Spectrum Sensing Using Software Defined Radio for Cognitive Radio Networks: A Survey

Cognitive radio (CR) network has emerged as a potential solution to the under-utilization problem of the allocated radio spectrum, where spectrum sensing (SS) plays a key role to enable the coexistence between primary and secondary users. It has attracted research interests, and several works have been reported in the literature. Nevertheless, the assumptions and simplifications introduced during the modeling of the communication system often yield misleading conclusions each time relevant aspects of their implementation on a testbed are omitted. Hence, prototypes are built to study their behaviour under real-world conditions, therefore software defined radio (SDR) has emerged as an ideal vehicle to allow researchers to experiment with prototypes of these CR approaches. In this survey, we provide an overview of the latest works in CR networks related to the spectrum awareness approaches and taking into account their implementation on testbeds. These approaches are classified from a practical point of view, where a detailed review of the existing works for each category is provided. A review of the existing SDR platforms is also exposed highlighting the main components and features of current architectures employed for experimental evaluation of CR approaches. Next, the challenges to implement current spectrum awareness approaches on SDR platforms are detailed. Finally, at the light of these reviews, research challenges and open issues are identified for future research directions.


I. INTRODUCTION
With the increasing demand of broadband wireless spectrum due to the incorporation of wireless devices requiring higher data rates, the allocation of spectrum has been carried out inefficiently, and its provision has been limited by the segmentation of the spectrum, and the allocated frequencies of standardized wireless systems.
The associate editor coordinating the review of this manuscript and approving it for publication was Pietro Savazzi . This shortage of the spectrum has motivated the conception of CR networks as a potential candidate to perform a complete dynamic spectrum access (DSA) by exploiting the available frequency bands called spectrum holes or white spaces [1]. It incorporates various techniques enabling the coexistence of licensed and unlicensed systems over the same spectrum, where primary users (PUs), also known as licensed or incumbers users, are defined as the users who have higher priority or legacy rights for using an specific part of the spectrum, while secondary users (SUs), also called cognitive users, exploit the spectrum in such a way that they do not VOLUME 10, 2022 This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/ FIGURE 1. SDR: past, present, and future. Evolution of SDR through successive generation and its adoption as a de facto industry standard for radios [3].
cause harmful interference to the normal operation of the license PUs. Although numerous approaches have been conceived for CR networks over the last decades, and despite of the significant efforts carried out by research institutions, regulatory and industrialization bodies, the final adoption of this technology is still facing numerous challenges [2]. For that end, the research of practical solutions towards a realistic implementation has become critical for the actual system deployment.
SDR has been a supporting technology that facilitates the assessment of novel approaches under more realistic environment. It allows to implement radio communication systems by shifting a hardware design to systems where most functional components are defined in software, as conceived by Joseph Mitola III [1]. Since its conception, different cognitive radio capabilities have been implemented, so that CR devices can acquire information from their operating environments and adapt their radio parameters autonomously in order to exploit the underutilized parts of the spectrum. In this way, the feasibility and validation of the novel approaches can be assessed through exhaustive experimental evaluations that can corroborate the expected results.
During the last years, the SDR adoption has been evident allowing the markets to move from hardware radio architectures (e.g. military communications) to SDR architectures, as it is depicted in Fig. 1. The advancement of radio frequency integrated circuits (RFICs) and programmable devices (Field-programmable gate array -FPGAs, DSP, system on a chip -SoC, etc.) has enabled it to become the dominant industry standard in 4G networks, and the current FIGURE 2. 5G technology enables diverse services and applications requiring access to different spectrum bands [7]. development of the wireless communications demands the validation of CR approaches.
In cellular networks, long term evolution (LTE) developed by third generation partnership project (3GPP) has extended its usage to WiFi unlicensed bands [4], where coexistence strategies using the sensing capabilities of the LTE network have been conceived. For instance LTE introduces a licensed-assisted access (LAA) feature to leverage the use of the unlicensed spectrum. It relies on different channel access procedures based on sensing the channel before transmissions. In fact, this mechanism known as listen-before-talk (LBT) is studied and validated on a SDR platform addressing the coexistence between LTE-LAA systems and wireless local area networks (WLAN) [5]. Moreover, it is also considered for 5G services such as the ultra-reliable and low latency communication (URLLC) [6], thus the incorporation of more advanced features to the current SDR-based platforms is required for the adoption and standardization of unlicensed communication systems.
5G networks have also introduced a massive MIMO technology, where a higher number of antennas at the base station provides a new dimension for opportunistic transmissions in addition to the time-frequency dimension, i.e. the spatial dimension. Hence, spectrum sensing techniques related to this dimension have been conceived and evaluated through multi-antenna SDR platforms [8]. Furthermore, with the increasing number of applications demanding higher data rate, higher spectral resources are required. Conventional narrowband SS schemes are not enough to explore the wide frequency bands, and wideband spectrum sensing (WSS) techniques are explored [9]. It is portrayed in Fig. 2, where the access to different frequency bands from lower frequencies to higher frequencies with large bandwidths are required by several applications/services. Nevertheless it imposes several challenges at the hardware level, along with a higher complexity for its implementation on current SDR platforms, as well as new challenges for the incorporation of reconfigurable antennas to handheld transceivers.
Last but not least, IoT networks are taking part of numerous applications such as monitoring purpose, e-health applications, smart homes, agriculture, among others giving rise to a high demand on spectral resources [10], [11]. Once more numerous SS mechanisms are being adapted and evaluated addressing their challenging implementation on large-scale testbeds. Thus highlighting the urgent need for the standardization of CR SS approaches.

A. RELATED WORKS
In Table 2, we cite the works related to the implementation and evaluation of SS techniques on SDR platforms. A brief review is provided along with their limitations in relation to the present survey.
In the upper part of the Table, the works concerning SS techniques are reported. In [12], the existing spectrum occupancy models based on measurement campaigns are investigated, without addressing the detection performance based on these models. Spectrum prediction algorithms based on occupancy models are reported in [14], remaining their assessment and validation on SDR platforms. Several spectrum algorithms are surveyed in [13], and practical considerations for their implementations are provided, however these are very barely covered for an energy detector without addressing SDR implementations and experimental evaluations. More recently, the authors in [15] focus on particular aspects such as full duplex and cooperative spectrum sensing (CSS), and in a similar way in [11] the recent narrowband/wideband SS techniques are reviewed. Nevertheless, the main scope of these latest works relies mainly on the mathematical models supporting the SS metrics.
On the other hand, SDR architectures have also been surveyed to cover in general the hardware needs for radio communication [22]. The second part of Table 2 is devoted to these works. In [16], different multi-core processor architectures are explored for the increasing complexity of new generation of mobile terminals, where SS is not covered but foreseen to require higher levels of dynamism and complexity. SDR and CR introduce new classes of security threats which are considered in [17], but they are out of the scope for SS approaches. In [18], the authors highlight the employment of radio software to provide a more flexible usage of the current heterogeneous hardware architectures suitable for CR. Nevertheless, it does not cover the challenges regarding the implementation of SS approaches. Moreover, in [19], [20], and [21] the implementation of CR mechanisms are not considered. In [19], the transmitter/receiver radio chain for a particular SDR board is studied, identifying the main bottlenecks when connecting a SDR board to a host PC via Ethernet. Wyglinski et al. [20] present some case studies about the employability of SDR platforms, while some SDR enhacements are reported in [21]. Finally, a comprehensive survey in [22] concerning the architecture, state-of-theart, and challenges is reported, while covering only general requirements for SDR platforms.

