A NEW PROPOSED ADAPTIVE COGNITIVE RADIO DETECTION
SYSTEM BASED ON MLP NEURAL NETWORK FOR
DIFFERENT MODULATION SCHEMES
Hadi T. Ziboon
1and Ahmed A. Thabit
2 1Univesrity of Technology, Department of Electrical Engineering, Baghdad, Iraq
2
Al-Rafidain University, Department of Communications Computers, Baghdad, Iraq
E-Mail: [email protected]
ABSTRACT
The frequency spectrum of the electromagnetic radio is crowded day by day due to the expansion in wireless devices and applications. It has been additionally found that the allocated spectrum is underutilized as a result of the static portion of the spectrum. Cognitive radio (CR) allows for usage of licensed frequency bands by unlicensed users. These unlicensed users need to monitor the spectrum continuously to avoid possible interference with the licensed users. Spectrum usage regulations not permitting unlicensed users to authorized in a licensed spectrum. It has been seen that the whole licensed spectrum is not used at all places constantly. An unlicensed user can exploit advantage of such a situation to communicate thereby increasing spectrum efficiency. This is the fundamental thought behind Cognitive Radio. Demand for spectrum is expected to increasing rapidly and it would get in future. As more and more technologies are moving towards fully wireless, demand for spectrum is enhancing. In this paper, a proposed adaptive CR detection system is designed based on statistical features using neural network (multi layer perceptron) for intermediate frequency stage. Matlab simulation program is used to obtain the results. In order to evaluate the performance of the proposed CR detection systems, different modulated digital signals (2FSK, 4FSK, BPSK, QPSK, 8PSK, 4QAM, 16QAM, 64QAM and 256QAM) are generated at low SNR values. Multilayer perceptron is better than single layer due to their speed and nonlinearity solving problem. This is clearly seen in the obtained results such as Pd =100% for Pf=0.1 at SNR= -16dB, also Pd=90% at SNR= -40 dB and 95% at SNR= - 24dB with sensing time −4 𝑒𝑐 at AWGN noisy channel.
Index terms: cognitive radio, MLP, statistical features.
INTRODUCTION
The usage of radio spectrum resources and the regulation of radio emissions are coordinated by national regulatory bodies like the Federal Communications Commission (FCC). The FCC assigns spectrum to licensed users, also known as primary users (PU), on a long-term basis for large geographical regions and According to the study of FCC of the spectrum utilization shows that licensed spectrum with utilization range is from 15% to 85% for the bands below 3 GHz [1].
The inefficient usage of the limited spectrum necessitates the development of dynamic spectrum access techniques, where users who have no spectrum licenses, also known as secondary users (SU), are allowed to use the temporarily unused licensed spectrum. In recent years, the FCC has been considering more flexible and comprehensive uses of the available spectrum, through the use of cognitive radiotechnology [1, 2].
CR is the key enabling technology that enables next generation communication networks, also known as dynamic spectrum access (DSA) networks, to utilize the spectrum more efficiently in an opportunistic fashion without interfering with the primary users. Figure (1) represents the basic idea behind the CR where it searches on the used and unused channels and care do not interface with the used channels. CR can change its transmitter parameters according to the interactions with the environment in which it operates. It differs from conventional radio devices in that a cognitive radio can equip users with cognitive capability and reconfigurability.
Figure-1. Idea behind the CR.
There are different methods proposed for spectrum sensing but they have limitations and requirements in real environment. Fading or increased distance between transmitter and receiver makes the primary signal power near to the ground. Very low Signal to Noise Ratio (SNR) e.g. -20dB makes detection harder, especially there is no prior information for the noise power and that makes all above situations more difficult in real environment. There are various spectrum sensing methods available such as Energy Detection, Matched Filtering, and Cyclostationary Detection etc. Energy Detection (ED) is a widely used method because of its low complexity and it does not need information of the primary signal. ED is vulnerable to noise uncertainty and performs poor as compared to other methods at low SNR [3, 4].
