Received 31 January 2023, accepted 14 February 2023, date of publication 22 February 2023, date of current version 27 February 2023.

Digital Object Identifier 10.1109/ACCESS.2023.3247519

Wireless Technology Identification Employing

Dynamic Mode Decomposition Modeling
AHMED ELSEBAAY AND HAZEM H. REFAI , (Member, IEEE)
Department of Electrical and Computer Engineering, The University of Oklahoma, Tulsa, OK 74135, USA
Corresponding author: Ahmed Elsebaay ([Link]-1@[Link])

ABSTRACT Significant growth in broadband wireless services, as well as ever-increasing demand on the
spectrum caused by the Internet of Things (IoT) have overstretched limited available spectrum space for
wireless services. Heterogeneous wireless networks (HetNets)—wherein multiple wireless technologies
(e.g., Wi-Fi, Bluetooth, Zigbee, LTE, and GSM) coexist and share spectrum—are a promising solution
for enhancing spectrum sharing. An essential element in developing coexistence protocols is correctly
identifying wireless technologies anticipated to share spectrum and to shift users between available wireless
technologies in an effort to optimize spectrum usage and minimize interference. For the coexistence research
reported in this paper, we analyzed the performance of our developed novel algorithm based on dynamic
mode decomposition (DMD) mathematical modeling to identify and differentiate among various wireless
technologies. More specifically, our technique identified GSM and LTE signals in the cellular domain,
IEEE802.11n, ac, and ax in the Wi-Fi domain, as well as Bluetooth and Zigbee. The proposed DMD-based
technique identifies the time domain signature of a signal by capturing embedded periodic features trans-
mitted within the signal. Performance and accuracy were tested and validated using an experimental dataset
collected for various time series, and raw-power measurements of the targeted technologies. Results showed
that the developed DMD-based algorithm can differentiate and classify individual and coexisting wireless
signals with high accuracy —greater than 90% for most cases. Furthermore, only a short time— less than one
second—is required for identifying a signal and enabling implementation in real-time practical networks.
The advantage of the developed technique over comparable techniques is lower complexity (i.e., shorter
processing and training time, no channel estimation, no time/frequency synchronization, and no need for
long observation-time intervals).

INDEX TERMS Dynamic mode decomposition, wireless coexistence, wireless identification.

I. INTRODUCTION congestion, which, in turn, has caused various levels of inter-

Escalating traffic demands coincide with significant growth ference among coexisting technologies.
in broadband wireless services. Vast connectivity and the Heterogeneous wireless networks (HetNets) employing
Internet of Things (IoT) are causes for a consistently over- smart cognitive radio devices have been introduced as
stretched spectrum space [1], [2], thus, increasing spectrum an effective solution to enhance network capacity, data
scarcity. Constrained licensed spectrum resources are sim- rate, coverage, and spectrum resource utilization. HetNets
ply unable to meet the ever-rising demand [1], [3]. Scarcity are composed of several coexisting wireless technologies
and high cost of licensed spectrum cause wireless technolo- sharing wireless spectrum [1], [6], [7]. Ensuring effec-
gies (e.g., Wi-Fi, ZigBee, and Bluetooth) to share spectrum tive coexistence across suitable wireless technologies is
in unlicensed bands [4]. The ISM band is one such unli- imperative.
censed and unrestricted band, and many technologies try to Wireless technologies must be identified within a fre-
coexist in it [5]. This trend has led to overutilization and quency range in order for intelligent cognitive radio devices
to analyze spectrum occupancy, identify available channels,
The associate editor coordinating the review of this manuscript and and model interference while attempting to coexist. Only then
approving it for publication was Hosam El-Ocla . can communication be effective and successful [5].

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see [Link]
VOLUME 11, 2023 18427
 A. Elsebaay, H. H. Refai: Wireless Technology Identification Employing Dynamic Mode Decomposition Modeling

A. RELATED WORK 3) CDF-BASED ALGORITHMS

Wireless signals are identified as likelihood-based (LB) and CDF-based algorithms mitigate the primary limiting issues
feature-based (FB). Conventional LB methods are based on in wavelet-based and cyclostionarity-based algorithms by
calculating maximum average probability for proper iden- processing signals that require a shorter observation interval
tification. Recent FB schemes rely on capturing common and can tolerate signal SNR variation. Researchers in [15]
features shared among similar signal types. Comparing LB introduced a novel identification method for distinguishing
methods with FB methods shows that although the latter has GSM signals from LTE signals. Amplitude of the observed
suboptimal performance, FB has a simpler implementation, signal samples was employed to extract technology-based
lower computational complexity, and relative robustness for features using statistics and signal structures obtained from
modeling mismatches among various operation cases [1], [8]. time and frequency domains. A Kolmogorov-Smirnov (K-S)
FB algorithms are based on wavelets, cumulative distribution test was then used to develop a decision principle. The same
functions (CDF), second-order cyclostationarity, machine technique was used by [16] to identify LTE, UMTS, GSM,
learning (ML), and deep learning (DL). Each technique is and CDMA2000 networks.
briefly described in the following sections and ordered from
oldest to most recent schemes. 4) DL-BASED ALGORITHMS
DL-based has been recently introduced as a useful method for
classifying wireless signals. Convolutional neural networks
1) WAVELET-BASED ALGORITHMS
(CNN) are the most popular DL architectures exploiting both
References [9] and [10] introduced a wavelet-based algo- modulation and wireless technology recognition [2]. High-
rithm for identifying GSM and UMTS signals relying on lighting the most recent work in the DL area, researchers
differences in their respective modulation schemes. The algo- in [17] and [18] chiefly constructed classification models
rithm applied wavelet transform to extract transient behaviors based on CNN to process time series signals for GSM,
within signals resulting from modulation types, and then UMTS, and LTE. Authors in [18] enhanced the trained model
utilizes template matching in the wavelet transform domain by utilizing both image and vector representations of the sig-
for identification. nals. The model achieved high accuracy for classifying sig-
nals, including GSM and LTE, with additive white Gaussian
2) SECOND-ORDER CYCLOSTATIONARITY-BASED noise (AWGN) and Rayleigh fading channels used to identify
ALGORITHMS UMTS, LTE, and 5G signals simulated by MATLAB LTE
Reference [11] utilized a second-order cyclostationarity- and 5G toolboxes. Reference [19] proposed a DL-based intel-
based algorithm to detect and identify cyclic patterns of GSM, ligent recognition method for identifying unlicensed band
wideband code division multiple access (WCDMA), and LTE and Wi-Fi signals generated in a laboratory environ-
orthogonal frequency division multiplexing (OFDM) mod- ment. CNN and recurrent neural network (RNN) models were
ulated signals. Cyclostationarity generated by cyclic prefix, trained using in-phase and quadrature (IQ) signals.
preamble, and pilot signals were exploited by [1] to classify Reference [20] used a CNN-based approach for classifying
GSM, LTE, and CDMA signals utilizing fast Fourier trans- 802.11 b/g, 802.15.4, and 802.15.1 signals that coexist in
form (FFT), autocorrelation function (ACF), power spectral 2.4GHz unlicensed bands. For SNR greater than or equal
density (PSD), and spectral correlation function (SCF) as to 5 dB, the model’s accuracy was greater than 95%. Authors
features for support vector machines (SVM). Reference [12] in [21] classified LTE, WiFi, and DVB-T technologies that
successfully employed a cyclostationarity-based technique share the same ISM band using Received signal strength
to classify LTE and GSM signals based on their perspec- indication (RSSI), IQ samples, and spectrogram features.
tive pilot signals. Researchers also investigated the effect of A CNN model was employed in [5] to distinguish individual
signal-to-noise ratio (SNR) and observation time on identi- and coexisting 802.11n, Bluetooth, and Zigbee signals using
fication accuracy. A classifier model developed by [7] was frequency domain features. The model was trained using
used for recognizing eight wireless standard signals, namely power-frequency measurements taken from the IQ compo-
WCDMA, LTE, GSM, CDMA, Digital Enhanced Cordless nents recorded at various SNR levels. Researchers com-
Telecommunications (DECT), WLAN, Bluetooth, and Dig- pared different ML models’ categorization accuracy. Results
ital Video Broadcasting (DVB). The technique is based on revealed that for signals with an SNR higher than 15dB, CNN
transforming second-order cyclostationarity to SCF, and then had the highest classification accuracy— greater than 90%.
utilizing it as an input to SVM for classification. Convolutional denoising autoencoders (CDAEs) were used
By utilizing hidden Markov Models to find second-order in a similar study published in [22] to recover distorted spec-
cyclic OFDM features, 802.11 a/g signals were cate- trograms before categorizing signals with CNN. The model
gorized in [13]. The authors used a software-defined achieved 91% accuracy when identifying IEEE 802.11a,
radio in [14] to present a combined energy detection IEEE 802.11n, IEEE 802.11ax, IEEE 802.11ac, and unli-
and cyclostationarity-based technique for detecting IEEE censed LTE signals. A WaveNet model-adapted neural net-
802.11g and IEEE 802.15.4 signals in the 2.4 GHz range. The work was created by authors in [23] to distinguish between
highest accuracy was 90% for signals with 1.6 dB SNR. 802.11n, 802.11ac, and 802.11ax Wi-Fi signals. Raw power