B. SCOPE AND OBJECTIVES OF THE SURVEY
The increasing amount of works reported in these surveys has also propitiated a vast amount of work assessing their performance in more realistic environments with the employment of SDR platforms.
In this survey, we provide an overview of the latest developments for CR approaches related to the SS while highlighting their practical implementation aspects on current SDR platforms. The most relevant functions for the successful deployment of CR networks are identified and studied. Consequently, a classification of these approaches from this perspective is portrayed, along with a review of the experimental evaluations carried out to validate them on a SDR testbed.
Unlike, the aforementioned SDR surveys and magazine articles, current platforms are examined based on CR VOLUME 10, 2022 requirements to alleviate the deployment of this technology. It is worth highlighting that the SDR implementation of each class of SS techniques faces different challenges in hardware and software aspects, which are addressed, and discussed.
Furthermore, an overview of the development tools along with the existing SDR architectures tailored for CR networks is exposed, by describing their main analog and digital components. Finally, future research directions and SDR enhancements are provided and discussed. Hence the ultimate goal of this article is to provide a bridge between the latest researchers tailored for SS concepts, and their deployment using SDR platforms.

C. CONTRIBUTIONS OF THE SURVEY
Hence the main key contributions distinguishing our paper can be outlined as follows: • A survey of SS techniques concerning their practical implementation, i.e. taking into account hardware impairments and limitations.
• Review of SDR platforms tailored for CR approaches, main components for a completed functional CR platform are identified and detailed.
• Deployment of SS approaches on current SDR platforms, where we highlight the main challenges for a realtime prototyping.
• Open research issues remained for the conception of CR approaches, and future SDR developments.
To the best of our knowledge, it is the first time that spectrum sensing approaches are reviewed taking into account the challenges of their implementation, i.e. hardware impairments, as well as advantages and weakness of current SDR platforms. Related surveys have provided partial answers from a theoretical and practical point of view when conceiving a CR approach. On the other hand, SDR architectures have been reviewed considering general requirements for the implementation of radio protocols. In this regard, this article aims to fill this gap by providing an up-to-date survey of current efforts for the deployment and adoption of this technology.

D. ORGANIZATION OF THE SURVEY
The content of this article is organized as follows: In Section II, SS algorithms are reviewed and discussed taking into account the aspects of their implementations. The existing SDR tools are provided in Section III, where the current hardware, software and SDR platforms are reviewed. Later, Section IV exposes and discusses the identified challenges related to the SDR implementation of SS approaches. Next, research challenges and future research directions for CR are given in Section V. Finally, our conclusions are presented in Section VI. The overall organization of this article can be depicted in Fig. 3.

II. TOWARDS A PRACTICAL IMPLEMENTATION OF SPECTRUM SENSING
The scope of this survey can be depicted in Fig. 4, where PUs and SUs coexist, and a SDR platform is used to emulate a SU in order to assess its performance under more realistic conditions. We briefly review the fundamental concepts when addressing the implementation of SS approaches on SDR platforms that will be employed as background to expose and discuss the reported works in the literature.
CR is a potential candidate to exploit the white spaces, where SS is probably the most employed mechanism for acquiring information about the spectrum occupancy. It enables the SUs to detect the presence or absent of the PU over a frequency channel of interest, 1 and it is often formulated as a hypothesis test, while evaluated based on the detection and false alarm probabilities P D and P FA [23], respectively.
Each of these spectrum sensing approaches face different challenges when addressing their performance employing real measurements that can deviate from what is reported P D 1 It is worth recalling that in the sequel of the present paper, we are interested to address interweave techniques, quite often reported in the literature. and P FA . In doing so, some detection problems are formulated by incorporating practical aspects to the signal model, so that a degradation of the performance, can be avoided or at least mitigated. Some of them have been studied resorting to computer-based simulations, while others are validated through experimental evaluations.
For that end, it is important to highlight the main concerns regarding the implementation of sensing mechanisms. It can be clarified with a description of the main components in a SDR receiver. In Fig. 5, it is depicted and is composed of two main parts: a digital and a radio frequency (RF) front end. The digital front end part is basically composed of a digital down converter (DDC), sample rate conversion (decimation) and low-pass filters, digital oscillators, among others. It is in charge of the rate adaptation, and channelization to operate between a wideband multichannel digital signal and independent baseband channels. On the other hand, RF front end is composed of elements such as low noise amplifiers (LNA), mixers, variable gain amplifiers (VGA), and automatic gain controller (AGC), where it carries out analog operations such as the down conversion from the operating RF signal to VOLUME 10, 2022  the baseband signal. 2 When a RF signal is received, first it is downconverted to a baseband signal, and the analogto-digital converter (ADC) provides digital samples to the DDC. To sum up, an SDR receiver is then composed of a RF, a digital front end, and finally a baseband signal processing.
SS approaches reported in previous classifications have been examined considering missing aspects of their implementation. It has been portrayed in Fig. 6, where main hardware impairments are described. Then, mathematical models overlooking these aspects prevent from obtaining analytical expressions, and thus resorting to experimental evaluations. Hence, SS approaches are surveyed regarding: • Signal uncertainties: noise uncertainty, non-Gaussian noise, time-varying noise floor.

