Spectrum Sensing for
Dynamic Spectrum Access
in Cognitive Radio
2 Abstract. The number of mobile devices is constantly growing, and the exclusive
static spectrum allocation approach is leading to the spectrum scarcity problem when some of the licensed bands are heavily occupied and others are nearly unused. Spectrum sharing and opportunistic spectrum access allow achieving more efficient spectrum utilization. Radio scene analysis is a first step in the cognitive radio operation required to employ opportunistic spectrum access scenarios such as the dynamic spectrum access or frequency hopping spread spectrum. The objective of this work is to develop and virtual prototype the subset of radio scene analysis algorithms intended to be used for deployment of opportunistic spectrum access in our target application: a cognitive radio network consisting of multiple software-defined radio nodes BitSDR. The proposed radio scene analysis algorithms are devoted to solving two radio scene analysis problems: 1. detection of vacant frequency channels to implement spectrum sharing scenarios; 2. waveform estimation including modulation type, symbol rate, and central frequency estimation. From the subset of two radio scene analysis problems two hypotheses are formulated: the first is related to the vacant band identification and the second to waveform estimation. Then seven research questions related to the trade-off between the sensing accuracy and real-time operation requirement for the proposed radio scene analysis algorithms, the nature of the noise, and assumptions used to model the radio scene environment such as the AWGN channel.
In the scope of this work, Hypothesis 1, dedicated to vacant frequency bands detection, has been proven. Research questions related to the selection of the observation bandwidth, vacant channels detection threshold, and the optimal algorithm have been answered. We have proposed, prototyped and tested vacant frequency channels detection algorithm based on wavelet transform performing multichannel detection in the wide band of 56 MHz based on the received signal observed during 500 microseconds. Detection accuracy of 91 % has been demonstrated. Detection has been modeled as a binary hypothesis testing problem. Also, energy detection and cyclostationary feature extraction algorithms have been prototyped and tested, however, they have shown lower classification accuracy than wavelets. Answering research question 7 revealed the advantage of using wavelets due to the potential of the results of wavelet transform to be applied for solving the waveform estimation problem including symbol rate and modulation type. Test data samples have been generated during the controlled experiment by the hardware signal generator and received by proprietary hardware based on AD9364 Analog Devices transceiver.
To test Hypothesis 2 research questions related to the waveform estimation have been elaborated. We could not fully prove Hypothesis 2 in the scope of this work. The algorithm and features that have been chosen for modulation type classification have not met the required classification accuracy to classify between five studied modulation classes 2FSK, BPSK, QPSK, 8PSK, and 16PSK. To capture more of the fine differences between the received signal modulated into different linear modulations it has been suggested to use the spectral features derived from the time series signal observed during 500 microseconds or less observation time in the scope of the future work. However, the binary classification between 2FSK and BPSK presented in Paper 1 could be performed based on instantaneous values and SNR input: ensemble boosted trees and decision trees have shown an average classification accuracy of 86.3 % and 86.0 % respectively and classification speed of 1200000 objects per second, what is faster than required 2000 objects per second.
The prototyping and testing of the proposed algorithm for symbol rate estimation based on deep learning have been performed to answer research question 2. Wavelet transform feature extraction has been proposed to be applied as a preprocessing step for deep learning-based estimation of the symbol rate for 2FSK modulated signals. This algorithm has shown an improvement in the accuracy of the symbol rate estimation in comparison with cyclostationary based detection. The validation accuracy of the symbol rate classification has reached 99.7 %. During testing, the highest average classification accuracy of 100 % has been observed for the signals with SNR levels 25-30 dB, while for signals with SNR 20-25 dB it was 96.3 %.
To my grandparents Evgenia and Hanis,
List of Papers
This thesis is based on the following papers, which are referred to in the text by their Roman numerals.
1. Inna Valieva; Mats Björkman; Johan Åkerberg; Mikael Ekström; Iurii Voitenko. (2019) Multiple Machine Learning Algorithms Comparison for Modulation Type Classification for Efficient Cognitive Radio. MILCOM
2019 - 2019 IEEE Military Communications Conference (MILCOM).
2. Inna Valieva; Mats Björkman; Johan Åkerberg; Mikael Ekström; Iurii Voitenko. Blind Symbol Rate Estimation for Cognitive Radio Using Wavelet Transform and Deep Learning for FSK Modulated Digital Signals.
3. Inna Valieva; Mats Björkman; Johan Åkerberg; Mikael Ekström; Bharath Shashidhar; Iurii Voitenko. Autonomous Detection of Vacant Frequency Bands for Cognitive Radio. Manuscript.
This thesis consists of seven chapters.
Chapter 1 provides an introduction to the field and motivation for this work driven by the growing market of mobile devices and spectrum scarcity problem arising from the static spectrum allocation.
Chapter 2 provides the theoretical foundation and historic perspective to the software-defined radio and cognitive radio communication paradigms used in this work to answer the research questions and propose radio scene environment sensing algorithms contributing to efficient frequency spectrum utilization. Also, in this chapter our target application hardware: cognitive radio network is introduced. Also, the main challenges related to the radio scene analysis and its hardware implementation are discussed.
Chapter 3 provides an overview of the related work in the field carried out by other researchers. Algorithms applied for blind detection of vacant frequency channels such as matched filtering, waveform-based sensing, cyclostationarity-based sensing, energy detector-based sensing, wavelet transform, and AI-based algorithms are discussed in Chapter 3.1. Chapter 3.2 describes algorithms applied for waveform estimation including non-linear energy operator, inverse fast Fourier transform and baseband shape-based, maximum likelihood, wavelet transform and cyclostationary for symbol rate estimation and cyclostationary, maximum likelihood estimation, wavelets, AI-based algorithms for modulation type classification using various features extracted from the received signal.
In Chapter 4 the main research goals are defined. Primary, two hypotheses are formulated. The first is related to vacant band identification and the second to waveform estimation. To address the formulated hypotheses, eight research questions related to the tradeoff between the sensing accuracy and real-time operation requirement for the proposed radio scene analysis algorithms, the nature of the noise, and assumptions used to model the radio scene environment such as AWGN channel are formulated.
In Chapter 5 the research methodology used to answer the research questions and test the formulated hypothesis is described.
8 Chapter 6 summarizes the main scientific contributions of this work including the study of the radio scene environment and nature of the noise, proposed radio scene analysis algorithms for vacant bands identification and waveform estimation. Chapter 6.2 provides the summary of the included papers and Chapter 6.3 discusses the relationship between the included papers and the research goals.
Chapter 7 provides the main conclusions and defines important research questions for future work.