Cognitive radio and SDR
CR including software-defined radio (SDR) as enabling technology was proposed in [5, 6, 7] to realize adaptable and effective usage of spectrum. The term
cognitive radio is gotten from “cognition”. Cognition is a
term referring to the mental processes involved in picking up information and comprehension, including thinking, knowing, recollecting, judging and problem-solving. These higher-level functions of the brain include language, imagination, perception, and planning [8].
CR may have significant impacts on both technology and regulation of use of spectrum leading to a revolution in wireless communication overcoming existing regulatory barriers [9].
CR came from number of technologies such as the improvement of digital signal processing (DSP), math tools and source coding of data, voice and image etc. CR has turned into promising strategy to solve the spectrum scarcity problem for supporting evolving wireless services and applications. In CR systems, the unlicensed users can use the licensed frequencies while the PU is not dynamic [10, 11].
CR was introduced at 1999 by Mitola [12]. It is supposed to change their operating band, if the currently used band becomes too occupied or the PU takes the band into use. The most important feature of the cognitive
radios is the capacity to sense the spectrum, whether to take a certain band into use or not.
It can adapt to their environment via varying their transmitter factors to different signaling systems. Depending upon the network and cooperation with other cognitive devices, they can trade information about their location and environment. CR can cooperate with other cognitive radios and offer information between each other. A common radio communication system is implemented in hardware. Figure (2) represents the detection cycle of CR that shows the steps of detection process.
Figure-2. Detection cycle of CR.
A software defined radio communication system places much of the signal manipulation into the digital domain where the computer is then able to process the data [1]. Figure-3 depicts the evolution from HW to SDR to SR to "Adaptive Intelligent-Software Radio" (AI-SR). Over time the number of system components performed in software is increasing [11, 14, and 15].
Figure-3. Evolution of SD.
Advantages of SDR
One of the basic advantages of SDR is the ability to receive and transmit various modulation methods using a common set of hardware. The ability to alter functionality by downloading and running new software at will. The possibility of adaptively choosing an operating frequency and a mode best suited for prevailing
Radio Environment
Sensing
Real time wideband monitoring
Analysis
Rapid characterization of environment
Adaptation
Transition to new operating parameters
Reasoning
conditions. The opportunity to recognize and avoid interference with other communications channels using programmable digital devices to perform the signal processing to transmit and receive base band information at radio frequency. SDR 3G cellular base stations are now possible and can match the efficiency performance of their HDR. Less infrastructure, less maintenance, easier deployment with reduced ownership costs. Support of multiple standards through multimode, multiband radio capabilities [17, 18].
Multi layer perceptron weighted neural network identifier
The most popular recognized feed forward neural network model is the multilayer Perceptron (MLP). This network consists of input layer, one or more hidden layers and output layer. The number of input and output variables determines the number of nodes in the input and output layer respectively [16]. The structure of the MLP neural network is shown in Figure-4. Any identification consists of learning phase and recall phase. The purpose of learning phase is to update the weights value of the network based on the learning algorithm [17].
Figure-4. MLP neural network structure.
The weights of the neural network are updated in a supervised mode using the most common algorithm known as the Back Propagation (BP) algorithm and according to Equation.(1) [17].
+ = − 𝜀𝜕𝐸𝑓
𝜕𝑤 (1)
Where is the weight between neuron i and neuron j, 𝜀 is the learning rate, and 𝐸𝑓 is the error function. In the learning phase, the output of the feed forward neural network is computed for each input training pattern. The error between the computed output and desired output is used to update the weight of the network by back propagation algorithm [17].
MLP is able to solve nonlinearly separable problems; a number of neurons are connected in layers to build a multilayer perceptron. Each of the perceptron is used to identify small linearly separable sections of the
inputs. Outputs of the perceptron are gathered into another perceptron to produce the final output.
Using hard-limiting (step) function for producing the output will prevents information on the real inputs flowing on to inner neurons. To solve this problem, the step function is changed with a continuous function- usually the sigmoid function.