A. ENERGY DETECTION
An energy detector (ED) does not need prior information about the PU signals, and has the lowest computational complexity among other methods [23]. It has been exhaustively studied and analyzed, under ideal assumptions, such as additive white Gaussian noise (AWGN) or known noise power. However, more realistic scenarios, where the noise uncertainty, a realistic wireless channel, and hardware impairments of a radio receiver limit its performance, or make it infeasible for some applications [24], [25], [26], [27], [28], [29], [30], [31], [32]. The required number of samples for detecting signals that span over a short period of time is studied in [28]. The noise uncertainty and non-Gaussian noise are taken into account in [26] and [27] to assess the performance of an ED. This uncertainty makes it challenging to compute a threshold in order to meet the detection requirements in CR networks. Furthermore, the time-varying noise floor in practical CR scenarios has motivated the proposal of adaptive mechanism for the threshold selection, where the estimation of noise power and the computation of signal-to-noise (SNR) ratio are required [29], [33]. Some of these aspects are depicted in Fig. 7, where the received baseband signal composed of a PU signal, an spurious and noise floor in time domain are plotted. Energy values are outlined showing the challenge to detect a PU signal under spurious signal, while the Gaussian/Laplacian noise distribution are common assumptions found in the literature.
On the other hand, the importance of modeling hardware impairments is exposed and addressed in [25], for accurate energy-based SS. In Fig. 8, some sources of these impairments are depicted for a direct-conversion receiver (DCR). The LNA, for instance, introduces signal distortions, while phase noise and frequency offsets come from the local oscillator (LO). IQ impairments comprising timing/quadrature skews and gain imbalance are also described, where probably IQ gain imbalance is the most cited in  the literature. Moreover, it is worth noticing that the joint of these impairments are present in a SDR board, and their impact on the detection performance appeals for their study. In this regard, the LNA non-linearities, IQ imbalance, and phase noise, are analyzed in [31] and [32], while assuming flat fading Rayleigh channels. Real measurements have been considered to validate the underlying assumption when conceiving a novel approach. In [26], the authors study the performance of an improved ED under noise uncertainty, that later is validated through experimental validations in [30].

B. MATCHED DETECTOR
A matched filter (MF) is implemented by the cross-correlation between the known transmitted signal and the receiver one. Under a low SNR regime, it has reported a robust and better performance than an ED, at the expense of higher computational complexity. Commonly, the MF performance has been employed as benchmarking to assess the obtained gain of a novel approach [34] through computer-based simulations. In addition, it has been proposed to assist an ED in [35], and some mechanisms for a dynamic threshold selection have been conceived in [36].
Although, the assumptions about full knowledge of the signal can be hard to meet in practice. For instance, due to aspects such as non-Gaussian noise, proper timing, hardware impairments, or real fading channel, that degrade the crosscorrelation. It is also reported for purposes of experimental comparisons, as it is shown in [37] comparing a matched and ED using a common SDR platform.

C. FEATURE DETECTION
Specific signatures of the PU signal can be identified in most practical communication systems, such as preambles, pilots, cyclic prefix, second-order statistics, etc. Among these features, it is worth highlighting the periodicity of the second-order statistics. This feature is exhibited in digital modulated PU signals due to the symbol rate, chip rate, channel code or cyclic prefix, for which its detection is quite useful and popular in CR networks reporting numerous cyclostationarity-based detectors in the literature. This detector is based on the cyclic autocorrelation function (CAF) [23], and it has shown to be reliable at low SNR. However, it has a higher complexity, and requires a large sensing interval. Simulation-based results are often employed to validate novel approaches, and a simple variant based on symmetry property of CAF has been introduced in [38] and with a low complexity.
Its performance has been studied under noise uncertainty [39], while offsets in frequency and sampling clock are considered on the detection performance in [40], where P D and P FA expressions are provided. In fact, frequency offsets has reported a noticeable degradation in the detection performance, and it can occur because of LO drifts. It has been characterized for a SDR board in [41] that we reproduce in Fig. 9. An evident lack of accuracy and stability for a given carrier frequency is observed, for which the latest SDR implementations address the problem.
The computational complexity is also reported in [39] for different cyclostationarity-based detectors, and the cost of its implementation in FPGA boards [42]. Furthermore, measurements campaigns have been carried out employing SDR platforms in [38], [43], and [33] showing that the use of theoretical models may result in detection errors under real conditions, where the detector in [38] shows to be insensitive to the problem of local oscillator drift.

D. EIGENVALUE-BASED TECHNIQUES
Eigenvalue-based spectrum sensing can make detection by capturing correlation features in space, time and frequency domains, and probably the spatial domain is the one that have received more attention. In fact, multiple antenna detectors have been reported in a significant number of works in the literature, thus being addressed in this section.
These detectors are based on the sample covariance matrix of the received signal. They exploit the fact that under the null hypothesis the signals received at the different antennas are spatially uncorrelated, whereas the presence of a PU induces some correlation and/or additional structure in VOLUME 10, 2022 FIGURE 9. LO drift of an off-the-shelf SDR board. It causes frequency mismatches at the receiver side degrading the performance of cyclostationarity-based detectors [41]. the spatial covariance matrix. Among these methods, the eigenvalues of the covariance matrix, have been so far explored [44], [45], [46].
However, in more practical scenarios, the detection performance is degraded due to the uncalibrated antennas. It has been described in Fig 10, where each DCR at each antenna has different RF and IQ impairments, and the noise variances measured at the input to the SS approach are different. IQ imbalance [47] has been considered for the proposal of novel detectors in [48] and [46], while the improperty of the received complex-value signals due to the IQ imbalance has also been addressed in [49], where a constant false alarm rate detection is reported. The multi-antenna signal detection has also been studied under PU-signal correlation due to oversampling [44], and novel detectors under interference are also conceived in [50].
Efficient implementations on FPGA boards for its incorporation to SDR platforms have been reported in [51], and several comparisons of eigenvalue-based detectors using real measurements have been presented in [52] and under timeselective channels. In addition, some covariance-based detectors tailored for uncalibrated antennas are assessed in [53].