18428 VOLUME 11, 2023

A. Elsebaay, H. H. Refai: Wireless Technology Identification Employing Dynamic Mode Decomposition Modeling

time series data were gathered for both standalone and coex- and coexisting IEEE 802.11b/g/n, 802.15.4, 802.15.1, and
isting signals at different throughputs. The approach had a Bluetooth Low Energy (BLE) technologies were categorized
high classification accuracy between 90 and 98%. Ten wire- by authors in [31] for the 2.4 GHz ISM band. They used a
less technologies, including Wi-Fi, ZigBee, and Bluetooth, variety of ML algorithms, including decision trees (DT), RF,
were classified in the 2.4GHz ISM band in [24]. Experimen- and SVM, demonstrating 90–97% identification accuracy for
tally, raw IQ samples were collected in an indoor lab. A DL signals with SNR greater than 0dB.
multi-task neural architecture was created by researchers The FB approaches detailed above are based on extracting
for identifying signals by their modulation properties. specific features from a certain signal, and then identifying
Reference [25] classified separate and coexisting Bluetooth, the signal using a classification model. The classification
Wi-Fi, and microwave signals using a pre-trained, Incep- decision is performed by analyzing the probability distri-
tionV3, CNN-based model. ComBlock’s commercial, off- bution function (PDF) of the feature vectors or minimiz-
the-shelf modules generated signals in the 2.4GHz ISM ing the error between the calculated and estimated values.
band. The model’s overall accuracy was 98% for 800 testing These approaches have the advantages of being a) simple
samples. to implement and b) proven to provide near-optimal per-
Researchers in [26] utilized CNN to identify various wire- formance. However, they are sensitive to noise level and/or
less signals based on their modulation. Constellation dia- might require prior information about targeted signals [8].
grams were generated for each signal category, and then used Cyclostationarity-based and wavelet-based schemes require
for training and testing several pre-trained CNN-based mod- long observation intervals. The same is true for DL-based
els, including AlexNet, VGG-16, and VGG-19. Classification algorithms. Additionally, DL algorithms are highly compu-
accuracy was higher than 85% for signals with SNR greater tationally complex and require increased time to converge.
than 5 dB. Accuracy was extremely low for signals with SNR ML-based algorithms require further data preprocessing and
less than 3 dB. An improved deep learning model (i.e., multi- rely on expert knowledge for understanding the data struc-
layer perceptron neural architecture) was proposed in [27] ture. Table 1 provides a summary of the recent algorithms
and [24] to classify received signals based on their mod- proposed in the literature and the pros and cons of each
ulation. Signals were classified with accuracy higher than technique.
95% when SNR was greater than 0 dB. Researchers in [28]
applied Deep Residual Network (ResNet), Convolutional B. RESEARCH CONTRIBUTION AND PAPER
Long Short Term Deep Neural Network (CLDNN), CNN, ORGANIZATION
and RNN on the RadioML dataset. The models successfully In this paper, we developed two novel techniques for wire-
classified 11 wireless technologies with SNRs ranging from less technology identification based on DMD (i.e., a data-
−20 dB to 18 dB. AutoML was employed to reduce time for driven modeling algorithm). We proposed a DMD mode
training and tuning hyper-parameters of the models. CNN amplitude-based identification (DMDA) and a DMD mode
accuracy was highest (85% for signals of SNR>2 dB) and oscillation frequency-based identification (DMDF) technique
RNN was lowest. Our developed DMD-based algorithms for extracting the unique periodic features embedded in
differ from those aforementioned and recent studies by two various wireless standard signals. The proposed algorithms
distinguishing features. First, the DMDF accuracy is indepen- process raw power-measured signals, capturing specifically
dent of signal SNR. The algorithm relies on tracking oscil- embedded periodic features within the targeted signals repre-
lation frequencies for various technologies. While DMDA sented in pilot, preamble, synchronization, and control sig-
algorithm implements power normalization among received nals. The classification was performed for both individual
signals to reduce its dependence on SNR, achieving accu- (or baseline) and different coexistence scenarios. To improve
racies above 90%. Second, the DMD algorithms were able user experiences with signal quality and boost coexistence
to track the periodic preamble transmissions within a signal and spectrum utilization, our schemes offer a strategy that
under poor channel conditions. While the other algorithms can be practically applied to smart radio devices within con-
track signal modulation requiring good channel conditions to temporary HetNets. Based on outcomes, our methods for
achieve high accuracy. using DMD in wireless signal detection provide the following
benefits over equivalent methods:
5) ML-BASED ALGORITHMS 1) Long observation intervals are not necessary. Fewer
Although DL approaches achieve high accuracy models with signal samples are required for the technique to operate.
the advantage of simple feature pre-processing or even raw 2) Truncated singular value decomposition (SVD) is a
data input, they also require large-scale training datasets, technique used by DMD that significantly reduces pro-
resulting in high implementation costs and large computa- cessing time and computational complexity.
tional time. As a result, ML techniques, such as SVM in [1] 3) When processing signals, there is no need for time
and [29] and Random Forest (RF) in [8] and [30], have synchronization.
been widely used in related research for identifying various 4) Identification and classification are direct, one-step
standards’ wireless signals. Researchers have demonstrated processes that do not require further classification tech-
promising results with reduced-size datasets [8]. Individual niques to extract features.

VOLUME 11, 2023 18429

 A. Elsebaay, H. H. Refai: Wireless Technology Identification Employing Dynamic Mode Decomposition Modeling

5) Accuracy is not affected by signal power and SNR Step 1: Compute the SVD of X:
variations, as DMDA normalizes signal power before
classification, and DMDF relies on comparing oscilla- X = Ur 6r Vr∗ (4)
tion frequencies.
where, Ur and Vr consist of r left/right singular vectors corre-
6) The techniques can identify signals in real time.
sponding to the r dominant singular values. 6r is the singular
The balance of this paper is structured as follows. The
values diagonal matrix. The non-negative diagonal elements
methodology and mathematical foundation of the identifi-
of 6r are the r singular values denoted by σi , which are sorted
cation algorithms are described in Section II. Section III
in descending order to satisfy the truncation approximation.
describes the embedded periodic features in the standard
Step 2: Calculate the reduced order matrix ÃϵRnxn .
PHY frame format for different wireless technologies.
Matrix à describes a low-dimensional, approximated linear
The experimental setup and dataset details are described
model of the system:
in Section IV. The results, algorithm evaluation, and valida-
tion are reported in Section V. Finally; section VI concludes à = Ur∗ AUr = Ur∗ X ′ Vr 6r−1 (5)
the paper.
Step 3: Find the eigenvectors and eigenvalues of à by
II. METHODOLOGY solving the equation:
A DMD data-driven modeling algorithm forms the founda-
tion of the identification schemes created for the research ÃW = 3W (6)
reported in this paper. DMD represents a perfect combina-
where, the columns of W are the eigenvectors and 3 is a
tion of proper orthogonal decomposition (POD), and Fourier
diagonal matrix containing the corresponding eigenvalues λk
transforms in the time domain [32]. DMD breaks down a
of both A and Ã.
dynamical system into a number of approximated Koopman
Step 4: Compute the eigenvectors of A (i.e., DMD modes),
modes. In addition to energy (or amplitude), DMD modes are
which are given by column vectors of matrix 8:
ranked by detected dynamics (or frequency). As a result, each
mode has a distinct amplitude and frequency property [33]. 8 = Ur W (7)
Without relying on presumptions, DMD finds dominant
frequencies that show repeating periodicity in signals or B. DATA FORMATTING
systems [34], [35]. We created two wireless technology Available raw power signals are prepared as algorithm input
identification schemes: 1) DMD mode amplitude-based during this stage. As a mapping procedure, we used the Han-
identification (DMDA) and 2) DMD mode oscillation
kel matrix staking method [32], [36] to transform a univariate
frequency-based identification (DMDF). The methods iden-
time series data with length n into a multidimensional matrix
tify and distinguish between various wireless technologies
of size (m x k). The form in (1) describes a time series signal
coexisting in a heterogeneous network utilizing resulting
of length n and fixed sampling time 1t:
DMD eigenvalues and eigenvectors (i.e., DMD modes).
Y = [y1 y2 y3 . . . yn ] (8)
A. THE STANDARD DMD ALGORITHM
Assume a nonlinear dynamical system is approximated by a We created data matrix X and its related one-time step
best linear-fit operator A, which evolves state X forward in evolution data matrix X ′ using the Hankel matrix stacking
time for each k = 1, 2,. . . ..,n-1 approach:
y3 . . .
 