Proposed adaptive CR detection system based on MLP
Feature selection is the most important step that should be takes in account in designing a CR. Conventional detection methods are based on instantaneous features such as energy. These methods are suffers from noise uncertainty. These systems are failed in the detection at low SNR values. Therefore, a new CR system using statistical features based on MLP is proposed. Moment and cumulant are suggested in this system as a signal feature. Two features are used in this system; they are (C11, C20). These features are selected according to their ability to separate or distinguish between signal and noise as clearly explained in the matlab simulation Figures (5 and 6, 7) that shows the ability of specified cumulant to distinguish between the noise and the signals. Therefore, these features are candidate to provide the best response. Figures (5 and 6) show the separation between the signal and noise for C20 and C11 respectively. Then Figure-7 represents the combinational of C20 and C11 by the relation 𝐶 −𝐶
𝐶 and
Figure-8 represents the Zoomed of Figure-7. It’s noted from Figure-8 that the decision is noise when [(C20-C11)/C20] is =< 0.05, else, the output is signal. This combinational relation is obtained according to designing requirements i.e. higher Pd at low SNR.
Figure-5. SNR versus C20.
-40 -35 -30 -25 -20 -15 -10 -5 0 5 10
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
SNR (dB)
|C
20
|
Figure-6. SNR versus C11.
Figure-7. SNR versus [(C20-C11)/C20].
Figure-8. Zoomed of figure-7.
Equations 2, 3, 4, 5 , 6 and 7 are used in the design
̅ = 𝐸{ } (2)
the received signal is complex baseband envelope, therefore the general expression for the + th moment for a complex random variable is [16]
𝐸 = 𝐸{ ∙ ∗ }
where ∗ denotes complex conjugate.
The + th moment of the received digital modulation signal is evaluated from its N samples by calculating the numerical mean of these samples after raising each sample to the power according to the moment order as in Equation.(4):
𝐸′, =𝑁∑𝑁= ∙ ∗ (4)
However, + th central moment is defined as:
𝐸, = 𝐸{ − ̅ ∙ − ̅ ∗ } (5)
𝐶, = 𝐸,
𝐶, = 𝐸,
These values of cumulants and moments are trained using MLP neural network. The system deals with 9 types of modulation levels. Due to the periodicity of the cyclostationary modulated signals, the system will be adaptive to detect each one of these signals and identify between the noise and the signal by CR principle. This is because of the using of more than one feature (cumulant and moment) in the same time and using more than one threshold instead of single decision region.
This design consists of input layer, two hidden layers and output layer. The number of input and output variables determines the number of nodes in the input and output layer respectively. Figure-9 represents the multilayer perceptron for the proposed system. It consists of one input, two hidden layers and one output. Each hidden layer consists of three nodes. The output consists of two nodes which represents either noise or signal.
Figure-9. Structure of MLP.
The table of designed system is shown below in Table-1 that shows the parameters that used in this design and their values.
-40 -35 -30 -25 -20 -15 -10 -5 0 5 10
0 5 10 15 20 25 30
SNR (dB)
|C
11
|
Noise Signal
-40 -35 -30 -25 -20 -15 -10 -5 0 5 10
0 50 100 150 200 250 300 350
SNR (dB)
[C
2
0
-C
1
1
]/
C
2
0
Noise Signal
-40 -35 -30 -25 -20 -15 -10 -5 0 5 10
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
SNR (dB)
[C
2
0
-C
1
1
]/
C
2
0
Table-1. Design parameters of CR based on MLP.
Parameters Values
Pf 0.1
Number of iterations 10
Message length 1000,…8000
SNR ranges -40dB → 0dB
Modulation type
2FSK,4FSK,BPSK, QPSK,8PSK,4QAM,16QAM,
64QAM,256QAM Carrier frequency 20 MHz
Sampling frequency 50 MHz
Sample per symbol 10
Hidden layer and nodes
2 hidden layer with 3 node in each layer
Simulation results
The performance of the MLP neural network for multi modulation system is shown in Figure-10 and Figure-11 respectively. As illustrated in these figures, the detection accuracy of received noisy channel is shown below for different message lengths at Pf=0.1 for 10 iterations.