E. WIDEBAND SPECTRUM SENSING
WSS techniques aim to determine the available channels in a wide range of frequency, and one of the key requirements is the higher sampling rate, as it can be appreciated with the numerous works found in the literature classified as Nyquist or Sub-Nyquist WSS.
Nyquist approaches require a high computational complexity due to the high sampling rate, for which Sub-Nyquist techniques have attracted research interest. It employs sampling rates lower than the Nyquist rate, while detecting spectral opportunities using partial measurements, thus suitable for being evaluated using SDR boards. However, wideband RF front-end circuits introduces impairments into the received signal. For instance, the nonlinear components, LNA and mixer, produce intermodulation terms, while the ADC converter introduces spurious signals, thus making the detection of unoccupied bands more difficult and degrading the detection performance if not mitigated [54]. It can be observed in Fig. 11, where spurious frequencies appear in occupied and non-occupied sub-bands and are more likely to occur as a large bandwitdh is sensed. Moreover, the noise floor may vary across all sub-bands, and it has been studied under nongaussian, and impulsive noise in [55].
The validation of Sub-Nyquist approaches has been addressed through experimental validations [56], [57], and compressive sensing is incorporated to reduce the computational complexity in [57]. In [58] a tunable WSS detector is implemented on a FPGA board. It operates in the frequency range 70 MHz-6 GHz with a bandwidth of 30.72 MHz, and incorporates a pipeline architecture to reduce the latencies. Thus, allowing to characterize the time-varying PU traffic.
The availability of SDR platforms has motivated the experimental validations under different impairments [9], [59], [60]. A novel WSS detection under colored noised and partial spectral overlap is evaluated in [9]. A calibration method to estimate the actual transfer function of a Sub-Nyquist sampling architecture (concretely a modulated wideband converter [61]) is conceived in [59]. Recently, analytical and experimental results for a cyclostationarity WSS detector at a 5G frequency band (3.5 GHz and 100 MHz) is studied in [60], where a degradation in the P D is observed. Fig. 12 summarizes the main sources impairments and detectors for WSS.

F. COOPERATIVE SPECTRUM SENSING
Unreliable and miss-detection rate of a single SU is often caused by fading, shadowing, hidden terminals. To mitigate these issues, CSS strategies are adopted to exploit the spatial diversity among the observations of SUs. It motivates the deployment of testbeds emulating these scenarios, as well as  campaigns of experimental evaluations to assess the performance of CSS approaches. Nevertheless, in addition to the DCR impairments already exposed, the implementation of these strategies might be affected by other impairments such as timing inaccuracies among the SUs for simultaneous local sensing, the different sensing capabilities of SUs, heterogeneous SNRs at each SU, the presence of external interference coming from other wireless networks, among others [62], [63], [64], [65].
In a CSS detection, cooperative users report their local measurements to a fusion center (FC) for making a decision, where the entire measurements are reported (soft combination scheme) or one-bit decision (hard combination scheme). A soft-combining decision rule for cooperative prediction is proposed and assessed employing real-world WiFi signals on a SDR platform [62]. In addition, the problem of soft-decision schemes perfectly matches the matching learning (ML) paradigm [65]. ML has been employed to mitigate the presence of narrowband external interference, and it has been validated by means of experimental evaluations [63].
Furthermore, the hidden terminal problem that provokes a miss-classification of spectrum occupancy by some SUs has also been addressed in [67]. In this work, the authors propose a strategy to place SUs, so that the whole area of the PUs can be covered. With this methodology, the performance of a ML based CSS scheme can be guaranteed, which is confirmed by resorting to experimental evaluations. Finally, the detection performance is also analysed when applying an energy detector under non-Gaussian noise in [66], which is validated by experimental measurements.

G. SPECTRUM SENSING AND SDR IMPAIRMENTS
In Table 3, we summarize the different SS techniques along with the addressed HW impairment. Some of them are more relevant for a particular type of SS such as the frequency offsets and RF-non linearities for cyclostationarity and WSS detectors respectively. Furthermore, it is also worth mentioning that these approaches take into account the mentioned impairments by modeling them, thus conceiving novel approaches. Nevertheless many of them are not validated through experimental evaluations confirming the attained performance in more realistic environments.

III. SDR PLATFORMS
The implementation of CR approaches follow designs that typically employ programming languages at low-level and/or at high-level. At low-level, hardware description languages (HDL), such as Verilog or VHSIC Hardware description language (VHDL), aim at designing the digital logic of the system with register-transfer level (RTL) abstractions, where the exchange of data between registers can be designed. The employment of HDL allows the implementation of low-level architectures providing more control in the design of its components. Nevertheless, it entails a large time in acquiring expertise to implement at low-level. In this regard FPGA-based SDR platforms can use dynamic partial reconfiguration in order to reduce reconfiguration time.
On the other hand, the time of implementation at low-level can be avoided by generating HDL from programming language at high-level. For instance, high-level synthesis (HLS) processes are incorporated to convert designs described by programming languages at high-level (C/C++, Python, etc) to HDL. In this way, it allows the researchers to address targeted design by just providing algorithmic specifications.

A. SOFTWARE TOOLS
The selection of the software corresponds to the general requirements of a CR implementation. Next, we list the most employed tools:

1) MATLAB AND SIMULINK
MATLAB and Simulink have become very common and widely adopted for most designers that start employing SDR tools [68]. The high-level of language and block diagram environment, named Simulink, alleviates the task for the design and experimental evaluation of wireless transmissions. It is able to work with hardware platforms such as universal software radio peripheral (USRP) [68], RTL-SDR [69], ADALM-PLUTO [70], Zynq SDR, among others.
On the other hand, the provided high-level tools may not be enough to the needs of our design. For instance, the implementation of more complex scenarios working in real-time, scalable designs and the lack of open-source code. In fact, it is often used for offline processing prohibiting its employment to more sophisticated emulations. 2) GNU RADIO GNU Radio is a set of open tools aiming to implement SDR systems [71]. It provides signal-processing blocks that are interconnected to form a flow graph representing the implemented transceiver in software. Basically, these blocks consist of source blocks (data files, audio files, etc), processing blocks (modulators, filters, multipliers and amplifiers), and sink blocks (FFT sink, constellation sink, oscilloscope sink). One of the main advantages is the capacity to define and add new blocks by employing C++ or Python. It is carried out by using a gr-modtool [72] script that allows to create our digital processing block. Besides, an intuitive graphical user interface (GUI) called GNU Radio Companion (GRC) is provided to alleviate the task of designing a new transceiver. Moreover, GNU Radio can be used with an external RF hardware (e.g. USRP [73], LimeSDR [74], RTL-SDR [69], ADALM-PLUTO [70]) or without it in a simulation-like environment.
Unlike Matlab, it allows the implementation of real-time radio system. However, it requires some expertise and knowledge to be more familiar with the developing tools. Moreover, although there exists a large number of projects available on the web, compatibility issues and a lack of updates prevent a rapid prototyping.

3) LabVIEW
National Instruments (NI) also offers a tool for the development of a SDR platform [75]. Likewise, GNU radio and Simulink, the design can be constructed schematically by connecting a chain of various blocks together in a visual programming environment. The different blocks of the system can be implemented using high-level languages such as C or MATLAB, and it is compatible with USRPs.
In addition, it provides application frameworks tailored for LTE and 802.11 implementations among others to easily get a real-time prototyping. Thus, allowing researchers to focus on radio aspects of their interest. Nevertheless, the products and licenses are not free, and the price for them could be not affordable.