Xk+1 = AXk (1) y1 y2 yk
 y2 y3 y4 . . . yk+1 
Operator AϵRnxn is the best linear fit operator, and it approx- X = .

.. .. .. .. 

(9)
imates Koopman operator. This operator satisfies (1) and  .. . . . . 
represents the solution of a Frobenius norm least-squares ym ym+1 ym+2 . . . ym+k−1
optimization between the one-step future state Xk+1 and the 
y2 y3 y4 . . . yk+1

expected future state AXk :  y3 y4 y5 . . . yk+2 
X′ =  . . .. .. ..  (10)
 
min∥Xk+1 − AXk ∥ (2)  .. .. . . . 
Consequently, we can write the equation in data matrix format ym+1 ym+2 ym+3 . . . ym+k
as
such that
X ′ = AX (3)
m=n−k +1 (11)
where, X is the temporal data matrix and X ′ is the data matrix
advanced one step 1t in the future [32]. where, m is the number of stacks. We defined column vectors
The standard DMD can be formulated in the following of data matrices as snapshots, such that k is the number of
steps: snapshots.

18430 VOLUME 11, 2023

A. Elsebaay, H. H. Refai: Wireless Technology Identification Employing Dynamic Mode Decomposition Modeling

TABLE 1. Recent Wireless technologies identification schemes.

C. DMDA IDENTIFICATION SCHEME

Utilizing eigenvalue matrix 3 and DMD modes matrix 8,
DMDA technique calculates a template-features matrix Fx
from a set of training samples of a certain technology x.
As Fig. 1 illustrates, we projected this matrix on testing sam-
ples of different technologies, then we classified the signal by
evaluating the resultant DMD mode amplitudes. Utilizing the
resultant DMD eigenvalues and modes [32], [37], we arrived FIGURE 1. DMDA classification scheme outline.
at the following:
xi = 83i b (12)
technologies. According to (8), power measurements in Xt
where, i=1, 2, . . . , k, and b is a vector containing the coeffi- affect the values of DMD modes amplitude. Therefore, power
cients of the initial condition x1 in the eigenvector basis, such measurements of various signals should be at comparable
that: levels. Thus, before using DMDA for classification, signal
power must be normalized. Input signals were normalized to
b = 8−1 x1 (13)
a minimum value of zero and a maximum value of one. The
From (12), we combine the two terms 8 and 3i to define the pseudo-code of the proposed DMDA technique is given in
feature matrix as: Algorithm 1.
Fx = 83i (14)
D. DMDF IDENTIFICATION SCHEME
Matrix Fx contains unique features of a wireless technology Frequency of the ith DMD mode can be defined
x, which is composed of eigenvalues and eigenmodes. This as [34], [37]:
matrix represents a model describing a specific technology
ln(λi )
class feature in the DMD domain. DMD modes amplitude fi = |imag( )| (17)
can be calculated as: 1t
fi represents the absolute oscillation frequency of a DMD
bt = Fx−1 Xt (15)
mode. The suggested notion recommends comparing the
Since b is a vector composed of complex values, we will oscillation trend of signals DMD modes, which can be
represent the amplitude by: accomplished by rearranging the fi values for each signal
class into descending order and plotting fi i against an index j,
St = bt b∗t (16)
where j=1, 2, 3,. . . , r. The received signal is then clas-
where, t represents a testing sample and b∗t
is the complex sified using the resulting plots. An overview of the pro-
conjugate of bt . cedure can be found in Fig. 2. To categorize the signals,
The resulting modes’ amplitude determines the detected we used the slope of the ensuing trends. The slope represents
signal class. Given that they include the same features, sub- the decay of the frequencies of the extracted DMD modes.
sequent mode amplitude derived from technology x train- Algorithm 2 provides a detailed illustration of the suggested
ing samples is anticipated to have larger values than other technique.

VOLUME 11, 2023 18431

 A. Elsebaay, H. H. Refai: Wireless Technology Identification Employing Dynamic Mode Decomposition Modeling

Algorithm 1 Wireless Signal Features Matrix Formation Algorithm 2 DMDF Scheme

(Matrix F) Input: first signal y1 , k, m, r, and 1t
Input: signal y1 , k, m, r, and 1t Output: figure
Output: F 1: Calculate required number of samples: n=m+k-1
Calculate required number of samples : 2: Extract required signal window: y1 = y1 (1 : n)
1: n = m + k − 1 Perform Hankel stacking:
Extract required signal window: 3: Initialize Y1
2: Y1 = y1 (1 : n) 4: for j = 1:m do
Perform Hankel stacking: 5: Y11 = [Y11 y1 (:, j : end − m + j)]
3: Initialize Y11 6: end for
4: for j = 1:m do Construct input matrices:
5: Y11 = [Y11 Y1 (:, j : end − m + j)] 7: X = Y11 (:, 1 : end − 1)
6: end for 8: X ′ = Y11 (:, 2 : end)
Repeat for all training signals Compute SVD of X:
Combine to form data matrix: 9: X = U 6V ∗
7: D = [Y11 , Y22 . . . Ynn ]T 10: Ur = U (:, 1 : r)
Construct input matrices for DMD: 11: Sr = S(1 : r, 1 : r)
8: X = Y11 (:, 1 : end − 1) 12: Vr = V (:, 1 : r)
9: X ′ = Y11 (:, 2 : end) 13: Ã = Ur∗ X ′ Vr 6r−1
Apply DMD: 14: ÃW = 3W
10: X = U 6V ∗ 15: 8 = Ur W
11: Ur = U (:, 1 : r) calculate eigenvalues and oscillation frequencies:
12: Sr = S(1 : r, 1 : r) 16: λi = diag(3);
13: Vr = V (:, 1 : r) 17: fi = abs(imag(log(λi /1t)))
14: Ã = Ur∗ X ′ Vr 6r−1 18: fi = sort(fi , descend); Define index j:
15: ÃW = 3W 19: j=[Link]r];
16: 8 = Ur W 20: figure : Plot(j, fi )
Construct features matrix: Repeat for all signals
17: Initialize F
18: for j = 1:k do
19: F = [F; 83j ]
20: end for the 2.4 GHz ISM band. The CC2530 development kit board
Repeat for all signals was used to create ZigBee traffic via channel 14 at a central
frequency of 2.42 GHz. We used two laptops equipped with
nRF52840-BLE Bluetooth 5.3 chipsets for BLE communi-
cation. A National Instruments (NI) vector network analyzer
and PXIe-1075 chassis were used to record the transmit-
ted signals’ raw power measurements. Power measurements
were gathered using a sampling frequency of 500 MHz.
Transmitted signals were recorded with an SNR range of
0 to 25dB for various transmitter and receiver locations and
heights.
To summarize, 660 timeseries signals containing 10.3 mil-
FIGURE 2. DMDF classification scheme outline. lion packets were captured from three scenarios—individual
BLE, individual ZigBee, and heterogeneous coexisting BLE
III. DATASETS AND EXPERIMENTAL SETUP and ZigBee—made up the dataset. For each case, we col-
A. ZIGBEE/BLE DATASET lected 220 signals (i.e., 3.5 million packets). Time domain
Experimental testing and data collection were conducted features of randomly selected signals are shown in Fig. 3.
at the wireless laboratory at the University of Oklahoma
campus in Tulsa. A semi-anechoic chamber was used for B. WIFI DATASET
testing to eliminate external noise. We set up a heteroge- We used a dataset collected by [23] at the wireless labo-
neous wireless network composed of the BLE and ZigBee ratory at the Oklahoma University Tulsa campus. For sig-
sub-networks. There were a pair of access point (Tx) and nals with maximum throughputs of 956 Mbps, 340 Mbps,
station (Rx) devices in each network. 220 separate—none and 250 Mbps, the data set included 450 raw power
sharing baseline—raw power timeseries signals of each tech- time series measurements containing 90 million pack-
nology were collected, in addition to 220 coexisting signals in ets for 802.11n, 802.11ac, and 802.11ax, respectively.

