Spectrum Management In The Context Of Cognitive Radio: Measurements, Analysis And Modeling

Marconi’s demonstration at the turn of the20th century revolutionized the way wecommunicate forever while also ushering the birth of radio communication. The growthof radio communication highlighted the problem of interference amongst transmittersoperating in the same geographical area. Earlier efforts aimed at resolving this issuehighlighted the importance of spectrum as a scarce and renewable natural resource.Regulatory bodies were then set up to manage the planning, allocation and overallmanagement of this scarce resource at the national level. Currently, the InternationalTelecommunication Union (ITU) along with the representatives of each country’sregulatory body converge at the World Radio Conference (WRC) periodically to mapout policies for international and regional usage as well establishing world standards.Spectrum management since the early days of radio communication has alwaysbeen about minimizing interference. Under this regime, spectrum bands are allocated toa particular service over a large geographical area in some cases a country, withprovision for band guards to alleviate the problem of interference. This approach hasbeen very efficient as far as maintaining non-interference communication is concerned.However, with the proliferation of several wireless standards over the years, the demandfor spectrum has grown tremendously so is its economic value. It has therefore becomenecessary to review the current command and control approach to spectrummanagement. One of the solutions being proposed is a regime whereby licensed andunlicensed users could share the spectrum in a non-interference fashion; this is knownas Cognitive radio. It was first envisioned by Mitola in1999. To realize this concept,the secondary user also known as the unlicensed user must be able to sense the channelprior to initiating transmission and also vacate the channel in the event the primary useror licensed user returns to transmit. This process termed as spectrum sensing is thecornerstone of successful deployment of cognitive radio.The successful deployment of the cognitive radio paradigm hinges on conciseknowledge of the spectrum; its behavior, usage patterns, and the availability ofspectrum holes. It is generally hoped in the research community that cognitive radiowill dramatically reduce spectrum usage inefficiency thereby increasing spectrumutilization. Knowledge of the spectrum has therefore become the first step towardsachieving this goal. Field measurements of the radio environment have been performed over the years to determine frequency channel usage information for national regulatoryauthorities which they have been using for formulating spectrum usage policies wherethe sole aim is interference minimization. These measurements provide a thoroughknowledge of network activities that can provide a platform for understanding the usagepattern which is invaluable in cognitive radio deployment. Prior to the cognitive radioera, the main aim of spectrum occupancy measurements have been to provideinformation for policy makers and regulatory bodies to formulate policies as well asminimize interference respectively. However, with the advent of cognitive radioresearch, the aim of spectrum occupancy measurements have broadened to includequantitative information on underutilized spectrum bands for possible cognitive radiodeployment and also the establishment of the degree of efficiency of spectrum usage.The economic and research communities might have different motives for conducting aspectrum occupancy measurement, with the economic community mostly interested inthe investments required to develop cognitive radio so that the investments are notmisapplied if they do not take into account the current realities. On the hand, theresearch community is mostly interested in determining the bands that experience lowutilization so that they can be analyzed and characterized in time, frequency and space.This information will be vital for the deployment of future cognitive radio systems.Current works on cognitive radio have all been theoretical with no practicaldemonstration as to the real applicability of these schemes. Spectrum measurement datacan be used to actually determine the feasibility of these schemes. Another importantapplication of these measurements lies in the development of spectrum predictionmodels for deployment in cognitive radio engines with the aim of saving both energyand sensing time. It can therefore be said that successful deployment of cognitive radiorequires a thorough understanding of the spectrum, spectrum occupancy statistics fromthe measurements and reliable models capable of predicting future usage with little orno errors.Spectrum prediction models have been researched over the years, where modelswere developed for the High Frequency HF bands. Due to the opportunistic nature ofspectrum, successful deployment of these models will depends on the performance ofthe secondary networks which will heavily rely on the spectrum usage patterns of theprimary user. It has therefore become necessary to investigate the nature and pattern ofspectrum usage so that successful models could be furnished. Recently, models aimed atpredicting the spectrum opportunities have been presented with varying degree of accuracy. Part of the objective of this work presented herein is to contribute to thiseffort by improving the prediction accuracy of these models.The first contribution of this work is to provide an insight into the spectrum usagein Nigeria with the aim of providing a platform for future measurements and thepossible deployment of the cognitive radio paradigm. Several spectrum occupancymeasurements have been performed all over the world; USA, United Kingdom, China,Singapore, Japan, are some of the countries were these measurements have been made.Few of these measurements can be found in Africa with the exception of South Africain the research community. It has therefore become necessary to perform thesemeasurements for the socio-economic benefit of the populace. The deregulation of thetelecommunication sector in Nigeria in2001has lead to a remarkable growth in thewireless communication services sector especially cellular communication. With thelaunch of the country’s first satellite the