ContentsList of Papers ... 6 Thesis Outline ... 7 Contents ... 9 Abbreviations... 10 1 Introduction ... 11 2 Background ... 15
2.1 SDR and Cognitive Radio ... 15
2.2 Spectrum Sensing and Radio-Scene Analysis. Challenges ... 18
2.3 Target Application Hardware ... 19
3 Related Work ... 22
3.1 Blind Detection of Vacant Frequency Bands ... 22
3.2 Waveform Estimation ... 24
3.2.1 Blind Symbol Rate Estimation ... 24
3.2.2 Modulation Type Classification ... 25
4 Research Goals ... 29 4.1 Hypothesis ... 29 4.2 Research Questions ... 32 5 Research Methodology ... 35 6 Thesis Contributions ... 37 6.1 Included Papers ... 37
6.2 Relationship Between Papers and Research Questions ... 37
6.3 Scientific Contributions ... 38
6.3.1 Contribution 1 ... 38
6.3.2 Contribution 2 ... 41
6.3.3 Contribution 3 ... 46
7 Conclusions and Future Work ... 56
ADC Analog to digital converter AI Artificial intelligence
AWGN Additive white Gaussian noise BER Bit-error rate
BPSK Binary phase-shift keying CNN Convolutional neural network CR Cognitive radio
CWT Continuous wavelet transform DL Deep learning
DSA Dynamic spectrum access
DSSS Direct sequence spread spectrum FHSS Frequency hopping spread spectrum FSK Frequency-shift keying
MAC Media access control layer ML Machine learning
NET Network layer PHY Physical layer PU Primary user
QPSK Quadrature phase shift keying RSSI Received signal strength indicator SDR Software-defined radio
SOM Self-organizing maps SU Secondary user
The global market of mobile devices and services actively using the electromagnetic spectrum is continuously growing. Global data traffic has shown an exponential growth from ten thousand petabytes per month in 2008 to a hundred and twenty thousand in 2018 (Bouwfonds, 2020). Already in 2014, the number of mobile devices has exceeded the Earth’s population (Boren, 2014) and keeps increasing. Advancements in digital electronics are successfully fulfilling this ever-increasing demand for processing power and mobile data capacity by keeping up with Moore’s law. Traditionally the utilization of the electromagnetic spectrum has been performed using a robust static approach developed almost a century ago. It rations access to the spectrum in exchange for the guarantee of interference-free communication: spectrum is divided into the rigid, exclusively licensed bands, allocated over large, geographically defined regions. In conditions when some of these license bands are being nearly unused, while the others are overwhelmed, the problem of spectrum scarcity arises (Rosker, 2020). To cope with the growing data traffic demand, the electromagnetic spectrum utilization policies have been reformed in recent years to allow the unlicensed secondary users to access licensed bands without causing interference to the licensed primary users (Xin, 2015).
12 From the occupancy perspective, the spectrum is classified into three broad categories. Black spaces are spectrum bands occupied by high power local interferes. Gray spaces: are partly occupied by low power interferers (Haykin, 2005). However, they are still considered to be good candidates for spectrum reuse. Frequency bands assigned to a primary user but at a specific time and in a current geographic location not utilized by the primary user and free of the RF interferers, except for the ambient noise, are referred to as white spaces or spectrum holes (Varshney, 2017; Haykin, 2005). White spaces are obvious candidates for secondary reuse.
Practical spectrum measurements have provided some quantitative description of the spectrum underutilization and have shown great potential for its reuse. In the work of Awe (2015) the spectrum occupancy of 13 % was measured in New York and below 35 % in Washington for bands 30 MHz - 3 GHz. Figure 1 shows the spectrum utilization in the frequency range 0...6 GHz measured by the Berkeley Wireless Research Center at downtown Berkeley. For the frequency ranges 2 - 3 GHz and 5 - 6 GHz, the spectrum utilization is less than 10 %, while for 3 - 5 GHz, it is less than 1 % (Cabri, 2007). The study by Valenta (2010) of spectrum usage in European cities in the 400 MHz - 3 GHz band has shown also significant spectrum reuse opportunities: overall utilization of 6.5 % was observed in Northern suburb of Brno, Chech Republic; 10.7 % in Eastern suburb of Paris, France and 7.7 % in the city of Paris near the “Place la Nation”.
Spectrum scarcity has driven generations of wireless technology development (Varshney, 2017). SDR and cognitive radio are communication paradigms that are actively addressing the scarcity problem. Spectrum sharing and dynamic spectrum access are techniques used in CR to optimize the use of the electromagnetic spectrum. Spectrum sharing enables interference-free access to the same frequency bands between multiple users’ categories. The most common example of spectrum sharing in cognitive radio networks is the reuse of the TV white spaces: spectrum allocated to TV broadcasters for other wireless communications (Xin, 2015). The mutually exclusive access to the spectrum and absolute privilege of the PU results in arbitrary disruptions, unstable, and unpredictable performance for the secondary user. Dynamic spectrum access addresses the arbitrary disruptions of the secondary user and provides simultaneous access of the secondary users and primary users by sensing the vacant frequency channels and allowing devices to communicate in underused parts of the spectrum. Intelligent signal processing and decision-making are used to dynamically select spectrum band, time diversity, spatial diversity options, and generate a cognitive waveform (Rondeau, 2009). Waveform information of the active transmitting users including both: symbol rate and modulation type are necessary to employ an autonomous spectrum sharing scenario, i.e., to identify the spectrum reuse opportunities in conditions where no information about primary users and signals is available (Morozs, 2015).
13 Figure 2 presents different DSA scenarios. In 2003 the Spectrum Policy Task Group of the FCC suggested (FCC, 2003b) the interference temperature concept for underlay spectrum sharing to allow low-power transmissions in licensed (used) bands. The FCC claims that often, between the original noise floor and the licensed signal of the incumbent radios, there lies an interval that could be exploited, which was identified as ‘new opportunities for spectrum use’ (FCC, 2003b).
Figure 2. DSA and reuse of vacant frequency channels: opportunistic spectrum usage with undelay spectrum (low power) and overlay spectrum sharing (high power, where vacant frequency channels exist in time and space) (Berlemann, 2009).
Nevertheless, having these bands available for reuse enables the immense economic success of wireless technologies like the popular WLAN IEEE 802.11. Opportunistic spectrum access to under-utilized spectrum, where cognitive radios use flexible spectrum access techniques for identifying under-utilized spectrum and avoiding harmful interference by for example using dynamic frequency selection, whether or not the frequency is assigned to licensed, primary services, is referred to as overlay spectrum sharing. Another spectrum sharing concept discussed in the literature and presented in Figure 2 above is vertical - horizontal sharing. The sharing of a licensed spectrum with primary radio systems is referred to as vertical sharing, and the sharing between radio systems with a similar regulatory priority (as in unlicensed bands) can be referred to as horizontal sharing (Berlemann, 2009).
The main challenges associated with the implementation of the dynamic spectrum access and spectrum sharing techniques are:
Control channel design: Reliable and dynamically changing control channels algorithms are required to ensure control signaling between the two CR users on either end of the link uninterrupted by the neighboring PU activity. This messaging is used to exchange the sensing information and coordinate the channel access.