18432 VOLUME 11, 2023

A. Elsebaay, H. H. Refai: Wireless Technology Identification Employing Dynamic Mode Decomposition Modeling

TABLE 2. Average values of WIFI signals duty cycles.

The collection included individual (or baseline) and coexisted

signals in the 5 GHz ISM band with a minimum and constant
noise level of less than −73 dBm. There were three different
Wi-Fi network setups, each with a pair of access points (Tx)
and station (Rx) equipment. To establish an 802.11ax net-
work with 160 MHz of bandwidth on channel 36 (central
frequency 5180 MHz), an Asus RT-AX88U device was used.
The 802.11n and 802.11ac networks were built using two
pairs of MikroTik router boards, and they shared a channel
with the 802.11ax network. LabVIEW software was used to
extract time domain IQ components from raw power mea-
surements of transmitted signals obtained by NI PXIe-5644R FIGURE 4. Raw power measurements of the collected WIFI signals with
RF vector signal transceiver (VST). The real-time bandwidth various duty cycles.

IQ sampling rate was set to 10 MS/s. Fig. 4 depicts the time

domain properties of randomly selected samples with various samples of the available time series signals obtained from the
duty cycles (DCs) for 20ms time interval. Table 2 shows DC dataset.
average values.

C. GSM/LTE DATASET
We used a GSM and LTE dataset provided by [38]. The
dataset consisted of GSM and LTE signals generated at
various SNR levels with a combination of non-line-of-sight
(NLOS) and line-of-sight (LOS) conditions. The exper-
iment was performed in the Wireless Research Labora-
tory in Tubitak, Belgium. Researchers deployed an Agi-
lent vector signal generator (VSG) E4438C as a transmitter.
An Agilent PSA series vector signal analyzer (VSA) E4440A FIGURE 5. Samples of GSM and LTE signals available in the dataset.
was employed as a receiver to capture transmitted signals.
The GSM-average received signal strengths ranged from IV. SIGNAL FRAME FORMAT AND FEATURES
−50 dBm to −45 dBm, while LTE varied between −50 dBm This section explains the signal model and frame structure
and −40 dBm. Sampling time 1t was set to 0.78µs and of the various wireless technologies included in our analysis.
0.14µs for GSM and LTE, respectively. Fig. 5 shows various The inherent unique periodicity in each signal, which serves
as the algorithm’s fundamental property for identification,
is highlighted by the frame structure.

A. GSM/LTE
GSM frame structure is a time division multiple access
(TDMA). Each frame consists of eight timeslots. Fig. 6
shows timeslot-per-frame for a normal burst [39], which
carries encrypted data transmitted between users. We noted
the periodicity of the pilot training signals, tail bits (TB),
and guard bits (GB). The dedicated 26 bits for the training
(i.e., pilot) signal used for channel estimation in each time
slot are repeated in the same instance per slot. Since the
duration of each timeslot is 577µs, the repetitive frequency
of the pilot sequence is 1/577=1733 Hz. Guard and tail
(i.e., synchronization) bits have the same value of repetitive-
ness. Other signaling GSM bursts (i.e., frequency correction,
synchronization, and access bursts) have similar repetitive
sequences with the same 1733 Hz frequency, although they
FIGURE 3. Zigbee and BLE Raw power signals. have a different duration.

VOLUME 11, 2023 18433

 A. Elsebaay, H. H. Refai: Wireless Technology Identification Employing Dynamic Mode Decomposition Modeling

Detailed frame structure of LTE FDD downlink can be

found in [40] and [41]. Every 10ms LTE frame is divided into
10 subframes, each with 1ms duration. Also, each subframe
contains two timeslots (or resource blocks [RBs]) character-
ized by 0.5ms duration and six or seven OFDM symbols,
depending on short- or long-cyclic prefix. The periodicity
of various signals demonstrates a constant repetition for the
following:
1) Reference/pilot signals (blue): Fixed location along the
time axis on the first and fifth OFDM symbol of every
RB. These are repeated once per RB (i.e., one time
per 0.5ms). FIGURE 7. 802.11n, 802.11ac, and 802.11ax packet structure.
2) PDCCH, PCFICH, PHICH (yellow, red, purple, see
below): Exist at the beginning symbol of each subframe
and are repeated once per subframe (i.e., one time
per 1ms).
3) PSS and SSS (green and orange): Repeated at fixed
locations once every 5ms.
where, PDCCH is Physical Downlink Control Channel;
PCFICH is Physical Control Format Indicator Channel;
PHICH is Physical HARQ Indicator Channel; PSSis Primary
Synchronization Signal; and SSS is Secondary Synchroniza- FIGURE 8. Zigbee and BLE packet structure.
tion Signal.
C. ZIGBEE/BLE
B. WIFI
Data is transmitted via packets by the Zigbee IEEE
Highlighting the physical layer (PHY) frame structure of 802.15.4 PHY. As indicated in Fig. 8, each packet is com-
802.11n, 802.11ac, and 802.11ax Wi-Fi standard signals in posed of a preamble (32 bits) for synchronization, the start
the 5 GHz bands, the 802.11 PHY employs bursted packets of packet delimiter (8 bits), physical header (8 bits), and
for transmission. There is a preamble and payload in each payload segment data unit (0 to 1016 bits) [42]. According
packet. Preamble enables synchronization of time and fre- to the BLE v5.1 standard created by the Bluetooth SIG,
quency, estimates channel parameters for equalization, and a PHY transmitted packet has four parts: the synchronization
gives receiver header details about the packet, such as con- preamble (8 bits), access address (32 bits), protocol data
figuration, format type, and data rates. Data from the user is unit (PDU)—advertisement or data packet (2-257 octets), and
transmitted in the payload. cyclic redundancy check (CRC), which is used to identify
Fig. 7 shows the packet formats for 802.11n, 802.11ac, packet errors (24 bits) [43].
and 802.11ax Wi-Fi standard signals. The 802.11n high
throughput (HT) mixed format begins with legacy pream- V. RESULTS AND DISCUSSIONS
bles: legacy Short Training sequence (L-STF), legacy Long A. SELECTION OF DMDA AND DMDF INPUT PARAMETERS
Training sequence (L-LTF), and legacy Signal Description
As shown earlier, algorithms 1 and 2 required four input
(L-SIG), which can be decoded by legacy 11a/g devices.
parameters for carrying out the proposed classification algo-
Legacy preambles are followed by the 11n specific HT
rithm, as follows.
preambles, and finally the user data. The 802.11ac frame
1) Number of snapshots k: Inputting the entire captured
format has the same beginning legacy preambles, followed by
raw signal into the algorithm results in unnecessary
the very high throughput (VHT) preambles, which are unique
large computational time. Instead, a minimum number
for 11ac devices.
of snapshots (i.e., number of column vectors of data
matrices) should be used in data matrix X. Based on the
discussion in Section II, for DMD to accurately capture
the periodic features embedded in the signals, we must
choose an adequate number of snapshots to represent
FIGURE 6. GSM timeslot structure for normal burst [23]. an ample number of packet timeslots. We used the
following formula to calculate the required k value:
Like 11n and 11ac, the 802.11ax packet begins with a tradi-
Nslot Tslot
tional preamble followed by a High efficiency (HE) preamble k= (18)
sequence that can only be decoded by 11ax devices. At the 1t
end of an 802.11ax frame, a packet extension (PE) with a where, Nslot is the number of the standard frame
length of 8 or 16 Bµs can be employed. packet duration and Tslot is the duration of each packet.
18434 VOLUME 11, 2023
A. Elsebaay, H. H. Refai: Wireless Technology Identification Employing Dynamic Mode Decomposition Modeling