4) XILINX VIVADO HLS
The Xilinx Vivado HLS tool [76] is a software suite produced by Xilinx. It provides a design environment for HLS that is widely employed on numerous Xilinx FPGA boards, where C, C++ and SystemC programs can be directly used into Xilinx devices avoiding the need to manually create RTL.
However, its usage is limited by the expertise and knowledge of the researcher about low-level languages and hardware architectures. For that end, it is often utilized when the radio design can not be addressed with the current highlevel tools.

B. HARDWARE PLATFORMS
The hardware of a SDR transceiver usually contains components such as a general purpose processor (GPP), digital signal processing (DSP), and FPGAs. The current SDR platforms are implemented employing a mixture of them, and the SDR-based prototypes require to review briefly some of these components.
GPPs are the preferable hardware platform by researchers in academia, since it can be used for several purposes. For instance generic x86/64 computer microprocessors, ARM architectures, and boards such as USRP [73] and LimeSDR [74]. However, the sequential processing of its instruction set (e.g. arithmetic and logic unit -ALU, data transfer, and I/O operations) limits its performance for realtime operations, where high-throughput and low latency are often required.
This limitation is overcoming with the incorporation of co-processors, such as the graphic processing unit (GPU), that are designed to process large blocks of streaming data in parallel for signal processing algorithms.
On the other hand, FPGA is an integrated circuit designed to be configured by a customer or a designer after manufacturing. It contains an array of programmable logic blocks, that can be configured using HDL. During the last years, it has significantly advanced and become more powerful computationally playing a remarkable role in embedded system development. However, the required prior knowledge about the FPGA hardware architecture can be time consuming for an efficient SDR implementation, and HLS tools have shown (e.g. Xilinx Vivado HLS) to alleviate this work.

C. SDR-BASED PLATFORMS FOR SPECTRUM SENSING
Next, we describe the current SDR platforms employed for CR networks.

1) USRP
USRPs are probably the most adopted and popular hardware platform. This board is generally composed of an RF frontend, ADC/DAC, and an FPGA, where the majority of operations (baseband processing) are offloaded to a GPP (host computer) via either USB, Ethernet, or peripheral component interconnect express (PCI-Express) connection. In addition, an embedded series comes with an internal GPP to operate in a standalone mode.
On the other hand, due to the fact that SDR tools such as GNU Radio, LabVIEW, and MATLAB Simulink support these boards, it is often reported in several research experiments motivating even more its employment. Nevertheless, USRP-based testbeds do not necessarily meet the requirements of communication standards. The bandwidth of the RF front-end, the data streams with a host processor, latencies, and the hardware impairments, are common aspects that usually affects the throughput and timing characteristics of the platform. For that end, each released product provides more advance features to overcome these concerns. For instance, the employment of SoC integrating an ARM processor with the hardware programmability of an FPGA (e.g. Zynq-7000 family), as well as supplementary accessories (e.g. GPS disciplined oscillator -GPSDO). An example of a SoC-based SDR platform is the portable stand-alone USRP E310 embedded SDR that does not need a host computer, thus being suitable for field deployments.

2) LimeSDR
A low cost option and open source SDR, LimeSDR, is also available on the market, it has a similar architecture to the USRPs, and is comprised of a field programmable RF transceiver, an Intel FPGA, and a microcontroller. Moreover, it is connected to a GPP via USB 3.0, where the transceiver has the task to transmit/receive the wireless data, while the GPP generates the data and process the incoming signal. LimeSDR comes with LimeSuite software including source code, firmware, and schematics, and it is also supported by GNU radio.

3) ADALM-PLUTO
ADALM-PLUTO is an active learning module aiming to introduce the fundamentals of SDR. It is composed of a Xilinx Zynq-7000 SoC, and an analog-device-based RF front-end. Like previous architectures, it is connected to a GPP via USB, and a variety of software can be utilized such as MATLAB, Simulink, GNU Radio, or custom applications (C/C++,C#,Python). Although, it has been designed for teaching purposes and has some limitations such as the RF coverage, the number of antennas, and bandwith frequency 20 MHz, it has been used for experimental evaluations whenever it fulfills the requirements.

4) RTL-SDR
RTL-SDR is the cheapest SDR receiver available today. It can be used as a based radio scanner for receiving radio signals covering a wide range of frequencies, it provides a custom driver to do some acquisitions and it is also supported by GNU radio.

5) bladeRF
bladeRF [77] is a promising powerful waveform development platform. It provides a 2 × 2 MIMO SDR, covering a large frequency range up to 6GHz, and connected to a GPP via USB. Supported by GNU radio and Matlab/Simulink, a HDL platform is also provided for the implementation of VHDL modem on FPGA meeting low latency and timing control required for the modulation/demodulation of 802.11 packets.

6) HackRF
HackRF One [78] is another SDR platform operating within a wide range of frequencies (1 MHz to 6 GHz). It is connected to a GPP via USB connector or programmed for stand-alone operation, and is compatible with GNU Radio. VOLUME 10, 2022

7) BB60C
BB60C is a real-time RF spectrum analyzer covering a frequency range from 9 kHz to 6 GHz, with 27 MHz of instantaneous bandwidth [79]. This super-heterodyne receiver is connected to a GPP via USB. It provides a calibrated streaming suitable for accurate measurements, and a kit of development tools for custom applications. Moreover it is also supported by GNU Radio.

8) AIR-T
Artificial intelligence radio transceiver (AIR-T) is a high-performance SDR [80]. It incorporates an Nvidia GPU, a Xilinx FPGA and a multi-core CPU tailored for developing autonomous applications based on artificial intelligence and machine learning. A 2 × 2 MIMO SDR covering a frequency range from 300 MHz to 6GHz can be implemented employed embedded or edge series. This SDR can also be connected to a GPP via USB or Ethernet, and provides development tools to integrate deep learning into SDR systems, such as Anaconda [81] for ML, a compute unified device architecture (CUDA) [82] platform, and is also supported by GNURadio.

IV. PROTOTYPING WITH SDR PLATFORMS
A first step towards the implementation of a SDR-based testbed is the understanding of the general requirements. Different types of applications may lead to totally different requirements related to the hardware and software components. Some of these aspects can be, for instance, the number of nodes (transceivers), the processing capabilities at each node, single or multi-antenna nodes, the network topology (cooperative schemes), stationary/non-stationary environments (scenarios), among others. For that end, these prototyping challenges hindering the implementation of SS are reviewed in this section.

