SDR-based Intelligent Cooperative Spectrum Sensing for Cognitive Radio Systems

In this paper


H I G H L I G H T S A B S T R A C T
• The real-time sensing performance of the new intelligent cooperative spectrum sensing based on the denoised mixed feature method paired with K-Medoids is verified. • Two RTL-SDR hardware receivers, a host laptop computer, and a USRP-SDR hardware transmitter are used. • The theoretical and practical sensing performance validity of the developed scheme is confirmed based on the SDR platform.
In this paper, the software and hardware of a software-defined radio (SDR) platform are used to implement and verify the blind real-time sensing act of intelligent collaborative spectrum sensing based on a new theoretical formula for constructing denoised mixed features named MSKU3 and paired with an unsupervised machine learning K-Medoids algorithm. Two low-cost RTL-SDR dongle hardware receivers are used as two cooperative unlicensed secondary users to capture the radio frequency of a licensed primary user channel. A host personal computer is used as a fusion center to run GNU-Radio software signal processing blocks to implement the developed method, and a single Universal Software Radio Peripheral (USRP) N210 hardware transmitter based on FPGA is used to take up unoccupied desired radio frequency bandwidth. Two scenarios of signal-to-noise ratio levels have been adopted to verify and test the sensing performance of the developed system. The first one occurs when unlicensed secondary users have equal signal-to-noise ratio values. The second occurs when unlicensed secondary users have different signal-to-noise ratio values since each secondary user has their location. The experimental results of detecting action in terms of the probability of detection for the proposed method show that the theoretical and practical results are very close to each other.