Empirically, we found that Nslot =4 was sufficient for 2) APPLYING DMDF TECHNIQUE
DMD to capture repetitive sequence frequencies and The DMDF identification method classified signals based
assign a signal (See next subsection). on evaluating the slope of DMD mode oscillation frequency
2) Number of stacks m for Hankel matrix formulation: trend, as explained in Section II. We applied the technique on
The Hankel matrix number of rows m significantly all the available GSM and LTE samples, and also plotted the
impacts DMD accuracy. The value of m is dependent oscillation frequency trend for 20 modes (i.e., value of r was
on the length of the time series signal n. We found that set to 40, and then we removed duplicate values, as resul-
when m was smaller than n/2 or approximate to n, error tant eigenvalues were complex conjugates). Fig. 11 shows
increased and accuracy decreased. Therefore, the value that LTE testing samples had a more pronounced, distinct
of m follows the threshold [34], [37]: trend than GSM testing samples (i.e., higher slope). Fig. 11
also shows that only one GSM signal could potentially be
n/2 < m < n (19) misclassified. When observing this signal, it is obvious that
the signal contained a high amount of noise when compared
We chose m to be approximately 60% of the value of n: with other signals, which affected algorithm accuracy. The
absolute value of slope α of the linearly fitted line for modes
m = 0.6n (20) oscillation frequencies was utilized for identifying each sig-
nal class. We developed the identification rule as:
3) Truncation value r: The optimal value of r can be found
from the inflection point in the decay of the singular αLTE > αGSM (21)
values of data matrix X [32], which are the elements where, αLTE and αGSM are the absolute values of slope of the
of the diagonal matrix 6 of the SVD process. For linearly fitted line for modes and their associated oscillation
example, Fig. 9 indicates the decay of the singular frequencies of LTE and GSM signals, respectively.
values of GSM signal. Singular values were normalized
with respect to the first singular value. Based on the
inflection point, six modes are ample to represent a
GSM signal. This result confirms the significance of
DMD, which required only a few numbers of extracted
modes (i.e., dominant modes) for identifying a signal.
4) Number of samples n: This parameter represents the
length of the timeseries signal, as indicated in (8). The
value can be calculated using (4) and (11), given the
value of k, as calculated in (18).

B. GSM/LTE IDENTIFICATION
1) APPLYING DMDA TECHNIQUE
This section describes results obtained using DMDA signal
identification. DMDA evaluates the resultant DMD mode
amplitudes. Fig. 10 shows the resulting modes of two test-
ing samples for a various number of transmitted timeslots
(or packets) Nslot . The algorithm achieved better distinc- FIGURE 9. Decay of singular values for GSM signal.
tion as Nslot increased. We evaluated six resultant DMD
mode amplitudes of GSM and LTE test samples relative to
features stored in matrix F, which was constructed using 3) EVALUATING PERFORMANCE OF DEVELOPED
GSM training signals (see Fig. 1). DMD mode amplitudes TECHNIQUES
were evaluated by comparing maximum and average values. We evaluated DMDF and DMDA technique performance
As expected, GSM test samples had higher amplitudes than using classification accuracy and processing time required
LTE test samples due to common features matching with to classify a signal. Concerning DMDA, signals were clas-
matrix F. Results confirm the ability of DMDA to identify sified by comparing the maximum value of DMD modes
signals with short time duration (i.e., only 5 packets/timeslots amplitude, as indicated in Fig. 10. GSM signals had higher
are sufficient). Regarding GSM, Nslot =4 was empirically modes amplitude than LTE signals. Classification accuracy of
found to sufficiently capture the repetitive sequence fre- both GSM and LTE was 90%. The processing time required
quency, as clearly shown in Fig. 10. Given Tslot =577µs to identify a signal is approximately 1.5 seconds. The time
and 1t=0.78 µs, k equals approximately 2950. For LTE, required to train the DMDA model with four GSM signals
Nslot =5 was empirically found to capture repetitive sequence was approximately 20 seconds. Employing DMDF scheme,
frequencies. Given Tslot =0.5ms and 1t=0.14 µs, k equals to where GSM acquires lower slope than that of LTE, DMDF
17850 approximately. technique achieved a 90% accurate detection for GSM

VOLUME 11, 2023 18435

 A. Elsebaay, H. H. Refai: Wireless Technology Identification Employing Dynamic Mode Decomposition Modeling

TABLE 3. GSM/LTE classification models accuracy.

C. WIFI STANDARD TECHNOLOGY IDENTIFICATION

Classification was performed on both individual (or baseline)
and various coexistence scenarios of 802.11n, 802.11ac, and
802.11ax Wi-Fi standard signals in the 5 GHz ISM band.
Processed time series raw power signals were collected in a
shared heterogeneous deployment [44].

1) INDIVIDUAL SIGNALS IDENTIFICATION

FIGURE 10. Resulted modes amplitudes for testing samples with We applied DMDA technique for classifying/differentiating
different values of Nslot .
between individual (or baseline) 802.11n, 802.11ac, and
802.11ax Wi-Fi standard signals in the 5 GHz ISM band.
Classified signals had the same duty cycle. In the avail-
able dataset, each technology contained 25 individual signals
divided into five groups. Each group contained five signals
having a specific duty cycle (see Table 2). In Fig. 12, we eval-
uated six resultant DMD mode amplitudes for three random
testing samples, regarding features stored in the 802.11ac fea-
ture matrix Fac . Mode amplitudes were evaluated by compar-
ing maximum and average values. Fig. 12 indicates that the
DMDA technique was able to clearly differentiate between
the targeted signals. The 802.11ac signals had the highest
amplitudes due to common features matching with the fea-
FIGURE 11. Resulted oscillation frequency trend for LTE and GSM signals. ture matrix Fac . The 802.11n signals had lower amplitude
than 802.11ac, while the 802.11ax signals had the lowest
amplitude. We empirically developed the identification rule
and 100% for LTE. DMDF technique required 3 seconds to expressed as:
identify a signal with no need for training.
Sac > Sn > Sax (22)
Our developed techniques were compared to [7], which
utilized second-order cyclostationarity to calculate spectral where, Sac , Sn , and Sax are the maximum mode amplitude for
correlation functions (SCFs). Resulting SCFs were utilized Wi-Fi-802.11ac, 802.11n, and 802.11ax signals, respectively.
as an input to SVM for classifying various signals, including Applying DMDF on the same dataset for each signal,
GSM and LTE. The method achieved 78% accurate detection we plotted the oscillation frequencies for 16 modes. Fig. 13
for GSM and 100% for LTE signals. In addition, we com- shows the resulted modes oscillation frequencies of a ran-
pared the technique’s performance with [15] and [16], which domly chosen signal for each duty cycle. More specifically,
distinguished GSM signals from LTE signals by applying the absolute value of slope α of the linearly fitted line for
CDF analysis using the same dataset used in our analysis. modes oscillation frequencies (or dashed line) was utilized
The magnitude of GSM signals follows a Rician distribution, for identifying each signal category. As Fig. 13 indicates, the
while the magnitude of LTE signals follows a Rayleigh dis- identification rule for all duty cycles can be defined as:
tribution. Using a distribution fitting MATLAB tool and CDF
calculation, we classified GSM and LTE based on their CDF. αac > αn > αax (23)
Results achieved 80% accurate detection for GSM and 70 %
for LTE. Table 3 provides a comprehensive summary of the where, αac , αn , and αax are the absolute value of slope from
overall classification accuracy for the compared techniques. the linearly fitted line for modes oscillation frequencies of
We selected the specific baseline methods for evaluation and Wi-Fi-802.11ac, 802.11n, and 802.11ax signals, respectively.
comparison due to their relevance in approach and recency This result indicated that the DMDF technique can extract
to the developed DMDA and DMDF methods. Cyclostation- unique features and accurately classify targeted signals.
ary methods track the periodicity of pilot signals embedded
in their transmission. Although this approach is similar to 2) COEXISTING SIGNALS IDENTIFICATION
our developed DMDA and DMDF methods, ours requires We created five coexistence scenarios, each coexisting two
fewer samples with less complexity to facilitate accurate Wi-Fi signals having the same duty cycle. Each scenario was
identification. repeated according to the duty cycles provided in Table 2.

18436 VOLUME 11, 2023

A. Elsebaay, H. H. Refai: Wireless Technology Identification Employing Dynamic Mode Decomposition Modeling

FIGURE 12. Individual Wi-Fi signals identification employing DMDA.

We applied a DMDF technique to differentiate between coex-

isting signals and individual signals of the same Wi-Fi tech-
nology. Figs. 14, 15, and 16 compared the resulted modes
oscillation frequencies of a randomly chosen signal from each
coexistence scenario with individual signals having the same
duty cycle. The slope of the linearly fitted line for modes FIGURE 13. Individual Wi-Fi signals identification employing DMDF.
oscillation frequencies (see dashed line) demonstrates that
the coexisted signal has the lowest slope of all cases. On the TABLE 4. Threshold values used for signals classification.
other hand, DMDA was able to classify ac, n, and coexisted
ac-n signals. However, DMDA was not able to differenti-
ate between ax and coexisted signals containing ax (e.g.,
coexisted ax-ac and ax-n signals) as they had similar ampli-
tude values without enough separating margin for accurate
classification.