A. SDR BOARD SELECTION
The technical specifications related to each SDR board are reviewed when selecting a SDR board. The frequency range and bandwidth are some of these parameters. For instance, if a board covers the ISM frequency bands and wider bandwidths are required. Moreover the receiver sensitivity turns out to be relevant for the detection of low-power signals, where the noise figure (NF) of the RF front-end allows us to determine the noise floor, thus establishing minimum detectable signals.
A higher accuracy of a LO allows a proper downconversion of the RF signal and sampling at the ADC, for which the accuracy of the clock (i.e. the LO) is also examined. It is expressed in terms of its frequency variation, i.e. in parts per million (ppm) or parts per billion (ppb) and reported in the datasheets of SDR boards. Furthermore, phase noise specifications can also be extracted from the manufacturers. Nevertheless, a better characterization of these features can be carried out through experimental measurements [83]. This is often the case to determine the spectral purity of the receiver. A clean spectrum without strong DC and image components as well as spurious signals is desirable, otherwise it will suggest us to apply some calibrations.
Each SDR board incorporates different hardware components, and the characterization of the receiver related to the aforementioned features becomes relevant for the assessment of the detection algorithms. In Table 4 we have reported some of these specifications in order to assist the selection of a SDR board.

B. SPECTRUM SENSING-BASED CHALLENGES
So far, we have addressed different SS techniques concerning their complexity, multi-antenna, bandwidth, and collaborative aspects. In this subsection, we discuss some of the challenges concerning the implementation of these approaches on a SDR platform, while reporting current efforts aligned to these aspects.

1) NOISE UNCERTAINTY
Noise uncertainty may come from several sources, internal components of a receiver chain and the external environment. Despite of common assumptions such as the stationary white Gaussian noise with zero mean, the distribution of this undesirable random process is not not accurately known. Then, the noise variance has to be estimated in practice.
For instance in an ED, the threshold calculation depends on noise variance. Besides, it has been assumed to be constant over time and frequency domain. Furthermore, variations in the noise power levels across the spectrum may occur. Experimental evaluations employing a fixed threshold confirm the degradation on the detection performance, and some adaptive schemes have been proposed in [84] and [29], where a benchmarking performance of an ED is provided. A real-time computation of the threshold shows to obtain significant gains in terms of detection and false alarm probabilities [29], and under low SNR conditions [84].

2) MULTIANTENNA
The experimental evaluations assessing the detection performance of multiantenna-based detectors have to deal with ignored problems in simulation-based environments. For instance, the transmission or acquisition of data streams through a given number of antenna requires to be executed simultaneously at the same time, while most available SDR boards are limited by the number of antennas. It has motivated the implementation of multi-antenna testbeds integrating several SDR boards [85], [86], [87]. In doing so, a synchronization module for sharing reference signals is incorporated. A timing synchronization for the transmission/reception of several aligned data streams is achieved by triggering the execution of transmissions/receptions at the same time at each SDR board. A clock signal is also shared among them, so that the LO can be obtained from a common source for all boards. An scheme of this synchronization among SDR boards is described in Fig.13, where a central controller is also included for the parameter settings of the boards, storage of sensed data, among others. Some examples of these multiantenna SDR platforms can be found in [85] and [86], where the timing synchronization is carried out by sharing a pulse-persecond (PPS) signal provided by the Octoclock, and a NI module (XIe-6674T [88]) is employed to generate a clock signal, crystal oscillator (10 MHz), that later is amplified and distributed by the Octoclock, [89].

3) WIDEBAND
The implementation of larger WSS schemes on SDR platforms imposes other challenging aspects. For instance, any SDR architecture requiring a connection to a GPP via Ethernet cable (e.g. 1 GbE), will be limited to stream up to 25 Msps (for 16-bit samples), thus creating a bottleneck for larger values of Msps and consequently for larger bandwidths. This is the case in [87], where the authors require a sustainable signal acquisition at very high sampling rates, and consequently high data rate throughput and storage (e.g. 400 Msps covering a bandwidth of 320 MHz). In this regard, a radio frequency network on chip (RFNoC) development framework provided by Ettus Research enables the FPGA processing in USRP devices (e.g. X310), so that data streaming between the FPGA and the host PC is used to implement an acquisition platform for off-line processing.

4) COOPERATIVE
The implementation of cooperative approaches requires a degree of synchronization among the nodes being deployed. In CSS, it is expected to have synchronized measurements among SUs, so that the activity of a PU in a given band can be detected simultaneously, for instance, in nonstationary environments. In this regard, a triggering signal is implemented using PPS or GPS signals, as well as network time protocols. Quite often, these approaches are reported for indoor environments where a CSS scenario emulates a real world scenario with a limited number of nodes [63], [67], as it is depicted in Fig. 14, while outdoor measurements require the deployment of several nodes monitored using GPS signals and network time protocols for the synchronization of the experiments [90].

5) MOBILITY
The mobility of PUs causes changes to the set of available channels for SUs transmissions. Even more SUs may change their physical locations modifying the vacant channels of SUs neighbors. On the other hand, it is also challenging to recreate these scenarios in a testbed, where a limited number of PUs VOLUME 10, 2022 FIGURE 13. Implementation of a multi-antenna receiver employing SDR boards. A synchronization module allows the streaming data of each SDR board to be aligned by sharing reference signals.
and SUs can be tested, and emulations help to recreate such scenarios [52].

6) CR SCENARIOS
A testbed to recreate a CR scenario is also required to validate an approach. For instance, a real-time testbed to evaluate the impact of the interference on the detection performance is implemented in [91]. A CR scenario may also require the deployment of a large number of nodes, and emulations are one of these approaches to recreate a large-scale scenario (e.g. using a USRP node to emulate several users by transmitting continuously packets). However, it overlooks other factors such as multipath, path loss, hidden node problems. In this regard, some efforts in building large-scale testbeds [92], aim to provide external users a virtual space to deploy their own approaches (e.g using Dockers containers and GNURadio in [92]). In fact, we can expect in future to have an online platform where all approaches can be compared under the same conditions avoiding users to face prototyping challenges.

C. SDR TRANSCEIVERS-BASED CHALLENGES
The imperfections of a SDR transceiver degrades and limits the performance of a CR system.