Figure-10. SNR verse Pd at message length of 3000 and Pf=0.1.
Figure-11. Pd vs. SNR for different Msg. lengths for MLP systems.
It is seen that when the message length increase, the detection probability increases also. It is an important to plot the ROC for MLP detection system as shown in Figure-12. From this figure, one can check the Pf value needed to get desired Pd.
Figure-12. ROC of MLP detection system for message length=2000.
also, the missed detection probability (Pm)is ploted in Figure-13 to show the occurred error during the detection. This error occurred when the detection is noise but it is a signal.
-40 -36 -32 -28 -24 -20 -16 -12 -8 -4 0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
SNR(dB)
P
ro
b
a
b
ili
ty
o
f
D
e
te
c
ti
o
n
(
%
)
SNR vs. Pd, numtrials=10, Pf=0.1
msg. length=3000
-40 -36 -32 -28 -24 -20 -16 -12 -8 -4 0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
SNR(dB)
P
ro
b
a
b
ili
ty
o
f
D
e
te
c
ti
o
n
(
%
)
SNR vs. Pd, numtrials=10, Pf=0.1
msg. length=1000 msg. length=2000 msg. length=3000 msg. length=4000 Msg. length=5000 Msg. length=6000
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
pf
P
ro
b
a
b
il
it
y
o
f
D
e
te
c
ti
o
n
(
%
)
Pf vs. Pd at different SNR, numtrials=10, Pf=0.1, msglength=2000
Figure-13. SNR versus Pm for MLP for different message lengths.
To check the sensing time of the proposed MLP detection system, Figure-14 show the sensing time versus Pd.
Figure-14. Ts versus Pd, for MLP for Pf=0.1 and 10 iteration.
It is noted that the designed system needs
∗ −4 second to obain 100% detection for SNR=-4dB
and needed 1 msec for 95% at -10dB. The rest cases is shown in the Figure. Different message lengths are shown below in Figure-15.
Figure-15. N-samples versus Pd, for MLP for Pf=0.1 and 10 iteration.
Two features are suggested for CR detector, namely C11 and C20 at carrier frequency of 20 MHz with different multilevel modulation schemes (2FSK, 4FSK, BPSK, QPSK, 8PSK, 4QAM, 16QAM, 64QAM and 256QAM). The output layer consists of two neurons, one for each CR conditions. Tan-sigmoid is the activation function of the hidden layers, while the linear transfer function is used for output layer. A Back Propagation algorithm is implemented for the learning of this network. The Mean Square Error (MSE) performance reached
.× 9 − 4 as best performance at 31 epochs as shown
in Figure-16. The probability of detection is 100 % when the network is tested by the same training set.
Figure-16. Training performance of MLP identifier.
CONCLUSIONS
It’s seen clearly from the simulation results of the
system, that the proposed system is more powerful than traditional detection systems. This is due to the statistical features that used to train MLP in this design. Numerical results shows that the proposed approach can guarantee a reliable sensing while enhancing the spectrum utilization greatly such as Pd =100% for Pf=0.1 at SNR= -16dB for
-40 -36 -32 -28 -24 -20 -16 -12 -8 -4 0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 SNR(dB) P ro b a b il it y o f M is s e d D e te c ti o n ( % )
SNR VS. Pm, numtrials=10 Pf=0.1
N=1000 N=2000 N=3000 N=4000 N=5000 N=6000
1.25 2.5 3.75 5 6.25 7.5 8.75 10
x 10-4
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 Ts P ro b a b il it y o f D e te c ti o n ( % )
Ts versus Pd at different SNR values
SNR=-20 dB SNR=-14 dB SNR=-12 dB SNR=-10 dB SNR=-8 dB SNR=-6 dB SNR=-4 dB
600 1200 1800 2400 3000 3600 4200 4800 5400 6000
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 N-samples P ro b a b il it y o f D e te c ti o n ( % )
N-Samples VS. Pd , variable SNR, numtrials=10 and Pf=0.1
SNR=-40dB SNR=-36dB SNR=-32dB SNR=-28dB SNR=-24dB SNR=-20dB SNR=-16dB SNR=-12dB SNR=-8dB SNR=-4dB SNR=0dB
0 5 10 15 20 25 30
10-10 10-5 100
Best Training Performance is 2.397e-14 at epoch 31
message length =6000 samples, also Pd=90% at SNR= -40 dB and 95% at SNR= - 24dB.