3) PERFORMANCE EVALUATION evaluate our developed DMDA with DMDF techniques.

DMDA classified the signals by thresholding the maximum Comparing obtained results with the WaveNet deep learning
value of DMD modes amplitude. DMDF classified the signals model performed in [23], researchers achieved lower clas-
by thresholding the values of slopes. Selected thresholding sification accuracies of 96%, 98%, and 98% for n, ac, and
values are indicated in Table 4. ax individual signals. For coexisted signals, their technique
Regarding individual signals classification, DMDA tech- achieved 90% accuracy for classifying coexisted ax-n and
nique correctly classified ac signals with 84% accuracy, n sig- 91% for coexisted ax-ac signals. The proposed DMDF tech-
nals with 88% accuracy, and ax signals with 92% accuracy. nique not only achieved higher accuracy but also, the compu-
DMDF achieved better accuracy, as it correctly classified tational time was extensively reduced by approximately 85%.
signals with overall accuracy of 98.6 %. DMDF classifi- In addition, a developed CNN architecture and ResNet9
cation accuracy of n, ac, and ax signals was 96%, 100%, CNN model [45] were trained to classify the collected sig-
and 100%, respectively. On the other hand, while classify- nals. We utilized gramian angular summation field (GASF)
ing individual and coexisted signals of the same technology transformation [46] to extract features. GASF was utilized
(as Figs. 16, 17, and 18 indicated), DMDF technique achieved to encode the collected, one-dimensional (1D) raw power
93.3% accuracy for classifying n, ac, and coexisted n-ac sig- timeseries signals into two-dimensional (2D) texture images
nals; 93.3% for classifying n, ax, and coexisted n-ax signals, for inputting into CNN classifier models. Generated images
and 97.3% for classifying ax, ac, and coexisted ax-ac signals. were 300 × 300 pixels. Because signal power measure-
DMDA was able to classify n, ac, and coexisted n-ac signals ments are distinct and to alleviate the effect of diverse SNR
with accuracy of 82%. values between technologies, a preprocessing normalization
CNN, ResNet, and WaveNet are indicative of recently step was performed on signals before applying the trans-
studied and published deep learning architectures used for formation. Training was conducted on 50% of the dataset
wireless signal identification. This type of approach is some- (see section III) for both individual and coexisting scenarios.
what relevant to our developed algorithms since data training with batch size of eight images using a cross-entropy loss
and model development are part of our approach. Hence, function along with an Adagrad optimizer [47]. Learning rate
we selected the aforementioned models to compare and was set to 0.001. Models were implemented, trained, and

VOLUME 11, 2023 18437

 A. Elsebaay, H. H. Refai: Wireless Technology Identification Employing Dynamic Mode Decomposition Modeling

FIGURE 14. Identifying coexisted ac-n Wi-Fi signals employing DMDF. FIGURE 16. Identifying coexisted n-ax Wi-Fi signals employing DMDF.

For CNN models, the best accuracy was attained for

50 epochs. It should be noted that CNN models are expected
to have higher accuracy for larger training samples when
compared with the sample trained and reported herein. How-
ever, this will be at the expense of additional time required to
train the models.

D. ZIGBEE-BLE IDENTIFICATION
We proposed DMDA and DMDF techniques to classify indi-
vidual (i.e., none sharing baseline) and coexistence scenar-
ios for experimentally collected BLE and Zigbee signals in
a shared heterogeneous deployment in the 2.4 GHz ISM
band [48].

1) APPLYING DMDA AND DMDF SCHEME

In Fig. 18, we evaluated six resultant DMD mode amplitudes
for three random testing samples with regard to features
stored in the Zigbee feature matrix Fz . Mode amplitudes
were evaluated by comparing maximum and average values.
Fig. 18 indicates that DMDA technique was able to clearly
FIGURE 15. Identifying coexisted ac-ax Wi-Fi signals employing DMDF. differentiate between the targeted signals. As expected, Zig-
Bee test samples had the highest amplitudes due to common
tested in the open-source Pytorch software library. The devel- features with Fz . BLE test signals had lower amplitude
oped CNN was structured with six transformation layers, than ZigBee; the coexisted signal had the lowest amplitude.
and the network was convolutional with batch normalization We empirically developed the identification rule expressed
and ReLU activation layers, a pooling layer, and dense fully as:
connected (FC) layers. The final FC layer was the output
Szig. > SBLE > Sco (24)
layer, which computes scores for each of the three class
labels. The label with the highest score was the one predicted where, Szig. , SBLE , and Sco are the maximum mode ampli-
by the model. Fig. 17 depicts the complete network structure tude for ZigBee, BLE, and coexisted ZigBee/BLE signals,
and Kernel filter sizes K. respectively.
Table 5 evaluated and compared model performance in In Fig. 19 we plotted the oscillation frequency trend for
terms of overall classification accuracy, processing time 10 unique modes based on two randomly chosen dataset
required to identify a signal, and model training time. samples. Each technology had a pronounced, distinct trend

18438 VOLUME 11, 2023

A. Elsebaay, H. H. Refai: Wireless Technology Identification Employing Dynamic Mode Decomposition Modeling

using the same set of testing samples. Table 7 provided a

comparison of the performance of the implemented models.
Results showed that the developed DMD-based techniques
achieved high performance for classifying various individual
and coexisted wireless signals. In most cases analyzed for
this paper, DMD-based techniques had the advantage of high
accuracy and short training/processing time required to iden-
tify a signal when compared to other techniques. The DMDF
scheme outperformed all compared techniques and did not
require training. DMDF also achieved the highest accuracy
and shortest processing time for identifying a signal.

E. COMPUTATIONAL COMPLEXITY
FIGURE 17. CNN network developed for classifying ISM band wireless Four algorithms (CNN, ResNet, WaveNet, and cyclostation-
signals. ary) were implemented to confirm the computational advan-
tage of our developed techniques over others. DMDA/DMDF
TABLE 5. Performance evaluation for various implemented models. computational complexity was attributed to the SVD calcu-
lation [32]. Notably, our algorithm was successfully imple-
mented using a truncated SVD solution. Doing so lim-
ited its complexity to O(rn2 ). Truncation value r was less
than 10, and n was limited to less than five wireless packet
samples. However, the implemented 3-layer CNN model
complexity was O(knd 2 ) per convolutional layer, where d
is the layer dimension [49]. Furthermore, the implemented
9-layer ResNet complexity was O(knd 2 ) per layer. Addition-
ally, the implemented WaveNet had a very high exponential
(i.e., slope). We empirically developed the identification rule complexity O(2L ), where L is the number of layers of the
expressed as:
αzig. > αBLE > αco (25)
where, αzig. , αBLE , and αco are the absolute value of slope
from the linearly fitted line for modes oscillation fre-
quencies ZigBee, BLE, and coexisted ZigBee/BLE signals,
respectively.

2) PERFORMANCE EVALUATION
Concerning DMDA, signals were classified by thresholding
the maximum value of DMD modes amplitude, as indicated
in Table 6. The technique correctly classified signals with
overall accuracy of 86.3%. Classification accuracy of Zig-
Bee, BLE, and coexisted signals was 86%, 81%, and 91%,
respectively.
Applying DMDF, signals were classified by thresholding
FIGURE 18. ZigBee-BLE identification employing DMDA scheme.
values of the slopes, as indicated in Table 6. DMDF acheived
an overall classification accuracy of 87.6%. Classification
accuracy of ZigBee, BLE, and coexisted signals was 91%,
85.5%, and 86.4%, respectively.
We evaluated the performance of DMDA and DMDF tech-
niques using overall classification accuracy, processing time
required to identify a signal, and model training time. Both
techniques were compared with the developed CNN model
and ResNet-9 model for validation (see subsection C). CNN-
based models training was conducted on 50% of the dataset
with a batch size of eight images using a cross-entropy loss
function along with an Adagrad optimizer. The learning rate
was set to 0.001. The accuracy of all models was calculated FIGURE 19. ZigBee-BLE identification employing DMDF scheme.