14 Radio scene analysis and adapting to PU transmission: Some PUs have specific transmission patterns, such as pre-determined spectrum usage times and durations, such as television broadcast stations, or may have occasional random access to the channel, such as public service agencies. At these times, the CR MAC protocol may infer the nature of the PU and adapt its own transmission to avoid both interference to itself and prevent conflict with the PUs. For this reason, dynamic power control and transmission scheduling schemes need to be devised (Cormio, 2009). Also, the PUs transmission and waveform should be identified. Challenges related to radio scene analysis are discussed in Chapter 2 more in detail.
This chapter provides the theoretical foundation and historic perspective to the software-defined radio and cognitive radio communication paradigms used in this work to approach the research problem of efficient frequency spectrum utilization. The main challenges related to radio scene analysis and its hardware implementation are discussed in Chapter 2.2.
2.1 SDR and Cognitive Radio
Software Defined Radio (SDR) is defined as “a class of reconfigurable/reprogrammable radios whose physical layer characteristics can be significantly modified via software changes” (Wyglinski, 2013). The term was first introduced in the 1970s by Joseph Mitola. The first demonstrated SDR prototype ICNIA was presented in 1988. It was a collection of several single-purpose radios in one box. However, the key milestone for the advancement of SDR technology took
Figure 3. Software Defined Radio (Wyglinski, 2013).
place in the early 1990s with SpeakEasy I/II research projects by the U.S. military (Wyglinski, 2013). It used a collection of programmable microprocessors for
16 implementing more than 10 military communication standards, with transmission carrier frequencies ranging from 2 MHz to 2 GHz, which at that time was a major advancement in communication systems engineering.
Cognitive Radio is an SDR based communication systems paradigm that focuses on employing highly agile, environmentally aware, intelligent wireless platforms to autonomously choose and fine-tune device operating parameters based on the prevailing radio and network environmental conditions (Wyglinski, 2013). The term “cognitive radio” was coined by Joseph Mitola in 1999 (Mitola and Maguire, 1999). In 2005, Simon Haykin had given a review of the cognitive radio concept and defined it as brain-empowered wireless communications (Abbas, 2015): “cognitive radio is an intelligent wireless communication system that is aware of its environment and uses the methodology of understanding by-building to learn from the environment and adapt to statistical variations in the input stimuli” (Haykin, 2005). Figure 4 presents three on-line cognitive tasks performed by cognitive radio. The first and second tasks are carried out in the receiver, and the third task is carried out in the transmitter (Haykin, 2005).
1. Radio-scene analysis, which encompasses: estimation of interference temperature of the radio environment and vacant frequency channels detection.
2. Channel identification, which encompasses the following: estimation of channel state information (CSI) and prediction of channel capacity for use by the transmitter.
3. Transmit-power control and dynamic spectrum management. Figure 4. Basic cognitive cycle (Haykin, 2005).
These three cognitive tasks could be mapped into tree control actions: perception, conception and execution (Rondeau, 2009) illustrated in Figure 5.
During perception, sensors collect data on both external factors: channel conditions, other radios, regulations, user needs, and internal factors: waveform capabilities, available computational resources, remaining battery power. This control action is related to both tasks radio scene analysis and identification of the CR’s internal factors. During conception: an intelligent core combines knowledge from the sensing mechanism to aid the adaptation mechanism. This control action is related to channel, link budget estimation, and predictive modeling.
17 Execution control action employs an optimization and adaptation mechanism that adjusts the radio’s behavior towards the environment and user-defined parameters. This control action is related to the tasks of dynamic frequency management and power control.
Figure 5.Cognitive radio: control actions.
Nowadays SDR and cognitive radio technologies are successfully making their way into modern wireless standards. Modern mobile phones have multiple cognitive radio features such as support of multiple waveforms, frequency bands, the ability to transmit various types of data including voice traffic, text messages, and data. They are also equipped with multiple communication standards such as Bluetooth and WiFi and have adaptive power control and modulation adaptation in response to signal quality. In 2009 IEEE 802.22 wireless regional area networks, WRAN has been introduced, and using white spaces in the television frequency spectrum
Table 1. Cognitive Radio Example of Operation Parameters and Optimization objectives.
Device Configurations: Knobs, (Layer)
Performance and Environment: Meters,
Target networking experience: Optimization objectives,
Packet size (NET) Packet rate (NET) Channel coding
rate, type (MAC) Frame size, type (MAC) Duplexing (MAC) Encryption (MAC) Transmitter power (PHY) Modulation type (PHY) Modulation index (PHY) Carrier frequency (PHY) Bandwidth (PHY) Symbol rate (PHY)
Packet delay (NET) Packet jitter (NET) Data rate (MAC) Signal power (PHY) Noise Power (PHY) Delay spread (PHY) Multipath profile (PHY) Power consumption (PHY) Fading statistics (PHY) Spectral efficiency (PHY) Path loss (PHY) BER (PHY)
Minimize delay (NET) Maximize throughput (MAC) Minimize bit error rate (PHY) Minimize power
consumption (PHY) Minimize interference (PHY) Maximize spectral
18 became the first full standard regulating cognitive radio functionality on MAC and PHY for deploying the dynamic spectrum access (IEEE, 2009; Giorgetti, 2012).
The operational parameters in cognitive radio are traditionally described in terms of knobs and meters. This terminology comes from conventional radio terminology. Knobs are defined as parameters affecting the link performance and radio operation. Some of them are assumed to be the design parameters setting the device configuration and some of them are remaining under real-time control. On the other hand, meters describe the system performance and network environment (Bates, 2006). Knobs are configured to reach target network experience in current network conditions. Target networking experience is usually defined in terms of optimization objective, for example: minimize bit error rate or maximize data throughput. To achieve the substantial level of real-time cognitive decision-making AI algorithms are used for perception in radio scene analysis, conception for channel estimation, and during execution in the cognitive controller. The focus of this work has been narrowed down to the perception task and, mainly, on the radio-scene analysis, specifically on detection of the vacant frequency channels, since it is the very first step in cognitive radio operation, defining one of the most critical operation parameters: the frequency channel transmitter and receiver agree to operate in. This work also covers some aspects of the blind waveform estimation of the received signal or PU. Some PUs has specific transmission patterns or may have occasional random access to the channel. The CR MAC protocol may infer the nature of the PU and adapt dynamically its own transmission to avoid both interference to itself and causing harmful interference to PU.
2.2 Spectrum Sensing and Radio-Scene Analysis. Challenges
The main challenges associated with perception control action and radio scene analysis cognitive tasks summarized by Ahmed (2010):
• Hardware Requirements. Software-defined radio-based hardware platform containing the analog to digital converters (ADCs) with high-speed signal processors, high resolution, and dynamic frequency operation range is required for accomplishing the spectrum sensing task fast enough to guarantee required QoS in cognitive radio networks (Arslan, 2007). Our target application and its operational characteristics, as well as the real-time operation requirements, are presented in Chapter 2.3 of this work.
