (Dense) (None, 25) 100
batch normalization 5 (BatchNorm) (None, 25) 100
dense 7 (Dense) (None, 50) 1300
batch normalization 6 (BatchNorm) (None, 25) 200
dense 8 (Dense) (None, 75) 3825
batch normalization 7 (BatchNorm) (None, 75) 300
dense 9 (Dense) (None, 100) 7600
dense 10 (Dense) (None, 20) 2020
Total Params: 31,297
Trainable Params: 30,491
Nontrainable Params: 806
performed with RMSprop, Adam and Stochastic Gradient
Descent (SGD) solvers. Since the network weights are updated
with an iterative perspective at every stage with Adam, more
successful results have been obtained compared to RMSprop
and especially SGD. The activation functions used in the
encoder and decoder layers are designed in a hybrid manner
by considering performance. therefore, it is decided to use tanh
and ReLU activation functions. Considering the convergence
time, different batch size values are tried and the experiments
are completed by gradually increasing them. As a result, the
batch size is determined as 256. After the design of the model
is completed, hyperparameter tuning is performed using a wide
parameter space to find the most optimized parameters for the
Autoencoder. Table I and Figure 5 show the final design of the
autoencoder network. The model and hyperparameter tuning
stages are designed with the Keras Deep Learning Library and
Google Collaboratory Notebook.
V. SIMULATION RES ULTS
Starting with Release 0.0.1 Random Forest algorithm, O-
RAN Anomaly Detection xApp is designed with IF algo-
rithm in Release 0.0.2. However, when O-RAN develops
this algorithm, it only evaluates the performance with the
average f1-score in the hyperparameter tuning phase. In fact,
the remaining Precision, Recall, and Accuracy performance
metrics should be checked separately within both classes
(anomaly/non-anomaly). Only in this way, the performance of
the algorithm is clearly revealed and then it can be compared
with the Deep-Autoencoder model that we built.
A. O-RAN Isolation Forest Classifier
The performance evaluation of O-RAN Anomaly Detection
xApp using the IF classifier should be done carefully. O-RAN

TABLE II
IF ALGORITHM ACCU RAC Y RES ULTS F OR VARIABLES n estimators AND
contamination PARAMETERS
contamination n estimators
200 400 600
0.05 75.80 75.80 75.96
0.10 79.08 79.08 79.20
0.15 75.78 84.60 84.56
0.20 84.60 85.40 87.04
0.25 85.76 87.28 86.04
0.28 83.48 83.40 83.56
0.30 81.96 81.76 81.71
TABLE III
CLASSIFICATION PER FOR MA NC E OF O-RAN ISO LATI ON FO RES T
ALGORITHM (CONTAMINATION=0.28)
Class Precision Recall f-Measure
Non-Anomaly 0.91 0.87 0.89
Anomaly 0.66 0.74 0.70
Accuracy 0.83
performs hyperparameter tuning using the labels (available
in the dataset), specific to f1-score and accuracy. However,
proceeding by evaluating f1-score, accuracy, precision and
recall performance metrics for both classes gives more ac-
curate results. That’s why, we evaluate the pretrained IF
algorithm in terms of f1-score, recall, precision and accuracy
for both normal and anomaly classes. As can be seen in
Table II, the experiments are completed by choosing the
two parameters that affect the performance the most. The
n estimators parameter refers to the trees in the forest of the
IF algorithm. It is a very valuable parameter for the model to
learn. When the n estimators parameter, which has a default
value of 100, is increased from 200 to 1000 as O-RAN does,
it has been observed that the learning curve decreases. For
this reason, performance results of up to 600 are examined.
As mentioned in Section III, the contamination parameter
specifies the percentage of anomalies in the dataset. When
this is examined, RandomizedSearchCV parameter estimator
makes the most successful choice with 28% among possible
values. However, this value is a biased result. Especially, as
seen in Table II, the best accuracy value is obtained when the
contamination parameter is 25% with our simulation results.
In fact, this result is also compatible with the class distribution
of the actual dataset.
The performance results of the contamination parameter de-
termined by RandomizedSearchCV are shared in Table III. The
imbalanced data shows that the IF algorithm has difficulties
in detecting the anomaly class. Particularly noteworthy is the
precision performance is a little over 50% in the anomaly class.
Such classification performance in communication systems
can lead to serious performance degradation in the network.
Much more capable pre-trained models should be used.
In Table IV, the performance results of IF algorithm with
the contamination parameter determined according to the
percentage of the actual anomaly of the dataset are shown.