Introduction
The enormous increase in the number of mobile devices, combined with the stationary handling of frequency bands, has resulted in a lack of accessible available bandwidth [1]. According to the findings of a study conducted by the Federal Communication Commission (FCC), the percentage of licensing spectrum that is being continuously utilized by the licensed Primary User (PU) is lower than thirty percent [2]. According to FCC reports, there are bandwidth vacancies, particularly in the TV band, sometimes known as TV white space (TVWS) [3]. Cognitive Radio (CR) is a telecommunication technique employed to address the issue of bandwidth shortages [4]. To exploit the bandwidth usage, in CR, an unlicensed Secondary User (SU) utilizes the licensed bandwidth of PU [5]. To avoid interference with PU bandwidth, Spectrum Sensing (SS) is a crucial task in CR networks that keep an eye on the licensing frequency band to identify the status of the PU and distinguish the locations of bandwidth holes [6].
In the academic literature, there are a variety of traditional narrowband SS feature detection algorithms that have been proposed. These algorithms are used to determine the state (occupied or unoccupied) of a PU in a certain frequency band. Some examples of these techniques include: energy detection (ED) [7], Cyclostationary Features (CFs) Detection [8], eigenvalue of the covariance matrix (CM) of random matrix theory (RMT)-based detection methods [9][10][11], and information geometry theory (IGT)-based methods [12]. Consequently, previous classical SS techniques are unable to adapt to the dynamic nature of the actual channel because they rely on a fixed detection threshold [13]. Lately, many Cooperative Spectrum Sensing (CSS) schemes based on machine learning (ML) and deep learning (DL) approaches combined with classical features methods of RMT have been proposed. They extract energy features (EFs) of RMT for sensing samples of SU, which are based on second-order moments, to train ML algorithms [14][15][16][17]. Unfortunately, all of the ML and DL-based CSS approaches that are based on EFs are more vulnerable to noise variations and perform poorly at low signal-to-noise ratios (SNR) [18][19][20]. As a result, previous CSS schemes had numerous issues, such as a high degree of overlapping between ML groups, improper categorization of desired features, and weak detecting performance [21].
In recent work published in [22,23], we provide a new blind intelligent CSS built on new denoised mixed features (DMFs) techniques and combined with unsupervised ML algorithms. We prove that the unsupervised ML-based CSS schemes that use denoised features (DFs) generated by high-order moments such as the Jarque-Bera (JB) formula or Skewness (Sk) and access Kurtosis (Ku-3) have less computational complexity and are more immune to noise fluctuation at low SNR. In [22], we established four novel denoised features (DFs) methods using the Jarque-Bera (JB) formula to improve the clustering of unsupervised ML algorithms (K-Means) and the sensing performance of CSS schemes. Similarly, in [23], we developed two new DMF methods using third and fourth moments such as Sk and Ku-3, respectively, to boost the classification process of unsupervised ML algorithms (K-Medoids) and enhance the detection performance of CSS schemes in terms of Probability of Detection (Pd).
To verify and implement the real-time detecting operation of the CSS scheme based on developed DMF methods, we use the Software Defined Radio (SDR) platform and select the maximum values of Sk and Ku-3 of the DMF vectors (MSKU3) [23] method as characteristic features for receiving sensed samples of SU. SDR has served as the cornerstone for the development of the SS methods. It is feasible to incorporate the SS techniques into communication equipment that is also adaptable and user-defined. SDR is a technique that implements communication capabilities with software operating on adaptable hardware [24,25]. In the literature, different studies have provided real-functioning implementations of SS methods using SDR. In the literature [26], one of the most widely used pieces of SDR hardware is the Universal Software Radio Peripheral (USRP), which is used to implement energy feature-based SS. The researchers of these publications evaluated the practical sensing performance of an EFs realization based on USRP hardware for their theoretical performance and discussed their findings. The paper [27,28]described the real-time realization of the SS method based on the covariance matrix of RMT employing USRP SDR hardware. In [29], a real-time adaptive SS method using EFs and CFs was realized by applying the USRP hardware platform and the GNU-Radio software throughout Rayleigh fading. Some of the studies on the development of SS make use of costly SDRs like USRP. However, some of the more recent work makes use of ultra-low-cost SDRs like RTL-SDR hardware. In the literature [30]. The author utilized SDR platforms such as RTL-SDR and GNU-Radio softwarebased IoT to implement real-time SS exploiting EFs. The literature [31] implements real-time sensing performance of the SS method based on EFs employing RTL-SDR as the receiver. In this work, we use GNU-Radio as a software toolkit to implement the actual-time detecting performance of the MSKU3 method-based CSS scheme [23], carrying on hardware made up of an RTL-SDR receiver and a USRP transmitter. In GNU-Radio Companion, a signal capture application is created, and samples collected are evaluated afterward to assess sensing accuracy. We explore the practical actual-time sensing act of a spectral domain of the MSKU3 method-based CSS scheme and compare it with its theoretical sensing performance. Based on our review of the relevant literature, we can confidently say that this is the first time the frequency domain MSKU3 methodbased CSS scheme has been implemented and analyzed using SDR.
Our novel contributions presented in this article are the following: Contributions [29,32].
Theoretical and practical real-time sensing performance analysis of a new denoised mixed-feature MSKU3 1) method-based CSS scheme in a real-channel environment have been performed. The characteristics of implementing an MSKU3 method-based CSS scheme using RTL-SDR hardware, 2) GNU-Radio software, and USR hardware is described in detail. The practical detecting capability of the MSKU3 approach is validated using two scenarios: first, the two 3) cooperative RTL-SDR receivers collect detected signals with the same SNR; second, each RTL-SDR receiver gathers up the signals with a different SNR since it is located in a separate location. The other sections of the paper are ordered as follows. The theory behind MSKU3 method-based cooperative spectrum sensing is described in Section 2. Section 3 provides details on the software model and hardware setup. The test results and associated analysis are provided in Section 4. The conclusion is given in Section 5.

Cooperative Spectrum Sensing Framework
In the CSS scheme of CR, there are two hypotheses, as shown in Eq. (1), that could be used to describe the signal that SU has captured [33] : Where i and j represent the indices of the number of SU and received samples in the CSS scheme, respectively. The null hypothesis, H 0 , is a channel with additive white Gaussian noise (AWGN) w i (j) signal, and no PU s(j) signal. The alternative hypothesis, H 1 , is a channel with noise w i (j)∼ (0, 2 ) signal plus PU s(j)∼ (0, 2 ) signal. Here, h i (j) is the channel gain between the PU signal and ith SU i , which is proposed here as h i (j) = 1 under no fading channel condition. Consequently, the complex samples of captured x i (j) signal which are detected by each SU, X= [ x 1 , x 2,…, x M ] T , can be described by the following Eq. (2) [33]: Where ∈ ℂ × , and T stands to a transposition operator. Two commonly used parameters to assess the detecting performance of CSS schemes are P d and the probability of false alarm (P f ). P D is the possibility of choosing which frequency to active when a signal is already being sent out by the PU as written in Eq. (3) [34].
P F is the possibility of deciding the frequency to be active when in actuality it is not, as written in Eq. (4) [34].
According to the IEEE 802.22 standard, P D should be equal to or greater than 0.9 and P F should be equal to or less than 0.1 as small as possible at SNR equal to -20dB so that underused wireless channels can be avoided [35].