• The Hidden Primary User Problem (Ganesan, 2005). It is caused by multiple factors like shadowing, severe multipath fading observed by the secondary user during the transmission scanning for the primary user: if one PU is “hidden” to a SU during the transmission/reception, some data can be lost when the appearance of the PU is not identified by one of the two SUs. This is particularly serious when the PU affects the receiving side: the receiver could not receive on a channel due to the PU interference while the transmitter continues to send data.
19 • Spread Spectrum Primary Users are the primary users (PUs) using spread spectrum signaling that are hard to identify as the power of the PUs is dispersed over a broad frequency range, while the real information bandwidth is significantly narrower (Cabric, 2004). The spread spectrum has two further types of technologies: frequency hopping spread spectrum (FHSS) and direct sequence spread spectrum (DSSS). FHSS devices adjust their operational frequencies vigorously to multiple narrowband channels. This is called hopping and is performed according to a sequence recognized by both the transmitter and the receiver. DSSS devices resemble the FHSS devices but they utilize a single frequency band or channel to spread their energy. An example of fixed frequency devices is IEEE 802.11a/g. Our target application cognitive radio has FHSS capabilities, operating in frequency range with a maximum possible instantaneous bandwidth of 56 MHz limited by optimal receiver sensitivity.
• Sensing Time and Frequency. The sensing frequency: how often sensing should be performed by the cognitive radio, and sensing time: the duration the sensing is performed are key parameters that should be chosen carefully. Sensing frequency requirements can be relaxed in case the status of the PU is known and changes slowly. For example, the sensing period for IEEE 802.22 standard is 30 seconds. While higher sensing times ensure the correct detection of the spectrum, this may result in a comparatively smaller duration for actual data transmission during the total time for which the spectrum may be used, thereby lowering the throughput. (Cormio, 2009). To achieve the optimum between sensing time and throughput in IEEE 802.22 has a two-stage sensing (TSS) mechanism that includes: fast sensing, done at the rate of 1 ms/channel, and fine sensing performed on-demand. The challenges related to sensing time are discussed in detail in Chapter 4 dedicated to the research goals.
2.3 Target Application Hardware
This chapter provides the brief introduction to our target application: cognitive radio network hardware, its operational characteristics and boundary conditions it applies to the proposed radio scene analysis algorithms to fulfill the real-time operation requirement.
Our target application is a software-defined radio-based network consisting of the digital radio nodes. SDR based nodes operating in the frequency band from 70 MHz to 6 GHz with up to 56 MHz of instantaneous bandwidth are used as the hardware platform for the cognitive functionality deployment with FHSS capabilities. Figure 6 below presents our target hardware, its principal schematics and a photo of transceiver AD 9364. Available computational resources are: Dual-core ARM Cortex A9 CPU, 2x512 MB of DDR3L RAM and 32 MB of QSPI flash memory. Operation system is Embedded Linux. The radio part is based on Analog Devices AD 9364 radio transceiver and Xilinx' Zynq-7020 FPGA. Operational
20 characteristics of AD 9364 are summarized in Table 2. It is optimized for a small size, light weight, and low power consumption allowing integration into power constraint systems and platforms.
Figure 6. Target application hardware, schematics and radio transceiver AD9364.
It is supporting multiple digital modulations including both linear: QPSK, BPSK, QAM, PSK and non-linear: FSK. Symbol rate could be also adjusted to generate cognitive waveform. In this work only three symbol rates have been studied: 10, 100 and 1000 KSymbol/second. Our target application predefines most of the boundary conditions and operational requirements such as required decision-making speed and computational resources available. In this work time required for radio scene analysis
tAnalysis is defined as a sum of time required for radio scene observation tObservation and
processing tProcessing of the observed signal on the receiver front end. Time allocated
for the active data transmission is data transmission time, tTransmission and total time is
the sum of observation and transmission time as illustrated on Figure 7. Table 2. AD9364 Operational Characteristics (Analog Devices, 2013).
TX band: 47 MHz to 6.0 GHz RX band: 70 MHz to 6.0 GHz
DACs and ADCs 12-bit, RF 2 × 2 transceiver integrated Tunable channel bandwidth: <200 kHz to 56 MHz
Dual receivers 6 differential or 12 single-ended inputs Receiver sensitivity noise figure 2 dB at 800 MHz LO
RX gain control Real-time monitor and control signals for manual gain Independent automatic gain control
Dual transmitters: 4 differential outputs Highly linear broadband
TX EVM: ≤−40 Db
TX noise: ≤−157 dBm/Hz noise floor
TX monitor: ≥66 dB dynamic range with 1 dB accuracy local oscillator (LO) step size 2.4 Hz maximum
t Analysis t Transmission
Figure 7. Optimizing sensing and transmission times.
Allocating the sensing time and transmission time at the MAC layer is involving a tradeoff between ensuring to the PUs user’s QoS requirements opposed to maximizing the data throughput. To meet real-time operation requirement on the target hardware the following real-time performance characteristics must be met by the proposed algorithms:
- Radio-scene observation time is tobservation = 500×10-6 seconds
- Processing time: tprocessing = 500×10-6 seconds.
Modulation classification must be performed for SNR values above the demodulation threshold of 12 dB, which corresponds to BER=10-8 and BER= 3.4×10-5 for 2FSK and BPSK respectively.
3 Related Work
Chapter 3 provides an overview of research dedicated to radio scene analysis. Algorithms used by other researchers for blind detection of vacant frequency channels including matched filtering, waveform-based sensing, cyclostationarity-based sensing, energy detector-cyclostationarity-based sensing, wavelet transform, and AI cyclostationarity-based algorithms are discussed in Chapter 3.1. Chapter 3.2 describes algorithms used for waveform estimation including non-linear energy operator, inverse fast Fourier transform, maximum likelihood, wavelet transform and cyclostationary for symbol rate estimation and cyclostationary, maximum likelihood estimation, wavelets, AI based algorithms for modulation type classification using various features extracted from the received signal such as for example, statistical features.