Nigersat-1which is basically a remote sensingsatellite the government signaled its intention to develop this vital area of the economy.NigComSat-1R a communication satellite was re-launched in2011after earlier attemptsin2007failed. The satellite was launched in partnership with China Great Wall IndustryCorporation with the aim of further improving the communication infrastructure.Spectrum scarcity has been a dilemma recently; African countries suffer more in thisregard because of their heavy reliance on wireless communication services due to non-availability of other vital infrastructure such as copper and fiber optic networks.Therefore it could be said that the transmission channel is not readily available thus theovercrowded nature of the spectrum allocated for broadband communication. InNigeria’s case, the population of over150million and over100million active lines hasaffected the network quality of these services, to increase quality, capacity must beincreased, to increase capacity there must be more spectra. Even though other servicescould be considered, the affordability across users must be taken into consideration asmost users won’t be able to afford expensive services. These factors and others lead tothe African region headed by Nigeria to demand for further spectrum allocation duringthe World Radio Conference2012. They argued that the allocation will improve thequality of broadband services and also increase mobile internet penetration amongst thepopulace. Currently, the ITU has granted the700MHz for efficient delivery ofbroadband services for this region. It has therefore become extremely important tostudy the spectrum utilization, behavior, and analyze the data obtained from these measurements in order to formulate and plan for the future due to the uniqueness ofNigeria’s case.The measurement setup involved an Aaronia AG HF-6060V4spectrum analyzerwith a range of10MHz-6GHz, an Aaronia AG OmniLOG90200antenna with a rangeof700MHz to2.5GHz, a laptop system that is connected to the spectrum analyzer via aUSB cable, and an MCS software specially designed to run on Aaronia AG spectrumanalyzers. The MCS software has a graphical interface whereby all the parametersrequired for accurate measurement are set. The software is also used for controlling themeasurements based on the parameters set. The measurement was conducted indoors inprimarily three different locations; at Gwarinpa District a primarily residential district inAbuja the capital of Nigeria, Wuse Zone4a residential/commercial activity centre alsoin Abuja, and Gafai quarters which is part of the ancient city of Katsina, a town in thenorthern part of Nigeria.. The measurement conducted in Gwarinpa Estate Abuja wasconducted during the summer holidays of2012i.e. July-August2012covering the700-2400MHz band over12hour periods. The measurements were further divided into twocategories i.e. daytime and nighttime. Daytime measurements cover the duration from9am-9pm while nighttime covers the duration from10pm-8am. Gwarinpa estate isapproximately19km from Nnamdi Azikiwe International Airport; we could therefore beable record some of the activity in the aeronautical bands due to the relatively closeproximity of the location to the airport.The second location was situated in the centralbusiness district of Abuja. It contains most of the government and businessestablishments with some residential areas sparsely situated. The dataset from thislocation consists of data from700-1200MHz band also obtained during the summer of2012(July-August2012). The third and last dataset was obtained from Gafai quarters inthe ancient city of Katsina a densely populated area of the ancient town. Themeasurements were conducted in the summer of2013(August) over the700-1000MHzband over a period of12hours from8am-8pm. The first band considered was the700-1000MHz. It comprises the800MHz band used for trunk radio services, emergencyservices, CDMA (fixed),900MHz for GSM and also the470-960MHz for analoguetelevision broadcasting. In the VHF band there are12channels where as the UHF bandconsists of49channels making a total of61channels. This band has the highestutilization level experienced at26%due to the activities of the analogue broadcasting(part of it to be precise) GSM operations and the radio trunk services. The1000-1500MHz band is mostly used for microwave point to point communication(1350-1550MHz), government agencies and oil companies in the Niger delta region andLagos. The1000-1300MHz and1300-1500MHz with a utilization level of2.13%and1.85%respectively are among the bands with the lowest utilization level. Apart frommicrowave point to point transmission observed around1350-1450MHz; there isvirtually no activity at all.Above1.5GHz, majority of the utilization can be observed in the3G mobilestandards. With a utilization level of around25.1%it has one of the highest utilizationlevel amongst the bands considered for this work. In Nigeria, there are currently fivemobile companies delivering3G mobile services in Nigeria: MTN, Globacom, Airtel,Etisalat and Starcomms. Networks employing UMTS use WCDMA technology, thespread spectrum nature of the signals where by the signals are modulated over a widebandwidth thus making them having a noise-like character due to the very lowtransmission power makes them difficult to detect. This makes it difficult for thespectrum analyzer to determine such signals. Similarly, since the measurements’ wereconducted indoors, the ability of the antenna to receive signals might be hindered.Above2.42GHz, with17%utilization, the ISM band shows considerable utilization butit could also provide some opportunity for secondary usage. As the measurement wasdone indoors it was able to detect much of the signals due to the short nature of signalsin this band. Some activity on the satellite uplink and downlink bands were alsodetected at a frequency of2.305-2.32MHz and2.335-2.36MHz.The utilization pattern over24hours was also investigated. For this analysis, bandscontaining the three most utilized bands i.e. TV broadcasting,2G cellular band and the3G cellular bands were investigated. The TV broadcasting band was found to betransmitting during the day time with the occupancy being highly unpredictable,however during the night especially after00:00there seems to be less transmissionswith the occupancy being predictable. This