Denoised Mixed Feature method -based Cooperative Spectrum Sensing
The theoretical analysis of a novel DMFs method-based CSS scheme is introduced in [23]. As shown in Figure 1 a new CSS scheme based on the DMF method MSKU3 consists of one PU signal on/off licensed channel, several SUs, and one fusion center (FC) unit. All SUs in the developed scheme is responsible for sensing the status of the PU signal if it is active or idle on the licensed channel. Then, collecting an adequate number of sensed samples of PU signal by each SU i to construct the vector. After that, in the "Pre-Features Extraction" section, the formula of 3 rd and 4 th high-order moments Sk and Ku-3, respectively, are employed to calculate the statistics of the collected vector x i (j) for each SU according to the following Eq. (5) and Eq. (6) [36] : Ku (x i )-3 of each SU i node are combined by a new formula as written in Eq. (7) to generate a new DMF i for each SU i [23]. The 438 Here, R( x − 3) represents the DMF i of ith SU i which is transmitted into the FC unit to construct the DMF-vector as expressed in Eq. (8) [23] : In the FC unit, the following newly developed formula, Eq. (9), is applied to Eq. (8) to obtain the maximum statistical value of the DMF-vector [23] : Here, the capital symbol F is an abbreviation of the word "Feature" and the subscript MSKU3 refers to the type of denoising mixed feature method that is used. According to the new DMF method, a training set of feature vector F � , F � = �F MSKU3 1 , F MSKU3 2 , … , F MSKU3 � are employed to teach the unsupervised ML KMD approach to establish a classifier based on (H0, H1) conditions of PU signal.

The SDR Hardware and Software Requirements
In this part, the real-time perception performance of a new intelligent CSS based on a novel MSKU3 method is confirmed, and the key parts that are used to test the real-time sensing performance of the developed scheme are described.

The SDR Hardware Specifications
The important SDR hardware parts that are chosen and used to design and test the real-time sensing performance of the proposed scheme (MSKU3-based CSS) are shown in Figure 2. They are coming with several pieces of hardware, such as the RTL-SDR USB dongle, the Universal Software Radio Peripheral (USRP®) from Ettus Research, and a laptop personal computer (PC) running the GNU-Radio Companion (GRC) software. As can be seen in Figure2 (a), the omni-directional antenna that is included with the RTL-SDR USB dongle functions as a wireless receiver. This receiver makes use of the Realtek RTL 2832U chip, which has a radio frequency (RF) range of 24 MHz to 1,766 MHz, a bandwidth of 3.2 MHz, and an analog-to-digital converter (ADC) resolution of 8 bits. The RTL-SDR hardware is capable of receiving FM radio, UHF band signals, ISM signals, GSM, 3G, and LTE mobile radio, GPS, and satellite signals, as well as any information that the designer can (lawfully) send.

Universal Software Radio Peripheral hardware
On the other hand, as can be seen in Figure2 (c), Ettus Research's USRP® hardware family consists of configurable FPGA with transmit and receive (Tx and Rx) SDR capabilities. One of the members of the USRP® family is the USRP® N210. The USRP N210 is capable of processing signals with both a large bandwidth and a wide dynamic range frequency. The USRP N210 is made with a Xilinx Spartan® 3A-DSP 3400 FPGA, 100 MS/s dual ADC, 400 MS/s dual DAC, 6GHz bandwidth, and a Gigabit Ethernet connection to make it easier for information to go to and from the PC.

Personal computer
Furthermore, as shown in Figure 2 (b), a PC, which is equipped with an Intel Core i5-8250U, 16GB of RAM, and 1TB of hard drive space, is used to run GRC software to program the SDR boards to act as a specific type of radio and to store and display information.