3.1 Blind Detection of Vacant Frequency Bands
Although precise and timely channel state information is desirable for spectrum access and primary user protection, continuous full-spectrum sensing is both energy inefficient and hardware demanding (Mao, 2014). There are various algorithms described in the literature, some of them perform screening-like estimation and others do more detailed detection based on channel modeling or demodulation. Yucek and Arslan have listed waveform/based sensing and matched filtering as the methods with the highest accuracy. These methods have not been studied in the scope of this work since they are not autonomous or blind and require demodulation and prior knowledge about the primary user signal: bandwidth, operating frequency, modulation type and order, pulse shaping, frame format (Yucek and Arslan, 2009). Cyclostationarity-Based Sensing. The signal detection is performed based on the cyclostationary property of periodicity detection in the digital signals. Modulated signals are cyclostationary: i.e. contain the periodicity in its statistics (Gardner, 2006). Periodicity in the mean is referred to as the first order cyclostationary, periodicity in autocorrelation is referred to as second-order cyclostationary. Common analysis of stationary random signals is based on the autocorrelation function and power spectral density. Jerjawi et al. (2015) applied cylostationary based sensing to develop an algorithm for LTE SC-FDMA signals detection. Jang et al. (2018) have proposed a blind spectrum sensing method using signal cyclostationarity. Munpreet
23 and Gagangeet (2014) classify cyclostationary as an effective algorithm for signals with low SNR. Also, it can differentiate between noise energy from received signal energy and can be used to find out the modulation type of the received signal. Among the named disadvantages: it requires longer observation times; it is computationally complex and detection performance degrades due to the poor estimate of cyclic spectral density.
Energy Detector-Based Sensing. The signal is detected by comparing the output of the energy detector with a threshold that usually depends on the noise floor: power measurements above this threshold are identified as signals. It requires low computational complexity and relatively easy to implement. It also allows blind detection: receivers do not need any knowledge about the primary users’ signals. However, the main practical challenge of energy detector-based sensing is the determination of the threshold. The decision threshold can be estimated in several ways such as 1. An empirical analysis of data (Ellingson, 2005; McHenry, 2005); 2. Computation of threshold from system properties such as noise floor (Ellingson, 2005); 3. Using a priori knowledge of the noise statistics (Govaert, 2009; Ellingson, 2005), and 4．Estimation of threshold directly from data (Yucek and Arslan, 2009). Another disadvantage mentioned by Munpreet and Gagangeet (2014) is the degradation of performance under low SNR or with noise uncertainty; it is classified as not suitable for spread spectrum techniques like direct sequencing and frequency hopping.
Wavelet Transform. The wavelet approach offers advantages in terms of both implementation cost and flexibility for signal detection over wideband channels, in adapting to the dynamic spectrum in contrast to the conventional implementation of multiple narrowband bandpass filters (BPF) (Tian and Giannakis, 2006). It could be used for both the symbol rate and modulation type estimation. M.K. Lakshmanan et al. have demonstrated the operation of a Wavelet Packet-based multi-carrier modulation (WPMCM) transceiver for dynamic spectrum access. The WPMCM receiver structure, which is used for demodulation of data, is also used for the analysis of the radio environment to identify active/idle bands - at no additional cost (Lakshmanan, 2009).
AI based Algorithms. Recent advancements in electronics and GPU design have made AI based algorithms an attractive tool to solve vacant band detection problem. AI based algorithms such as deep and machine learning-are capable of implicitly learning the surrounding environment. They can reliably detect PU activity without requiring any prior knowledge of the environment or PU. Nie and Haykin, (1997) have pioneered applying reinforcement learning to spectrum allocation problems. They have developed a mathematical model for DSA problems based on the Q-learning algorithm, which is the best-known algorithm among reinforcement Q-learning techniques (Abbas, 2015).Multiple improved learning algorithms, deep
network based spectrum allocation algorithms and other recently proposed deep learning algorithms are summarized by Wang (2019). Nowadays, large labeled data sets such as Nsynth are available for training the deep networks. (Thilina, 2012; Mikaeil, 2014) have used supervised machine learning KNN for spectrum allocation. Markov chain has been used by M. Pesce et al. (2016) to model correlation patterns of the primary users’ activity. Channel modeling has extensively used Markov models in research. Probably the most famous is the two-state Gilbert-Elliot model (Oksanen and Koivunen, 2015) that describes a channel as in either vacant or occupied (Rondeau, 2009). Mark (2018) have proposed an algorithm identifying the primary user activity over a wide spectrum band and provides a statistical characterization of the primary user signals in the band. The algorithm applies hidden Markov modeling to a hierarchically partitioned representation of the spectrum band, together with a recursive tree search (Mark, 2018).
3.2 Waveform Estimation
Waveform estimation consists of estimation of three signal parameters: symbol rate, modulation type and central frequency.
3.2.1 Blind Symbol Rate Estimation
There are various algorithms suggested for blind symbol rate estimation described in the literature including non-linear energy operator (Khan S.Z., 2011), cyclostationary (Phukan and Bora, 2014; Julien, 2010; Guner and Kaya, 2011; Suesser-Rechberger and Gappmair, 2018; Stuwart, 2015; Jallon, 2008; Elgenedy, 2013; Majhi, 2015; Sheng-en, 2010), Inverse Fast Fourier Transform and baseband shape based (Xu, 2007; Flohberger, 2006); Monte Carlo based (Stuwart, 2015); Wavelet Transform based (Sawai, 1999; Xu, 2012; Xu, 2005; Sun, 2008; Chan, 1997; Hatoum, 2014) and Maximum Likelihood (Lodge and Moher, 1990; Yu and Pasupathy, 1995). Most of the symbol rate estimation algorithms are applicable for the certain modulation types. For example, for the linear modulations often cyclostationary and wavelet transform based algorithms are suggested. Gardner’s method also based on signal cyclostationarity (Gardner, 1988) was intended for BPSK and QPSK (Al-Hamiri, 2017). Non-linear modulations including FSK, however, received significantly less attention in the literature. Wavelet transform has been listed by Hatoum et al. (2014) as a high capacity algorithm to detect discontinuity structures and zoom on the signal abrupt changes. While the traditional symbol rate estimation methods, such as the cyclostationary and the envelope spectrum method, show a dramatic decline of performance in low SNR, the wavelet transform method proposed by Z. Yang et al. has demonstrated high performance at low SNR for FSK modulated signals (Yang, 2019).
3.2.2 Modulation Type Classification
The literature study on the modulation classification topic includes both aspects classification algorithms and features used for classification. Modulation classification features could be classified into spectral-based and cyclostationary features. The spectral-based features exploit the unique spectral characters of different signal modulations in three physical aspects of the signal: the amplitude, phase and frequency. Zhu et al. (2015) have summarized some of the well-recognized spectral features designed for modulation classification including the following and suggested a decision tree classification described by Figure 8 (Zhu, 2015):
1. the maximum value of the spectral power density of the normalized and centered instantaneous amplitude of the received signal (Azzouz and Nandi, 1996a),
2. the standard deviation of the absolute value of the non-linear component of the instantaneous phase,
3. the standard deviation of the non-linear component of the direct instantaneous phase,
4. an evaluation of the spectrum symmetry around the carrier frequency, 5. the standard deviation of the absolute value of the normalized and centered
instantaneous amplitude of the signal samples,
6. the kurtosis of the normalized and centered instantaneous amplitude,
7. the standard deviation of the absolute value of the normalized and centered instantaneous frequency,
8. the standard deviation of the normalized and centered instantaneous amplitude,
9. the kurtosis of the normalized and centered instantaneous amplitude.