VOLUME 11, 2023 18439

 A. Elsebaay, H. H. Refai: Wireless Technology Identification Employing Dynamic Mode Decomposition Modeling

network (i.e., large number) [50]. Finally, cyclostationary identifying three wireless technologies simultaneously oper-
complexity was O(2n[4 + 2log2h + 4n + 2h + hlog2(4n/h)]), ating in the same environment. In future work, the authors
where h is the FFT number of points [1]. In comparison, would like to evaluate the scalability of the algorithms to track
our approach required fewer signal samples and achieved more networks with more diverse coexistence scenarios.
higher accuracy with less computational time, as shown
in Tables 5 and 7. REFERENCES
[1] K. Tekbiyik, O. Akbunar, A. R. Ekti, A. Gorcin, and G. K. Kurt,
TABLE 6. Threshold values used for ZigBbee/BLE classification. ‘‘Multi–dimensional wireless signal identification based on support vector
machines,’’ IEEE Access, vol. 7, pp. 138890–138903, 2019.
[2] X. Li, F. Dong, S. Zhang, and W. Guo, ‘‘A survey on deep learning tech-
niques in wireless signal recognition,’’ Wireless Commun. Mobile Comput.,
vol. 2019, pp. 1–12, Feb. 2019.
[3] N. Kato, B. Mao, F. Tang, Y. Kawamoto, and J. Liu, ‘‘Ten challenges
in advancing machine learning technologies toward 6G,’’ IEEE Wireless
Commun., vol. 27, no. 3, pp. 96–103, Jun. 2020.
[4] G. Naik, J. Liu, and J.-M. Park, ‘‘Coexistence of wireless technologies in
the 5 GHz bands: A survey of existing solutions and a roadmap for future
TABLE 7. Performance evaluation for various implemented models. research,’’ IEEE Commun. Surveys Tuts., vol. 20, no. 3, pp. 1777–1798,
3rd Quart., 2018.
[5] N. Bitar, S. Muhammad, and H. H. Refai, ‘‘Wireless technology identifica-
tion using deep convolutional neural networks,’’ in Proc. IEEE 28th Annu.
Int. Symp. Pers., Indoor, Mobile Radio Commun. (PIMRC), Oct. 2017,
pp. 1–6.
[6] S. Sun, L. Gong, B. Rong, and K. Lu, ‘‘An intelligent SDN framework
for 5G heterogeneous networks,’’ IEEE Commun. Mag., vol. 53, no. 11,
pp. 142–147, Nov. 2015.
[7] K. Tekbiyik, O. Akbunar, A. R. Ekti, G. K. Kurt, and A. Gorcin, ‘‘On
the investigation of wireless signal identification using spectral correla-
VI. CONCLUSION tion function and SVMs,’’ in Proc. IEEE Wireless Commun. Netw. Conf.
Wireless identification technologies and effective coexistence (WCNC), Marrakesh, Morocco, Apr. 2019, pp. 1–5.
management are desperately needed, as the number of wire- [8] C. Skiribou and F. Elbahhar, ‘‘V2X wireless technology identification
using time–frequency analysis and random forest classifier,’’ Sensors,
less technologies operating in both licensed and unlicensed vol. 21, no. 13, p. 4286, Jun. 2021.
bands continues to increase. The research highlighted in [9] B. Liu and K. C. Ho, ‘‘Identification of CDMA signal and GSM signal
this paper reports the development of two novel techniques using the wavelet transform,’’ in Proc. 42nd Midwest Symp. Circuits Syst.,
for wireless technology identification, namely DMDA and Las Cruces, NM, USA, Aug. 1999, pp. 678–681.
[10] K. C. Ho, H. Liu, and L. Hong, ‘‘On improving the accuracy of a wavelet
DMDF, which are the first to employ a DMD algorithm based identifier to classify CDMA signal and GSM signal,’’ in Proc.
for identifying signals. Our identification algorithms are IEEE Int. Symp. Circuits Syst. VLSI, Orlando, FL, USA, May 1999,
based on capturing periodic features embedded in targeted pp. 564–567.
[11] M. Oner and F. Jondral, ‘‘Cyclostationarity based air interface recognition
signals by utilizing a DMD mathematical platform. The for software radio systems,’’ in Proc. IEEE Radio Wireless Conf., Atlanta,
DMDA technique is based on evaluating a template-features GA, USA, Sep. 2004, pp. 263–266.
matrix obtained for a specific signal class, and then clas- [12] E. Karami, O. A. Dobre, and N. Adnani, ‘‘Identification of GSM and
sifying signals under evaluation according to the values of LTE signals using their second-order cyclostationarity,’’ in Proc. IEEE
Int. Instrum. Meas. Technol. Conf. (I2MTC), Pisa, Italy, May 2015,
the resulting projected DMD mode amplitude. The DMDF pp. 1108–1112.
technique classifies signals based on evaluating the slope [13] M. Kulin, T. Kazaz, I. Moerman, and E. De Poorter, ‘‘End-to-end learn-
of the DMD modes oscillation frequency trend. The pro- ing from spectrum data: A deep learning approach for wireless signal
identification in spectrum monitoring applications,’’ IEEE Access, vol. 6,
posed techniques were implemented on experimental datasets pp. 18484–18501, 2018.
captured under various channel conditions. When compared [14] G. C. Zanuz, J. M. Winter, I. Müller, J. L. T. Garzon, J. C. Netto, and
with various similar techniques implemented in the litera- C. E. Pereira, ‘‘Identification of IEEE 802.11 g and IEEE 802.15.4 signals
using energy and cyclostationarity detection approach,’’ in Proc. 1st Int.
ture, results showed that DMD-based algorithms had lower Symp. Instrum. Syst., Circuits Transducers (INSCIT), Belo Horizonte,
complexity and achieved higher performance. The algorithms Brazil, Aug. 2016, pp. 55–60.
can differentiate and classify wireless signals with high accu- [15] Y. A. Eldemerdash, O. A. Dobre, O. Ureten, and T. Yensen, ‘‘Fast and
racy by employing short observation intervals (e.g., four-time robust identification of GSM and LTE signals,’’ in Proc. IEEE Int. Instrum.
Meas. Technol. Conf. (I2MTC), Turin, Italy, May 2017, pp. 1–6.
slots or packets), which enable them to be implemented in [16] Y. A. Eldemerdash, O. A. Dobre, O. Üreten, and T. Yensen, ‘‘Identification
real-time for identifying various operating wireless technolo- of cellular networks for intelligent radio measurements,’’ IEEE Trans.
gies coexisting in heterogeneous networks. DMDF scheme Instrum. Meas., vol. 66, no. 8, pp. 2204–2211, Aug. 2017.
[17] H. Xia, K. Alshathri, V. B. Lawrence, Y.-D. Yao, A. Montalvo, M. Rauchw-
outperformed all compared techniques, as it is a direct iden- erk, and R. Cupo, ‘‘Cellular signal identification using convolutional neural
tification scheme (i.e., no training required). This technique networks: AWGN and Rayleigh fading channels,’’ in Proc. IEEE Int. Symp.
achieved the highest accuracy—greater than 90% for most Dyn. Spectr. Access Netw. (DySPAN), Nov. 2019, pp. 1–5.
cases—and the shortest time required to identify a signal— [18] K. Alshathri, H. Xia, V. Lawrence, and Y.-D. Yao, ‘‘Cellular system
identification using deep learning: GSM, UMTS and LTE,’’ in Proc.
less than one second. The proposed DMD-based identifica- 28th Wireless Opt. Commun. Conf. (WOCC), Beijing, China, May 2019,
tion algorithms showed promising performance in accurately pp. 1–4.