REFERENCES
[1] J. ma, g. ye li, and b. h. juang. 2009. Invited paper: signal processing in cognitive radio. Journal proceedings of the IEEE,97(5): 805-823, 2009. IEEE doi:10.1109/jproc.2009.2015707.
[2] B. wang and k. j. ray liu. 2011. Advances in cognitive radio networks: A survey. IEEE journal of selected topics in signal processing. 5(1): 5-23.
[3] Ahmed a. thabit, hadi. t. ziboon. 2014. Improving probability of detection using cfar and adaptive threshold for cognitive radi (cr). International journal of engineering and advanced technology (ijeat). 4(5).
[4] A. mate, k.h. lee, and i-tai lu. 2011. Spectrum sensing based on time covariance matrix using gnu radio and usrp for cognitive radio. IEEE long island systems, applications and technology conference. pp. 1-6. IEEE doi:10.1109/lisat.2011.5784217.
[5] Venkata r. reddy. 2011. Resource allocation for ofdm-based cognitive radio systems. Department of electronics & communication engineering national institute of technology Rourkela.
[6] Prashob r nair, a. p. v., anoop, kumar Krishna. 2010. An adaptive threshold based energy detector for spectrum sensing in cognitive radios at low snr. IEEE.
[7] James o. neel. 2006. Analysis and design of cognitive radio networks. Ph.D thesis, and distributed radio resource management algorithms
[8] Simon haykin. 2005. Cognitive radio: brain-empowered wireless communications. IEEE journal on selected areas in communications. 23(2).
[9] Aravind puttupu. 2013. Improved double threshold energy detection in cognitive radio networks. M.sc. thesis, department of electrical engineering national institute of technology, rourkela.
[10]Ashish bagwari, geetam s. tomar and shekhar verma. 2014. Cooperative spectrum sensing based on two-stage detectors with multiple energy detectors and adaptive double threshold in cognitive radio networks. Canadian Journal of electrical and computer engineering. 36(4).
[11]Danijela c. artem, t. and robert w. brodersen. 2007. Experimental study of spectrum sensing based on energy detection and network cooperation.
[12]Joseph mitola iii. 2000. Cognitive radio an integrated agent architecture for software defined radio. Royal institute of technology (kth), teleinformatics Ph.D dissertation.
[13]Daniela m. m. plataa, ángel g. a. reátiga. 2012. Evaluation of energy detection for spectrum sensing based on the dynamic selection of detection-threshold. International meeting of electrical engineering research, eniinvie.
[14]Anurag bansal 1, ms. rita mahajan 2. 2009. Building cognitive radio system using matlab. International journal of electronics and computer science engineering.
[15]Mithun chakraborty1, r.bera2, p.pradhan2, r.pradhan2, s.sunar2. 2010. Spectrum sensing and spectrum shifting implementation in a cognitive radio based IEEE 802.22 wireless regional area network’’ international journal on computer science and engineering. 02(04).
[16]M. azarbad, s. hakimi and a. ebrahimzad. 2012. Automatic recognition of digital communication signal. International journal of energy, information and communications.3(4): 21-34.
[17]Ivan A. Hashim. 2016. Automatic Digital Modulation Identification for Software Defined Radio Based on FPGA. Ph.D Dissertation.