Kubankova et al. (2011) have applied instantaneous amplitude, instantaneous phase and spectrum symmetry together with the set of new features from both spectral and time domain including: Linear Predictive Coefficients, Adaptive Component Weighting, Zero-Crossing Ratio, Linear Frequency Bank Spectral coefficient for the classification of commonly used digital modulations including ASK, FSK, MSK, BPSK, QPSK, PSK, FSK4 and QAM-16. A classifier based on Gaussian mixture models was used to analyze the features and classify the modulations. Average classification accuracy of 96.13 % has been reached using the optimal set of features. Azzouz et al. (1995) have demonstrated modulation classification accuracy of 90 % at SNR = 10 dB using the spectral-based features.
Cyclostationary based features. Gardner pioneered the area of cyclostationary signal analysis (Gardner, 1994). Gardner and Spooner first implemented cyclostationary analysis for modulation classification problems in (Gardner, 1988). Spectrum correlation feature is often used as a second order cyclostationary feature to classify between manmade modulated signals that contain periodicity and noise.
26 Maximum likelihood estimation is not classified as a blind method since it requires channel state information but some adaptations to the non-cooperative environment exist. Xu et al. (2011) have performed an extensive survey on the automatic modulation classification methods based on likelihood functions. The common approach of a likelihood-based modulation classifier consists of two steps. In the first step, the likelihood is evaluated for each modulation hypothesis for the observed signal samples. The likelihood functions are derived from the selected signal model and can be modified to fulfil the need of the reduced computational complexity or to be applicable in non-cooperative environments (Zhu, 2015). During the second step, the likelihood of different modulation hypothesizes are compared to provide the classification decision. Dulek, (2017) have proposed two online methods facilitating close to optimal classification performance with reduced computational complexity are employed for modulation classification over unknown nonidentical flat block-fading additive white Gaussian noise channels using multiple sensors.
Figure 8. Decision tree for modulations classification using spectral-based features (Zhu, 2015).
Wavelets are classified as relatively low complexity algorithms suitable for blind modulation classification. Continuous wavelet transforms of the received signal is defined as the integral of times the conjugate transpose of the wavelet function over time. Among different mother wavelet functions, namely Morlet, Haar and Shannon, most researchers selected the Haar wavelet function because of its simple form and
27 computational convenience. The performance comparison of different wavelet families for classification of chaos-based digital modulation techniques using wavelets has been performed by Türk, (2011). Kumar et al., (2017) have proposed denoising method based on the continuous wavelet transform (CWT) with Haar wavelet.
Machine learning and deep learning are popular AI based algorithms in the literature. The practical deployment of autonomous machine learning (ML) based classifiers is difficult due to relatively high computational complexity, also they are using the manually extracted features, what requires some expertise. ML classifiers are typically classified into supervised and unsupervised. Supervised machine learning requires the training data set to be labeled, while the unsupervised are performing clustering of unlabeled data and have proved to be efficient in modulation type classification at low SNR. Unsupervised machine learning algorithms have proven to give above 90 % accurate spectrum sensing rate (Awe, 2015) when they have been tested in highly interfered environment. Bari et al. (2015) have been using self-organizing maps (SOM) and support vector machines (SVM) based on three features: the standard deviation of the signal, the standard deviation of the time averaged signal and the standard deviation of the derivative of the signal to perform classification of FSK, BPSK and 16-QAM. In this work we have applied multiple supervised machine learning for modulation classification. The summary of the several classifier families is given below.
Decision trees is a very popular supervised learning algorithm in data mining due to its simplicity and transparency (Rocach and Maimon, 2008). Classification is based on classes, features and attributes. According to Mathworks (2020) it is referred as fast in terms of prediction speed and has relatively small memory usage.
Support vector machines (SVM) is a supervised machine learning algorithm, that performs classification based on the support vector learning algorithm (Haykin, 1999) as a product of the support vector, drawn from the input space. The support vectors consist of a small subset of the training data. There are multiple types of learning machines depending on the kernel method: polynomial, radial basis function networks, two-layer perceptron’s, etc. According to Mathworks (2020) SVM is referred to be slow in terms of prediction speed for multiclass predictions and has medium to large memory usage.
k-Nearest neighbor (KNN) is a supervised learning algorithm based on searching nearest k samples from the existing training data when a new sample appears and classifies the appeared sample according to the most similar class (Mitchell, 1997). Mathworks, (2020) has classified it as medium to slow depending on the kernel type in terms of the prediction speed for and has medium memory usage.
28 Ensemble methods are supervised learning algorithms that use multiple learning algorithms to obtain better predictive performance than could be obtained from any of the constituent learning algorithms alone. There are two most common sampling methods in ensembles: bagging and boosting. In bagging each model is trained in parallel with the others in the ensemble using a randomly chosen subset of the training data set. Boosting involves incrementally building an ensemble by training each new model instance to emphasize the training instances that previous models misclassified (Govaert, 2009). Ensemble bagged trees are referred to as medium in terms of the prediction speed and boosted trees are fast in terms of the prediction speed and have high and low memory usage respectively (Mathworks, 2020).
Discriminant analysis is a supervised learning classification algorithm. It assumes that different classes generate data based on different Gaussian distributions. During the training process, the fitting function estimates the parameters of a Gaussian distribution for each class. To predict the classes of new data, the trained classifier finds the class with the smallest misclassification cost (Mathworks, 2020). The discriminant is considered as fast classification method with the small memory usage.
Logistic regression. The classifier models the class probabilities as a function of the linear combination of predictors. It is listed as a fast classification method with medium memory usage (Mathworks, 2020).
The recent advancements in electronics and GPU design have opened a possibility of applying end-to-end deep learning using convolution neural networks (CNN) to solve waveform estimation problem. Hands on practical implementation of deep learning-based modulation classification has been demonstrated by Mathworks in (Mathworks, 2020) and Peng et al. (2017).
4 Research Goals
In this chapter, the main research goals are defined. From the subset of two radio scene analysis tasks performed by the cognitive radio for implementation of the opportunistic spectrum access, two hypotheses are formulated: the first is related to the vacant band identification and the second to waveform estimation. Seven research questions related to the trade-off between the sensing accuracy and real-time operation requirement for the proposed radio scene analysis algorithms, the nature of the noise, and assumptions used to model the radio scene environment such as, for example, the AWGN channel model.
It is possible to perform the real-time detection of the vacant frequency channels in the wide band of 60 MHz (effective 56MHz) within the observation time of 500 microseconds or less.
This hypothesis is related to the perception action of the CR: real-time radio-scene analysis. There are multiple objectives behind the spectrum allocation problem in CR networks summarized by Wang (2019):
1.Spectrum efficiency; 2.Power efficiency; 3.Interference minimization; 4.Fairness optimization; 5.Througput maximization; 6.Connectivity optimization; 7.Delay minimization; 8. Reliability maximization.