The SDR Software Requirements
Besides that, the GRC serves as a framework for developing the software blocks of receivers and transmitters of the developed scheme. In addition to make it possible to develop new function blocks, GRC makes it feasible to handle 439 discontinuous signals by mixing a variety of functional blocks [37]. Figure 3 displays a snapshot of the experimental equipment.

Experimental Setup
On the receiving side, one of two RTL-SDR receivers is allocated to each SU and connected to the PC through a USB port in order to accomplish the sensing performance of the developed method. In our experiment, RTL-SDRs work together to sample and digitize PU's RF signals that are sent between 25 MHz and 1.75 [37,38] Here, the PC acts as a ring link between the transmitter side and the receiver side and works as an FC unit. The PC is using a virtual machine (VMware® Workstation 16 Pro) to run GRC software version 3.8.1.0 based on the operating system (Ubuntu 20.04 LTS). The GRC is software that provides input/output blocks and enables the implementation of the DSP blocks and SDR algorithms of the proposed method, which is presented in experiment [39]. On the transmitting side, a single USRP N210-SDR transmitter is connected to a PC via Gigabit Ethernet. It is used to take up an empty PU RF channel as decided by the FC unit. Figure 4 shows the experimental setup of the GNU radio block diagram software for the entire developed scheme, which includes the transmitting side, the new denoised mixed feature method, and the receiving side. Each RTL-SDR Source GNU Radio block provides the python software codes to interface between GRC and RTL-SDR hardware. The GNU Radio RTL-SDR Source block software is used to set up important tuning parameters like carrier frequency fc and sample rate fs (bandwidth) so that the antenna and RTL-SDR dongle can pick up the desired PU RF signals from the actual channel. The fc is set to any frequency in the coverage range of the RTL-SDR antenna by the "Ch0" parameter, and the "Sample Rate" parameter sets the fs to 2 MHz.The multiplexer block controls the connection of two RTL-SDR hardware sources between the training and sensing stages. Once the training process is done, the sensing stage is activated to accomplish sensing operation, and vice versa. The training and sensing blocks are equivalent. They contain the essential Python codes to perform all the theoretical equations of the developed scheme. Following the KMD algorithm, the optimal classifier block based on two optimum medians of PU conditions is constructed (H0, H1). The role of the classifier block is to enable or disable the USRP-SDR hardware transmitter according to the PU RF channel. If a PU RF signal is present (H1), the classifier disables transmission of USRP-SDR hardware. If there is no PU RF signal (H0), the classifier lets the USRP-SDR transmitter use an empty channel. On the receiver side, the USRP-SDR Sink block provides the FPGA software codes to interface between GRC and the USRP N210 -SDR hardware transmitter. The USRP N210 -SDR hardware transmitter employed to occupy an unoccupied PU RF channel.

Results and Discussion
In this part, the practical sensing performance of the developed method-based CSS system in a real-channel environment is confirmed. In addition, a comparison between the practical and theoretical measurements of the developed method (MSKU3)based CSS scheme with different SNR conditions has been performed.

Performance Evaluation of The MSKU3-Based CSS Scheme in The Real Channel
To further verify the effectiveness of the real-time sensing performance of the newly developed DMF method (MSKU3)based CSS scheme in a more challenging wireless communication environment, the indoor real-channel environment has been selected. The indoor channel measurements are performed within the lab room on the middle floor of the communication research center in the Ministry of Science and Technology in Baghdad, Iraq (latitude 33° 16' 49" North, longitude 44° 23' 15" East, altitude 15 m), as depicted in Figure 5. The chosen site room is a challenge for our developed framework because it has walls, floors, and other surfaces that reflect electromagnetic waveforms. These surfaces introduce the complex patterns of a wireless indoor environment, such as RF interference, degradation, fluctuation, and variation against the RF of the PU signals [40]. The real-time activities of the occupied and unoccupied (holes) portions of the RF bandwidth of the PU from 25 MHz to 1.766 GHz are recorded by the spectrum analyzer device Tek RSA6100A as shown in Figure 6. Then, during sensing operation, the tuning frequency fc of both RTL-SDR receivers is set to any frequency between 24 MHz and 1,766 MHz, which is the RF range that the RTL-SDR device bandwidth can cover, and the sample rate fs is set to 2 MHz, while the USRP N210-SDR hardware transmitter utilizes the frequency fc of the unoccupied channel (hole) to broadcast a test signal, as illustrated in Figure 7. The real recorded measurements of RF ranging from 25 MHz to 1.6777 GHz for the indoor site channel when the noise floor is about -80 dBm are presented in Table 1. It contains the probability of detection (Pd) of the sensed frequency fc for each active PU signal and hole.
Then, based on the results of Pd in Table 1, the average probability of detecting the MSKU3-based CSS scheme in an indoor channel is 0.96 dBm. As a result, the practical measurements support the theoretical aspect of the developed method and confirm its sensing capability in an actual environment.