18440 VOLUME 11, 2023

A. Elsebaay, H. H. Refai: Wireless Technology Identification Employing Dynamic Mode Decomposition Modeling

[19] H. Gu, Y. Wang, S. Hong, Y. Xu, and G. Gui, ‘‘Deep learning based intel- [41] S.-W. Kim and K.-Y. Kim, ‘‘Physical layer verification for 3GPP
ligent recognition method in heterogeneous communication networks,’’ in LTE (FDD),’’ in Proc. 11th Int. Conf. Adv. Commun. Technol., 2009,
Proc. IEEE/CIC Int. Conf. Commun. China (ICCC), Chongqing, China, pp. 1095–1100.
Aug. 2020, pp. 478–482. [42] L. Chaari and L. Kamoun, ‘‘Performance analysis of IEEE 802.15.4/Zig-
[20] M. Schmidt, D. Block, and U. Meier, ‘‘Wireless interference identification Bee standard under real time constraints,’’ Int. J. Comput. Netw. Commun.,
with convolutional neural networks,’’ in Proc. IEEE 15th Int. Conf. Ind. vol. 3, no. 5, pp. 235–251, Sep. 2011.
Informat. (INDIN), Emden, Germany, Jul. 2017, pp. 478–482. [43] Bluetooth Low Energy—It Starts With Advertising. Accessed: Apr. 2021.
[21] J. Fontaine, E. Fonseca, A. Shahid, M. Kist, L. A. DaSilva, I. Moerman, [Online]. Available: [Link]
and E. De Poorter, ‘‘Towards low-complexity wireless technology classi- [44] A. Elsebaay and H. H. Refai, ‘‘Wi-Fi standard technology identifica-
fication across multiple environments,’’ Ad Hoc Netw., vol. 91, Aug. 2019, tion employing dynamic mode decomposition modeling,’’ in Proc. IEEE
Art. no. 101881. Global Commun. Conf., Rio de Janeiro, Brazil, Dec. 2022, pp. 4788–4793.
[22] E. Almazrouei, G. Gianini, N. Almoosa, and E. Damiani, ‘‘Robust [45] Using ResNet for Image Classification. Accessed: Mar. 2021. [Online].
computationally-efficient wireless emitter classification using autoen- Available: [Link]
coders and convolutional neural networks,’’ Sensors, vol. 21, no. 7, p. 2414, 4b3c42f2a27e
Apr. 2021. [46] C.-L. Yang, C.-Y. Yang, Z.-X. Chen, and N.-W. Lo, ‘‘Multivariate
[23] R. J. Estes, D. Willis, and H. H. Refai, ‘‘Classifying Wi-Fi from raw power time series data transformation for convolutional neural network,’’ in
measurements using a neural network adapted from WaveNet,’’ in Proc. Proc. IEEE/SICE Int. Symp. Syst. Integr. (SII), Paris, France, Jan. 2019,
IEEE Int. Conf. Commun., Montreal, QC, Canada, Jun. 2021, pp. 1–6. pp. 188–192.
[24] J. Gong, X. Qin, and X. Xu, ‘‘Multi-task based deep learning approach for [47] J. Duchi, E. Hazan, and Y. Singer, ‘‘Adaptive subgradient methods for
open-set wireless signal identification in ISM band,’’ IEEE Trans. Cognit. online learning and stochastic optimization,’’ J. Mach. Learn. Res., vol. 12,
Commun. Netw., vol. 8, no. 1, pp. 121–135, Mar. 2022. no. 7, pp. 1–39, 2011.
[25] M. He, S. Peng, H. Wang, and Y.-D. Yao, ‘‘Identification of ISM band [48] A. Elsebaay and H. H. Refai, ‘‘Identifying coexisting Bluetooth and
signals using deep learning,’’ in Proc. 29th Wireless Opt. Commun. Conf. ZigBee technologies employing dynamic mode decomposition,’’ in
(WOCC), May 2020, pp. 1–4. Proc. IEEE Globecom Workshops, Rio de Janeiro, Brazil, Dec. 2022,
[26] H. S. Ghanem, R. M. Al-Makhlasawy, W. El-Shafai, M. Elsabrouty, pp. 432–437.
H. F. A. Hamed, G. M. Salama, and F. E. A. El-Samie, ‘‘Wireless mod- [49] A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez,
ulation classification based on radon transform and convolutional neural and I. Polosukhin, ‘‘Attention is all you need,’’ in Proc. Adv. Neural Inf.
networks,’’ J. Ambient Intell. Humanized Comput., vol. 2022, pp. 1–10, Process. Syst., vol. 30, 2017, pp. 1–16.
May 2022. [50] S. Hussain, M. Javaheripi, P. Neekhara, R. Kastner, and F. Koushanfar,
[27] C. T. Fredieu, A. Martone, and R. M. Buehrer, ‘‘Open-set classification ‘‘FastWave: Accelerating autoregressive convolutional neural networks on
of common waveforms using a deep feed-forward network and binary FPGA,’’ in Proc. IEEE/ACM Int. Conf. Comput.-Aided Design (ICCAD),
isolation forest models,’’ in Proc. IEEE Wireless Commun. Netw. Conf. Nov. 2019, pp. 1–8.
(WCNC), Austin, TX, USA, Apr. 2022, pp. 2465–2469.
[28] K. S. Durbha and S. Amuru, ‘‘AutoML models for wireless signals classi-
fication and their effectiveness against adversarial attacks,’’ in Proc. 14th
Int. Conf. Commun. Syst. Netw. (COMSNETS), Bangalore, India, Jan. 2022, AHMED ELSEBAAY received the [Link]. degree
pp. 265–269. in electrical power and machine engineering and
[29] L.-X. Wang and Y.-J. Ren, ‘‘Recognition of digital modulation signals the [Link]. degree in electrical engineering from
based on high order cumulants and support vector machines,’’ in Proc. Helwan University, Cairo, Egypt, in May 2012 and
ISECS Int. Colloq. Comput., Commun., Control, Manage., Sanya, China, May 2018, respectively. He is currently pursuing
Aug. 2009, pp. 271–274. the Ph.D. degree in electrical and computer engi-
[30] X. Wang, Z. Gao, Y. Fang, S. Yuan, H. Zhao, W. Gong, M. Qiu, and Q. Liu, neering with The University of Oklahoma, Tulsa,
‘‘A signal modulation type recognition method based on kernel PCA and OK, USA. From 2013 to 2019, he worked as
Random Forest in cognitive network,’’ in Proc. Int. Conf. Intell. Comput., a Research Assistant with the Faculty of Engi-
2014, pp. 522–528. neering, Helwan University. From 2019 to 2020,
[31] S. Grimaldi, A. Mahmood, and M. Gidlund, ‘‘Real-time interference iden- he also worked as a Researcher with the Public University of Navarra,
tification via supervised learning: Embedding coexistence awareness in Pamplona, Spain. Since 2020, he has been with The University of Oklahoma,
IoT devices,’’ IEEE Access, vol. 7, pp. 835–850, 2019. where he is also working on multiple research projects focusing on wireless
[32] J. N. Kutz, S. L. Brunton, B. W. Brunton, and J. L. Proctor, Dynamic coexistence, signal processing, mathematical modeling, machine learning,
Mode Decomposition: Data-Driven Modeling of Complex Systems. and deep learning. His research interests include the Internet of Things and
Philadelphia, PA, USA: SIAM, 2016. Accessed: Jan. 28, 2022, doi: data collection by deploying wireless sensor networks.
10.1137/1.9781611974508.
[33] Z. Wu, D. Laurence, S. Utyuzhnikov, and I. Afgan, ‘‘Proper orthogonal
decomposition and dynamic mode decomposition of jet in channel cross-
flow,’’ Nucl. Eng. Des., vol. 344, pp. 54–68, Apr. 2019. HAZEM H. REFAI (Member, IEEE) is currently
[34] P. J. Schmid, ‘‘Dynamic mode decomposition of numerical and experimen- the Williams Chair of Telecommunication and
tal data,’’ J. Fluid Mech., vol. 656, pp. 5–28, Jan. 2010. Networking with the School of Electrical and
[35] A. A. Kaptanoglu, K. D. Morgan, C. J. Hansen, and S. L. Brunton, Computer Engineering (ECE) Telecommunica-
‘‘Characterizing magnetized plasmas with dynamic mode decomposition,’’ tion Program, The University of Oklahoma-Tulsa
Phys. Plasmas, vol. 27, no. 3, Mar. 2020, Art. no. 032108. (OU-Tulsa), Tulsa, OK, USA, where he is also the
[36] K. Fujii and Y. Kawahara, ‘‘Dynamic mode decomposition in vector- Founder and the Director of the Wireless Elec-
valued reproducing kernel Hilbert spaces for extracting dynamical struc-
tromagnetic Compliance and Design (WECAD)
ture among observables,’’ Neural Netw., vol. 117, pp. 94–103, Sep. 2019.
Center. WECAD’s mission is to conduct basic and
[37] S. Zhou, G. Lin, Q. Qian, and C. Xu, ‘‘Binary classification of floor
applied research examining wireless coexistence.
vibrations for human activity detection based on dynamic mode decom-
position,’’ Neurocomputing, vol. 432, pp. 227–239, Apr. 2021. He has published more than 190 refereed papers for national and inter-
[38] O. Akbunar, A. R. Ekti, G. K. Kurt, and A. Görc. Standard- national conferences and journal articles. His research interests include
Based Wireless Signals. Accessed: Feb. 2022. [Online]. Available: the development of optical wireless communication, physical and medium
[Link] access control layers to enhance wireless coexistence, and cognitive radios
[39] GSM Technical Specification. Accessed: Jan. 2022. [Online]. Available: and networks. He is the past IEEE ComSoc Tulsa Chapter President and
[Link] served as the Organization’s North American Distinguished Lecturer Tour
[40] LTE Frame Structure Made Simple. Accessed on: Jan. 2022. [Online]. Coordinator.
Available: [Link]

VOLUME 11, 2023 18441