Although a better performance is seen according to the O-
TABLE IV
CLASSIFICATION PERFORMANCE OF O-RAN ISOL ATIO N FOR ES T
ALGORITHM (CONTAMINATION=0.25)
Class Precision Recall f-Measure
Non-Anomaly 0.80 0.95 0.92
Anomaly 0.82 0.77 0.75
Accuracy 0.89
TABLE V
CLASSIFICATION PERFORMANCE OF THE PROP OSE D DEE P
AUTO EN CO DER MO DE L
Class Precision Recall f-Measure
Non-Anomaly 0.93 0.93 0.93
Anomaly 0.91 0.92 0.91
Accuracy 0.93
RAN implementation results, it is still open to development.
B. The Proposed Deep-Autoencoder Model
By starting the experiments with the simple autoencoder
model, the structure of the autoencoder has been made more
complex and advanced. As the stacked layers are added, the
learning ability of the autoencoder and the related performance
results improve. While the overall accuracy score obtained
with the simple autoencoder model is 85%, the accuracy score
of the deep-autoencoder model after hyperparameter tuning is
93%, as seen in Table V. A momentous result that should
be considered while obtaining the final accuracy score is
that Precision, Recall and f1-score performance results are
quite high for the minority anomaly class compared to the
IF algorithm. The Deep Autoencoder model is validated bi-
directionally during the experiments. MSE loss is checked
during the training/test against the possibility of overfitting
or underfitting the model.
VI. CONCLUSION
In this work, we have designed a deep autoencoder for
RF anomalies at the UE side in order to handover UE to a
neighboring cell and thus, to manage the network traffic seam-
lessly through the O-RAN Near-RT RIC applications that run
on RAN side. We have visualized the dataset first. Then, we
applied the existing IF algorithm for detecting anomalies. Due
to the insufficient accuracy rates of IF, we have proposed to use
a deep autoencoder in order to better track the RF anomalies
that UE experiences. We have demonstrated the improved
performance of the designed deep autoencoder compared to
IF algorithm for efficient TS. In particular, although the IF
algorithm is an unsupervised learning technique, performance
measurement has been made using the labels in the dataset in
the O-RAN implementation. In this case, the performance of
the algorithm is provided with predefined labels. Designing a
deep autoencoder model with a more alternative and innovative
semi-supervised learning perspective has enabled us to achieve
better results in many respects. In this way, a deep learning
model that complies with O-RAN’s 5G Network AI/ML
Workflow requirements has been obtained. First of all, the
performance values are high in both classes. Especially, the

low classification performance in the minority class seen in the
IF algorithm is not seen in the deep autoencoder model. On the
other hand, the deep autoencoder model designed with a semi-
supervised learning technique is very suitable for working with
larger and unlabeled datasets.
REFERENCES
[1] N. Panwar, S. Sharma, and A. K. Singh, “A survey on 5G: The next
generation of mobile communication,” Physical Commun., vol. 18, pp.
64–84, 2016.
[2] S.-Y. Lien, D.-J. Deng, and B.-C. Chang, “Session management for
URLLC in 5G open radio access network: A machine learning ap-
proach,” in Int. Wireless Commun. and Mobile Comput. (IWCMC), 2021,
pp. 2050–2055.
[3] A. Thantharate, A. V. Tondwalkar, C. Beard, and A. Kwasinski,
“ECO6G: Energy and cost analysis for network slicing deployment in
beyond 5G networks,” Sensors, vol. 22, no. 22, 2022.
[4] O-RAN Alliance, “O-RAN: Towards and open and smart RAN,” O-RAN
Alliance Whitepaper, Oct. 2018.
[5] ——, “O-RAN use cases and deployment scenarios,” O-RAN Alliance
Whitepaper, Feb. 2020.
[6] O-RAN Working Group 2, “O-RAN AI/ML Workflow Description and
Requirements - v1.01,” O-RAN Alliance Technical Specification, 2021.
[7] H. Zhang, J. Xin, S. Xu, and S. Xiong, “Artificial intelligence based
architecture and implementation of wireless network,” in 2nd Int. Conf.
Electron., Commun. and Inf. Technol. (CECIT), 2021, pp. 273–278.
[8] A. Giannopoulos et al., “Supporting intelligence in disaggregated open
radio access networks: Architectural principles, AI/ML workflow, and
use cases,” IEEE Access, vol. 10, pp. 39580–39 595, 2022.
[9] V. R. Chintapalli et al., “WIP: Impact of AI/ML model adaptation on
RAN control loop response time,” in IEEE 23rd Int. Symp. World of
Wireless, Mobile and Multimedia Networks (WoWMoM), 2022, pp. 181–
184.