Comparison of The Theoretical and Practical Measurements of The MSKU3 Method
This part compares the practical and theoretical measurements and the extent of their affinity for the developed MSKU3 method-based CSS scheme in a real environment. Figure 8 and Table 2 show the all theoretical and practical experimental findings of the receiver operation characteristics (ROC) curves for the MSKU3-based CSS method in parameters of Pd and Pf when using RTL-SDR receivers at various SNR values. As can be seen from Figure 8 (a) and the second column of Table 2, when the practical components are adjusted to N = 1000 samples, M = 2 SUs, Pf = 0.1, and at SNR = [-16dB, -16dB], there is a convergence between the theoretical and practical performances of CSS based on the MSKU3 method in terms of Pd, which is equal to 0.965 and 0.930, respectively. Also, in Figure 8(b) and the third column of Table 2, when N = 1000 samples, M = 2 SUs, Pf = 0.1, and at SNR = [-20dB, -20dB], the theoretical and practical performances of CSS based on the MSKU3 method are converged in terms of Pd and equal to 0.942 and 0.900, respectively. Furthermore, Figure 8(c) and the four columns of Table 2 confirm the convergence between the theoretical and practical performances of CSS based on the MSKU3 method for the same settings and at SNR = [-16 dB, -20 dB], where Pd is equal to 0.911 and 0.850, respectively. Finally, Figure 8 (d) and the five columns of Table 2 prove the convergence between the theoretical and practical performances of CSS based on the MSKU3 method for the same settings and at SNR = [-20 dB, -24 dB], where Pd is equal to 0.900 and 0.845, respectively. Even though the method that is developed is tailored to the noise channel, experimental and theoretical results match up. This is because of the idea of learning, which is achieved by using the unsupervised ML KMD.This indicates that the practical results that have been obtained under uncertain noise conditions and in real channel environments prove the validity of the theoretical methods that have been proposed for the CSS system.

Comparison of The Developed Method With Some Other Methods
As shown in Table 3, the key difference between our developed work [23] and other works [26,27,29] is that we use highorder moments (Sk and Ku-3) to build the DMF method and the unsupervised ML algorithm (KMD) for the trained classifier, whereas other works use features based on second-order moments such as CF, EF, and EF of RMT to derive a fixed threshold. Except work [31], which employed supervised ML algorithms such as Artificial Neural Networks (ANN), Support Vector Machines (SVM), decision trees (TREE), and K-Nearest Neighbors (KNN). On the other hand, in comparison to supervised ML algorithms, the unsupervised ML algorithm does not need previous knowledge about the actual PU channel. This would reduce the complexity of the proposed method and enhance the blind detection capability of the developed system [23]. In addition, deriving the threshold from the second-order moment is fixed, has a lot of computational complexity, and makes it hard to adapt to how the real channel of the PU signal changes over time [26,27,29]. Therefore, our proposed method has more benefits when it uses an unsupervised ML algorithm. For example, it avoids the computational complexity of fixed-threshold derivation and instead builds the classifier through the clustering process [23]. Moreover, the classifier construction is more prone to noise uncertainty, especially at low SNR, when using EF and RMT; on the contrary, when we use the classifier built on high-order moments (Sk and Ku-3), the proposed classifier is more immune to noise uncertainty.

Conclusion
In this study, the hardware and software of the SDR platform have been used to verify the real-time sensing performance of a new MSKU3 method-based CSS scheme. Various SDR hardware and software are used to implement the developed system, such as RTL-SDR hardware as a receiver, GNU-Radio software to implement digital signal processing blocks of the proposed method, and USRP N210 hardware as a transmitter. The experimental results of sensing performance in terms of the probability of detection based on the SDR platform show the effectiveness of the proposed method in the current environment.