In the scope of this work, multiple algorithms proposed in the literature for blind detection of vacant frequencies including energy detectors, wavelets, and cyclostationary have been studied and compared. Also, three main research challenges related to the radio scene analyses and vacant channel detection identified in the literature have been studied and addressed: 1. The threshold value for the vacant frequency channels detection (Cormio, 2009). There are multiple approaches discussed in detail in Chapter 6. In this work, the level of the threshold for the vacant band detection has been estimated empirically from the captured noise
30 data in the wideband spectrum. However, this empirical approach is limiting the results of this study towards our hardware. 2. Optimal sensing duration: In our time-slotted target application the maximum radio scene environment sensing duration corresponds to 500 microseconds what is the size of one time slot. Allocating the sensing and transmission time at the MAC layer is involving a trade-off between ensuring the PUs user’s QoS requirements as opposed to maximizing the data throughput (Cormio, 2009). 3. Sensing coordination and sharing of the sensing information: for accurate sensing, the measurements on the channel must be undertaken during the time periods when the other CR users in the neighborhood are quiet to ensure correctly attributing the measured power to the radiation caused by the PUs alone. This coordination is difficult to achieve at the MAC layer in the absence of time synchronization. For accurate sensing, the spectrum sensing results from several different CR users may be integrated. However, the exchange of this information may possibly interfere with the data packets sent by the other CR users (Cormio, 2009). The cooperative sensing by multiple CR nodes is not covered in the scope of this work. In this work, energy detectors, wavelets, and cyclostationary based algorithms have been adapted to perform the initial detection of vacant frequency channels based on the input signal captured on the receiver front end in the wide band of 56 MHz (maximum usable instantaneous bandwidth of AD9364 transceiver).
It is possible to perform autonomous (blind) waveform estimation including symbol rate, the modulation type and central frequency of the received signals to deploy dynamic spectrum access using resources available on the target hardware AD9364 within the observation time of 500 milliseconds or less.
This hypothesis has been formulated from the subset of both research and engineering problems. It is related to the perception action and radio-scene analysis task of the cognitive radio: waveform estimation of the active transmitting signals in the channel. Waveform estimation requires the estimation of tree signal parameters: the modulation type, the symbol rate, and the central frequency. Once the waveform is known, there is a possibility to employ the dynamic spectrum access for spectrum reuse by adjusting transmitters waveform to occupy the vacant frequencies without causing harmful interference to the primary user’s transmission. A literature review dedicated to modulation classification has shown that most of the proposed algorithms are based on statistical features derived from the received signal. In the scope of this work, we have tested whether the instantaneous values of the in-phase and quadrature components of the time domain received signal could be utilized for modulation classification directly to achieve at least 85 % accuracy without any pre-processing and feature extraction. As a first step, in Paper 1, the classification has been performed between non-linear modulation 2FSK and linear BPSK. Modulation type classification has been performed using supervised ML algorithms. The suggested workflow for waveform estimation could potentially contribute to the
31 improvement of the throughput by shortening the sensing time for 2FSK modulated signals. Since we are not doing any pre-processing by using the instantaneous values during the fast modulation type estimation: this could potentially lead to shortening of the signal sensing time, meaning that more time is going to be available for the active transmission, what could potentially result in the increased throughput. Also, some of the algorithms suggested for symbol rate estimation in the literature are more suitable for certain modulation types.
Figure 9. Waveform estimation algorithm
Estimated waveform for DSA Input Signal PU or other CRs Waveform:
- Modulation - Symbol Rate - Center frequency
Linear? Fast modulation classification. Classification features:
- instantaneous values of in-phase and quadrature components - RSSI
Fine modulation classification Classification features:
- features extracted from the time domain signal time series recorded during observation time
- RSSI QPSK QAM 8PSK 16PSK 2FSK
Symbol Rate Estimation: -Wavelets
Symbol rate estimation: -Cyclostationary
32 For example, symbol rate could be estimated using cyclostationary-based algorithms for the signals with linear modulations such as Gardner’s method intended for BPSK and QPSK. Symbol rate of FSK modulated signals could be estimated using wavelets, that is less computationally demanding and may require less observation time. Therefore, the fast estimation of the modulation type may potentially allow to select more accurate and less computationally complex symbol rate estimation algorithm for the identified modulation type. Figure 9 presents the proposed waveform estimation algorithm. The possibility of using instantaneous values for the fast modulation classification has been investigated in this work. However, the second step of modulation classification is not covered and is going to be carried out in the scope of the future work.
4.2 Research Questions
RQ1: How to achieve an accurate estimation of the modulation type of the received
signal within the spectrum observation time of 500 microseconds or less?
This research question covers both the selection of the input data to be used for modulation classification and the algorithms to be used. Could the instantaneous values of the time-domain received signal captured as in-phase and quadrature components on the receiver front end be utilized for modulation type classification directly or is pre-processing or feature extraction from the time-sampled signal is required to reach classification accuracy of 85 % or higher? What classification accuracy could be reached in conditions of AWGN channel with SNR varying from 30 to 1 dB for classification based on instantaneous values? Also, we have available SNR value as a measured RSSI level by the transceiver. Could it be a valuable input parameter for the modulation classification? Is it possible to reach the target classification speed of 2000 objects/s? To approach this set of questions the study has been conducted to test the most common supervised ML algorithms and compare them in terms of accuracy and classification speed. The details are described in Chapter 6 “Contributions”, and in the included Paper 1.
RQ2: How to perform an accurate estimation of the received signal’s symbol rate? In the scope of this research question various algorithms described in the literature for the symbol rate estimation have been studied and compared including cyclostationary based and wavelet transform. Also, the effect of SNR on the symbol rate estimation accuracy has been studied. The results of this study for FSK modulated signals are summarized in Paper 2 “Blind Symbol Rate Estimation for Efficient Cognitive Radio Using Wavelet transform and Deep Learning for FSK Modulated Digital Signals”.
RQ 3: Is AWGN nature of noise an appropriate assumption when it comes to RQ1
33 We have elaborated on the noise sources existing in the frequency channel according to their origin, nature, and statistical properties. The statistical properties of the received signal captured on the receiver’s front end have been compared with the statistical properties of AWGN. Also, second order cyclostationary properties of captured AWGN have been studied including the spectrum correlation.
RQ4: How to perform the real-time detection of the vacant frequency channels
within the observation time 500 microseconds or less?