[10] Y. Cao, S.-Y. Lien, Y.-C. Liang, and K.-C. Chen, “Federated deep
reinforcement learning for user access control in open radio access
networks,” in IEEE Int. Conf. Commun. (ICC), 2021, pp. 1–6.
[11] A. K. Singh and K. Khoa Nguyen, “Joint selection of local trainers
and resource allocation for federated learning in open RAN intelligent
controllers,” in IEEE Wireless Commun. and Netw. Conf. (WCNC), 2022,
pp. 1874–1879.
[12] H. Lee, Y. Jang, J. Song, and H. Yeon, “O-RAN AI/ML workflow
implementation of personalized network optimization via reinforcement
learning,” in IEEE Globecom Workshops (GC Wkshps), 2021, pp. 1–6.
[13] O. Orhan, V. N. Swamy, T. Tetzlaff, M. Nassar, H. Nikopour, and
S. Talwar, “Connection management xAPP for O-RAN RIC: A graph
neural network and reinforcement learning approach,” in 20th IEEE Int.
Conf. Mach. Learn. and Appl. (ICMLA), 2021, pp. 936–941.
[14] B. Haryo Prananto, Iskandar, and A. Kurniawan, “O-RAN intelligent
application for cellular mobility management,” in Int. Conf. ICT for
Smart Society (ICISS), 2022, pp. 01–06.
[15] B. Agarwal, M. A. Togou, M. Ruffini, and G.-M. Muntean, “QoE-driven
optimization in 5G O-RAN enabled HetNets for enhanced video service
quality,” IEEE Commun. Mag., pp. 1–7, 2022.
[16] A. Huff, M. Hiltunen, and E. P. Duarte, “RFT: Scalable and fault-tolerant
microservices for the O-RAN control plane,” in IFIP/IEEE Int. Symp.
Integrated Netw. Manage. (IM), 2021, pp. 402–409.
[17] L. Bonati, M. Polese, S. D’Oro, S. Basagni, and T. Melodia, “Intelligent
Closed-loop RAN Control with xApps in OpenRAN Gym,” in Proc.
European Wireless, Dresden, Germany, Sep. 2022.
[18] D. Canastro, R. Rocha, M. Antunes, D. Gomes, and R. L. Aguiar, “Root
cause analysis in 5G/6G networks,” in 8th Int. Conf. Future Internet of
Things and Cloud (FiCloud), 2021, pp. 217–224.
[19] K. Boutiba, M. Bagaa, and A. Ksentini, “Radio link failure prediction
in 5G networks,” in IEEE Global Commun. Conf. (GLOBECOM), 2021,
pp. 1–6.
[20] T. Sundqvist, M. Bhuyan, and E. Elmroth, “Uncovering latency anoma-
lies in 5G RAN - A combination learner approach,” in 14th Int. Conf.
Commun. Syst. and Netw. (COMSNETS), 2022, pp. 621–629.
[21] P. Li, X. Wang, R. Piechocki, S. Kapoor, A. Doufexi, and A. Parekh,
“Variational autoencoder assisted neural network likelihood RSRP pre-
diction model,” in IEEE 33rd Ann. Int. Symp. Pers., Indoor and Mobile
Radio Commun. (PIMRC), 2022, pp. 554–559.
[22] T. Karamplias, S. T. Spantideas, A. E. Giannopoulos, P. Gkonis, N. Kap-
salis, and P. Trakadas, “Towards closed-loop automation in 5G open
RAN: Coupling an open-source simulator with xApps,” in Joint Eur.
Conf. Netw. and Commun. & 6G Summit (EuCNC/6G Summit), 2022,
pp. 232–237.
[23] 3rd Generation Partnership Project (3GPP), “Study on new radio access
technology: Radio access architecture and interfaces,” 2017. [Online].
Available: http://www.3gpp.org/DynaReport/38801.html
[24] O-RAN Working Group 2, “O-RAN AI/ML workflow description and
requirements 1.03,” O-RAN.WG2.AIML-v01.03 Technical Specification,
Jul. 2021.
[25] L. Van Der Maaten and G. Hinton, “Visualizing data using T-SNE,” J.
Mach. Learn. Res., vol. 9, pp. 2579–2605, 2008.
[26] F. Pedregosa et al., “Scikit-learn: Machine learning in Python,” J. Mach.
Learn. Res., vol. 12, pp. 2825–2830, 2011.
[27] D. E. Rumelhart and J. L. McClelland, Learning Internal Representa-
tions by Error Propagation. MIT PRESS, 1987.