1) TRANSCEIVER IMPAIRMENTS
The non-linearities of the power amplifier (PA) due to the saturation region introduce non-linear distortions provoking a regrowth of the output signal resulting in adjacent channel interference (ACI). This problem is more evident with the multicarrier distorsion faced by the transmission of orthogonal frequency-division multiplexing (OFDM) waveforms. The multiple parallel data flows transmitted on different subcarriers cause a large variation of amplitude in the time domain that requires to have an enough dynamic range of the PA. Several methods to reduce this variation (also known as peak-to-average power ratio -PAPR ) have been proposed, In [93], the authors proposed a method based on a filtering FIGURE 14. Emulation of a CSS scenario for a more real world experimentation. Real wireless channel, obstacles and interference are considered, while the FC is also emulated by a PC for controllability and repeatability of the experimental evaluation. and clipping of the OFDM transmissions to improve the PAPR which is evaluated by modeling the PA, and eventually resorting to experimental evaluations on a SDR platform. In a similar way, a digital pre-distortion is designed and evaluated experimentally to compensate for PA impairments [94].
Although, the conversion of an analog RF signal to the digital form introduces some quantization errors at the ADCs, it does not deviate significantly the obtained performance from that obtained with infinite precision. In fact, the ADC resolution (e.g 12 bits) is set by default with the board and similarly for the ADC sampling frequency (sample rate per second). Despite of it, an interesting approach related to the ADC resolution is reported in [95]. The authors show the feasibility for the simultaneous acquisition of two signals with different powers, where an intra-quantum signal is acquire due to the contribution of a strong transmission when it crosses the quantization levels of the ADC, which is corroborated using one the latest SDR boards.
In practice several types of synchronization errors in frequency, time, phase drifts, introduce offsets degrading the performance of spectrum awareness approaches. A common source of LO is employed to provide these signals, and consequently its frequency stability turns out to be relevant. For instance a quartz oscillator with an output frequency of 1 MHz and 5 ppm will have a variation in frequency of 5 Hz. This problem is more evident for some approaches such as a cyclostationarity detector very sensitive to the mismatch of the cyclic frequency offsets due to these drifts. What is more, some applications (e.g. multiantenna approaches) require to have a common source of oscillator among the different RF receive chains of each antenna.

2) SDR CALIBRATION
Calibration of the RF hardware is often required when doing measurements since imperfections caused by e.g. the PA, LNA, the filtering and filter transients, can cause the measured values to differ throughout the deployed nodes. This calibration can be done in the software by correcting the values of the received or transmitted signal based on a table of measured corrections factors. A simple calibration procedure can be made by connecting a signal generator directly to the board, and set up the corresponding gains. Some calibration procedures have been reported in [32] and [59]. However, the fact of having reconfigurable SDR and the different parameters related to CR approaches can be difficult to comprehensively and cost-effectively test with traditional methods. It is motivating the development of software defined synthetic instrument (SDSI) [97] to automatize the test and validation of SDR transceiver, even enabling the intermediate steps within the radio chain.

3) SDR RESPONSE DELAYS
High and unpredictable delays are commonly found in SDR platforms, and several studies have been carried out for its characterization [62], [98]. In this way, and with a better understanding of the SDR latencies, novel approaches can be conceived meeting the requirements of the targeted application. It is related to the time a frame takes to be transmitted or received when it goes through the transceiver chain. These latencies are often grouped and identified as: transceiver latency, communication link latency and host latency, as it is shown in Fig.15, where significant delays have been reported concerning the communication link and host.
They have been measured and analyzed for the exposed platforms, such as GNU Radio [99], [100], employing USRPs and LimeSDR boards. Moreover, the integration of CR principles in spectrum access schemes, has been studied taking into the introduced latencies by the SDR platforms, as it is depicted in Fig. 15b. In this spectrum accesses scheme, the time after a sensing period and the transmission of packet (after determined that the channel is vacant) T p is described, as well as the time to start receiving an ACK packet from a SU, T w , both of them including SDR latencies.
In [62], the authors aim to assess the performance of a hidden markov model (HMM)-based prediction approach, for which the delay response of the SDR platform for packet transmissions after a previous sensing period slots turns out to be critical. This time between the end of the time slot for SS and the beginning of a time slot for a transmission is reported T p , and the approach validated using real WiFi signals. More recently, the authors in [98], report long and variable delays when waiting for an ACK packet, T w , after a packet transmission, for which a new sense-transmit scheme is proposed.

D. PROTOTYPING CHALLENGES AND EXPERIMENTAL VALIDATIONS
The implementation of an SDR-based testbed faces several prototyping challenges. In table 5, we summarize them when targeting the implementation of multiantenna, wideband, and cooperative SS on SDR-based testbeds. In addition, the corresponding approaches found in the literature are described.
It is worth highlighting that the validation of an SS technique aims to validate the proposal employing real measurements. In doing so, a main obstacle for some of these SS proposals is the implementation on an SDR platform, while for others it is straightforward allowing, in any case, the study of these approaches under different conditions. In Table 6, we show the numerous efforts for several SS techniques along with the employed SDR platform, implementation aspects and the validation addressed.

V. ENHANCEMENTS AND FUTURE RESEARCH DIRECTIONS
In this section we will discuss the enhancements that can be considered for the experimental evaluation of reported works, as well as the research gaps found to address for the support of the future generation of wireless communication networks.

A. ENHANCEMENTS FOR THE EXPERIMENTAL EVALUATIONS
At the light of exposed review, and with the current development of SDR platforms, novel approaches require to address more challenging CR scenarios under realistic conditions.

1) ENHANCED MODELS
Exhaustive measurement campaigns based on SDR platforms have been carried out to extract different statistical features of the spectrum occupancy. It has enabled the conception of several spectrum occupancy models that can be used to improve the accuracy of SS approaches. Nevertheless, these models have not been employed with the current experimental evaluations based on collecting noise or signal.
In addition, a perfect synchronization between the SU and PU has been so far assumed, so that two assumptions regarding the presence of a PU is only considered, i.e. under a null and alternative hypotheses. This assumption, however, is very hard to satisfy in a real word environment, where an asynchronous traffic of a PU is expected to be sensed by a SU. In this context, its impact on the sensing performance has been, for instance, addressed in [103], thus motivating the detection performance of SS approaches through more realistic experimental evaluations.

2) SDR IMPROVEMENTS
One of the challenges for GPP-based SDRs is to provide a generic runtime environment (e.g. to meet the performance for 5G Networks), for which low latencies and high throughput are among the desirable enhanced features of SDRs. Although, the new generation of SDR platforms is incorporating advanced features such as LabVIEW FPGA module and RF Network on Chip (RFNoc), the advantage of this computational power is still underused. An interesting work is the employment of these tools to provided an embedded solution (e.g. using RF Network on Chip). In this way, it allows to introduce our designs into the SDR boards instead of implementing them on the host PC, thus mitigating the communication bottlenecks between the SDR board VOLUME 10, 2022  and host PC, for achieving lower latencies and/or higher throughput.
In addition, the adaptability of CR devices using the reconfiguration of radio parameters has been barely studied. For instance, modulation, coding schemes, power control, the operating carrier frequency, and bandwidth among others requires to be adaptive in real-time. Although, some of the software tools (e.g GNU Radio) incorporate this ability, it is still missing for other platforms limiting most of the experimental evaluations to off-line assessments.