While higher spectrum observation times ensure more accurate radio spectrum sensing, this may result in a comparatively smaller duration for actual data transmission in the total time for which the spectrum may be used, thereby lowering the data throughput. The implementation of the radio scene environment sensing in the IEEE 802.22 protocol has a two-stage sensing (TSS) mechanism (Cormio, 2009). To reliably attribute the source of the received power to the PUs, the standard enforces quiet periods throughout the CR network called channel detection time. The TSS consists of two-stages with different durations and goals: fast sensing is done at the rate of 1 ms/channel, and the sensing results are used to decide if a subsequent fine sensing stage is needed. The sensing is completed quickly though the accuracy is low. Fine sensing is performed on-demand, which allows CR networks to meet the strict quality of service (QoS) requirements by decreasing the rate of false alarms. The duration for this is much larger than the fast sensing and gives a tradeoff for improving the sensing accuracy at the cost of transmission time. In this work, maximum observation time has been selected 500 microseconds to meet real-time operation requirements for the time slotted communication system, where 500 microseconds correspond to one time slot. Is it possible to reach 85 % vacant channel detection accuracy from the received signal observed during 500 microseconds? The results of the vacant channel detection are summarized in the included Paper 3. However, more research work dedicated to vacant channel detection and effect of the radio scene observation time on the vacant channels detection accuracy is planned in the scope of future work. Also, the potential of the application of two-step spectrum observation (fast and fine scanning) like in IEEE 802.22 is going to be studied in the scope of the future work.
RQ5: Could the vacant channels detection be performed in the wide band of 56-60
MHz corresponding to all available instantaneous bandwidth for our transceiver?
Our target application could be operated in 56 MHz instantaneous bandwidth, modeled as a wideband channel divided into K-non-overlapping narrowband subchannels. In a particular geographical region and time, some of the K-subchannels might not be utilized by the primary users and are available for opportunistic spectrum access. The advantage of the wideband detection is the potential to maximize opportunistic throughput. Spectrum sensing in CR networks could be done as primary transmitter detection, primary receiver detection, or cooperatively. The primary transmitter detection is based on the detection of the weak signal from a primary transmitter through the local observations of CR users.
34 The primary receiver detection aims at finding the PUs that are receiving data within the communication range of a CR user. Cooperation allows independently faded radios to collectively achieve robustness to severe fades while keeping individual sensitivity levels close to the nominal path loss. Furthermore, a small number of radios (10–20) are enough to achieve practical sensitivity levels (Mishra, 2006). However, this work has been limited to single radio sensing and cooperative sensing has been suggested for future work.
RQ6: How to determine the threshold or the noise floor for the vacant frequency channel detection?
During the literature review, multiple ways of the threshold estimation have been studied and summarized in Chapter 3, dedicated to the previous work. The empirical approach for the threshold estimation has been chosen since it provides the most realistic estimation for the fixed input power level not calibrated at the receiver. Receiver Operating Characteristics (ROC) has been plotted for multiple thresholds and the optimal value has been selected as a compromise between the hit rate and the number of false-positive detection. This method, however, limits the results towards the transceiver hardware used in our target application. The selected value of the threshold cannot be lower than the noise floor. The noise floor is predefined by multiple factors such as the bandwidth, the sample rate, amplifier gain, modulation The noise floor is predefined by multiple factors such as the bandwidth, the sample rate, amplifier gain, modulation.
RQ7: Could one set of feature extraction algorithm and classifier be used to answer multiple questions about radio scene environment?
Many feature extraction algorithms have been applied in the literature to solve multiple radio scene environment sensing tasks in the CR, for example, wavelet transform has been applied for vacant band detection, modulation classification and denoising. Cyclostationary detectors have been also applied for modulation classification, vacant bands detection, and symbol rate estimation. Could the binary hypothesis testing used to model the vacant band detection problem be extended to M-ry hypothesis testing where the same set of features and the same algorithm applied to accomplish both radio scene environment sensing tasks: vacant frequency channel detection and waveform estimation. In the scope of this work the study has been limited towards the literature study of the feature extraction algorithms their performance and application potential.
5 Research Methodology
This chapter describes the research methodology followed in this work to answer research questions and test formulated hypothesis. Also, methods used for data collection are summarized here. The main objectives of this work are proposing and virtual prototyping the subset of radio scene analysis algorithms, including 1. detection of vacant frequency channels to implement FHSS scenario; 2. waveform estimation including modulation type, symbol rate, and central frequency estimation for our target application: real time portable power constrained CR application. Therefore, the proposed algorithms are required to demonstrate compatible with real-time operation performance in terms of speed, computational complexity, and accuracy. The deductive approach presented on Figure 10 has been used to address the research questions. Primary the extensive literature review has been conducted to identify the state of the art and possible contributions, evaluate methods, tools and theoretical foundations used to address the research questions by the other researchers. Some of the research questions have been extensively addressed in the literature and are even described in the wireless standards for cognitive radio such as IEEE 802.22.
36 Multiple research questions have been formulated to identify the radio environment our application is operating and propose the algorithms providing the best performance in terms of operational speed and accuracy.
Two hypotheses have been formulated: the first one is related to identification of the vacant frequency channels and the second to real time waveform estimation. The hypothesis implementation and testing has been limited to the virtual prototyping in the Mathworks environment in scope of this work.
Research methods used for data collection include both the software simulation and controlled experiment. Artificially generated data sets have been used for training and testing of the proposed machine and deep learning algorithms for waveform estimation. Artificial data has been generated to simulate the AWGN channel only with SNRs ranging from 1 dB to 30 dB. No fading phenomena has been considered. Algorithms proposed for the vacant frequency channels detection have been tested on data collected during the controlled experiment by our target application hardware.
6 Thesis Contributions
Chapter 6 summarizes the main scientific contributions of this work including the study of the radio scene environment and nature of the noise, proposed radio scene analysis algorithms for vacant channels detection and waveform estimation. Chapter 6.1 provides the summary of the included papers and Chapter 6.2 discusses the relationship between the included papers and the research goals.
6.1 Included Papers
The order of papers is chronological.
Paper P1 “Multiple Machine Learning Algorithms Comparison for Modulation Type
Classification for Efficient Cognitive Radio”
Published, IEEE MILCOM, 2019.
Paper P2 “Blind Symbol Rate Estimation for Efficient Cognitive Radio Using
Wavelet transform and Deep Learning for FSK Modulated Digital Signals”
Paper P3 “Blind Detection of Vacant Frequency Bands for Cognitive Radio using
6.2 Relationship Between Papers and Research Questions
Contribution 1 is dedicated to the study of the internal noise nature and the research question 3. Contribution 2 is dedicated to blind vacant frequency channels detection and research questions 4, 5, 6, 7 elaborated in paper 3. Contribution 3 is related to the waveform estimation and research questions 1, 2, 3, 7 elaborated in Paper 1 and Paper 2.
Table 3: Mapping of the research questions to the included publications Paper 1 Research Questions 1, 3, 7; Hypothesis 2 Paper 2 Research Questions 2, 3, 7; Hypothesis 2 Paper 3 Research Questions 4, 5, 6, 7; Hypothesis 1