3) REAL DEPLOYMENTS
A complete spectrum management framework showing its operability, while coexisting with other communication technologies, is not quite often reported in the literature. Although, there exists open platforms providing a full implementation of the protocol stack such as OAI [104], and srsRAN [105], these platforms require more efforts to learn its usage and key features. On the other hand, the incorporation of communication system components into the SDR platform is facilitating the deployment of more realistic testbeds. For instance, modules and toolboxes specifically designed for cellular GSM, LTE, or Wi-Fi enable a rapid prototyping that can be considered for assessing the reported SS approaches.

B. COGNITIVE RADIO FOR 5G
5G networks will bring higher bandwidth and download speeds (in the order of multi gigabits per second), and several VOLUME 10, 2022 underlying technologies for the radio access technology (5G NR) are expected to support them. Aiming to this goal, several gaps are expected to be addressed for the deployment of this technology.

1) MASSIVE MIMO
The effectiveness of current algorithms tailored for exploiting the diversity for massive MIMO networks will require to be re-examined. Massive MIMO incorporates a large number of antennas increasing its complexity, as well as the hardware cost. Although low-cost components enable the deployment of this technology, it comes with an increase of HW impairments. SDR platforms bring the ideal tool to assess the impact of these impairments in order to mitigate them. For instance, with the introduction of new signal models (as it has been exposed in this article), as well as with the design of compensation algorithms.
Exhaustive measurement campaigns not only for indoor, but also for outdoor environments require to be carried out to assess the massive MIMO performance. These measurements can allow to evaluate the viability of these systems in shared spectrum scenarios, so that new rules concerning the coexistence between PUs and SUs can be conceived. Nevertheless, these testbeds are not still quite reported due to the complexity of their prototyping with the current limitations of SDR boards.

2) WIDEBAND APPROACHES
The emerging bandwidth-hungry applications will be supported by WSS approaches. Nevertheless, wideband channel impairments such as frequency-selective fading, interference from close frequency bands, nearby transmitters, colored noise, or fast large-scale channel effects require to be taken into account on the sensing performance.
This evident gap can be fulfilled with the employment of SDR platforms where calibration methods for HW impairments related to wideband RF front-end (e.g. for millimeter waves) need to be addressed. In addition, the complexity of the digital front-end to deliver a higher throughput meting lower latencies require more efforts for its rapid prototyping, and experimental evaluation with large volumes of data.

3) IoT NETWORKS
IoT over 5G cellular networks is one of the technologies considered for massive connectivity and better efficiency. Nevertheless, the evaluation of novel approaches (such as CSS) has been barely assessed in large-scale scenarios. In fact, the performance of these solutions is typically evaluated for indoor environments and generalized for large-scale scenarios leading, for instance, to an inefficient spectrum reuse.
The employment of empirical channel models fills the lack of accuracy of theoretical models, so as to improve the accuracy of a detection approach, for example taking into account the mobility of sensors. The main limitation to fill this gap, is the cost and manpower to deploy these networks at large scale. In this regard, a promising prototype overcoming all these challenges is described in [106], where a cheap SDR board, composed by an FPGA board and a microcontroller, is configured over the air, while having common features of an IoT mote such as low power sleep modes.

4) MACHINE LEARNING
ML techniques have being successfully employed in cognitive radio networks, where SS can be formulated as a classification problem [107]. Deep learning is one of these classes of ML algorithms that using multiple layers extract relevant features from the raw data and it is reporting significant gains in comparison to conventional approaches based on model assumptions [108].
These approaches based their performance on the ability to learn the spatial and temporal features of the PU signal such as the energy-correlation features, or PU activity patterns. Nevertheless, the reported gains have not been experimentally verified. A gap to be covered should address the learning of these features using real measurements introducing spurious signals due to the HW impairments. Moreover, we can expect to validate and conceive more realistic DL-based approaches for MIMO, wideband and cooperative approaches.
On the other hand, the introduction of GPU-based SDR platforms promises a rapid prototyping. An interesting initiative is the development of AIR-T, that is specifically designed for high-performance computing applications. Concretely, the creation of ML-based wireless systems whose feasibility can be examined for spectrum sharing services in 5G networks.

5) ANTENNA DESIGNS
The design of antennas constitutes an important part in the development of new wireless devices. For instance, new frequency bands, wider channel bandwidths, higher antenna gains, and more antennas for MIMO systems are required for the 5G technology [109].
In massive MIMO, a realistic performance assessment requires to consider the features of a large antenna array (e.g. shape and size), so that it can capture the 3D propagation environment. Conventionally, patch antennas, omnidirectional dipoles, or 2D planar arrays have been employed for testbeds suitable for beamforming applications, and it is not clear which antenna pattern provides the best performance.
For massive IoT applications, the employment of compact antenna arrays along with the digital control of radiation pattern more directive, allows us to validate novel approaches when scanning surrounding sources. For instance, the accuracy for determining PU beams can be assessed dealing with multipath propagation, and under hardware imperfections (such as non-calibrated arrays or phase synchronization) due to the low-cost SDR boards.
We expect to have more antennas for millimeter-wave frequencies simplifying the deployment of large array of antennas (unlike the size for those arrays working below 5GHz), thus allowing the assessment of adaptive approaches [110] in new frequency bands and with wide tuning capabilities.

C. STANDARDIZATION
Many standardization groups are working on incorporating CR technologies to communication system such as WLAN, ZigBee, and wireless personal area networks (WPAN), to exploit TV white spaces. In this regard, the experimental works provide accuracy channel models that later can be employed to assess system-level and link-level performances of advanced signal processing techniques for standardization purposes. The proof of concepts and experiments, part of the standardization phases, can be supported by the development of SDR-based platforms. Finally, different standards can be evaluated to decide with one is the most suitable for a given application.

VI. CONCLUSION
The scarcity of the radio spectrum has motivated the development of CR networks attracting research interests and reporting numerous approaches. Nevertheless the adoption of this technology is still facing several challenges. In this survey, we provide an overview of the latest works in CR networks, while taking into account relevant aspects of their implementation. Current SDR platforms have been reviewed and surveyed, and the main challenges for the deployment of CR approaches have been identified on SDR platforms. Finally, potential research directions as well as open issues are provided.