observation is in tune with the normaloperating manner of such bands as most of the TV stations switch off their transmissionafter00:00. The2G and3G cellular bands indicate a random occupancy from beinghighly unpredictable during the day time to being moderately predictable at night. Thispattern is also understandable with people utilizing the bands during the day and lesspeople using the services at night.The conversion of the raw data which is a collection of power levels across700-2400MHz considered as a time series containing yields the channel occupancy. It indicates how long a channel is actually free at a given time before the primary userresumes transmission. Channel vacancy duration can be defined as the number ofconsecutive0’s in the channel occupancy series; this is the stage after thresholdingwhich is the conversion of power level to0’s and1’s. Channel vacancy durationanalysis was performed, the channel vacancy distribution was found to follow anexponential-like distribution even though the channels are not independently distributed.The coefficient of determination R2is well approximated by an exponential-likedistribution with values around0.93in all the locations considered. The significance ofthis analysis lies not in the empirical distribution of the channel vacancy distribution butin the distribution of vacancy timeslots amongst the bands analyzed. It was found outthat despite abundant opportunities the amount of timeslots/spectrum opportunitiesactually presented are much lower than initially expected. In essence, the spectrum isheavily fragmented and scattered across time thus greatly reducing the amount ofspectrum available for dynamic spectrum access.While channel vacancy duration provides long term information about the averageinformation for a particular channel, the Service Congestion Rate SCR metric providesthe short term or instantaneous picture of the spectrum at a particular instant. A situationmight arise whereby a secondary user needs to select between two services say3G1800Uplink and2G900Uplink, assuming their overall occupancy for the day is similar andalso their CVD is identical, a metric is needed to provide instantaneous information sothat the secondary user might be able to make an informed decision. For the purpose ofthis analysis, data from cellular band (both2G and3G) were used to determine the SCRvalues of these services both in the uplink and downlink paths. Five channels wereselected randomly from each service for the purpose of this analysis.The need to examine the correlation among frequency channels is of paramountimportance as it will provide valuable information on the similarities/differences inusage/behavior amongst different services analyzed. The correlation should be over alonger period of time due to the unpredictable nature of usage as this will provide acalculated insight into the actual correlation of the channels in question. For spectralcorrelation analysis, an hour long data from five channels were randomly selected andprocessed. From the results obtained from the spectral correlation analysis, it can beseen that there is average correlation coefficients of around approximately0.5in allcases considered. This implies a degree of independence in the spectrum utilizationpatterns across the services. It should be noted that some of channels used especially in the cellular bands are idle during the whole time because they serve as band gaps whichare used to avoid interference amongst services adjacent to each other. However, forthe temporal correlation analysis spectrum occupancy results from Katsina town andAbuja city are compared and correlation analysis performed amongst the servicescommon to both cities. The2G uplink,2G downlink and the broadcasting bands wereconsidered, results indicates that there is little or no correlation amongst the services.The correlation results were found to be0.1023,0.4999, and0.2303for2G uplink,2Gdownlink, and broadcasting band respectively. These findings are in tune with thegenerally known fact that spectrum utilization is dependent on location.The main attraction of using neural network based spectrum prediction lies in thefact that cognitive radios can learn and train itself from historical information obtainedby the cognitive radio without redesigning the whole system completely as is the casewith other methods. Unlike other models such as the Markov chain approach tospectrum prediction, the neural network model need only to be updated with the mostrecent data as its input. This approach saves power, sensing time and manpower.Machine learning has already been proposed as an integral part of future cognitiveradios, the high prediction accuracy realized in this model will greatly reduce the timerequired to sense whole bands. In addition, the processing power required at the basestation will be also reduced. Two factors generally influence modeling a network duringthe learning and training session. One is the initial interconnecting weights of thenetwork, and another is their modified quantities. Generally the initial interconnectingweights of BP ANN are often stochastically and blindly produced, this might cause thenetwork to run into partial optimization and therefore decrease the probability to obtainthe optimal solutions. Moreover, because the Delta rule is always adopted to modify theinterconnecting weights of BP ANN, the convergence velocity is always slow, orsometimes the network does not even converge. These shortages of BP ANN are quitenecessary to be optimized and improved. The problem of partial minimum of a BPANN can be solved by adjusting the initial interconnecting weights of the network. Thiscan be achieved through the application of Genetic algorithm because the problem is anon linear problem. GA is a nonlinear optimization method that has very strong abilityof global searching. To the best of our knowledge this problem has not been addressedin neural network based spectrum prediction. For the purpose of this work, five popularservices were considered. The GSM900uplink channels licensed to Etisalat had a meanprediction error of about0.035over five channels that were selected randomly. We observed a mean prediction error of about0.005in the GSM900downlink band,0.0179prediction error was recorded in the3G downlink band. Errors of0.0004and0.0007were observed in the broadcasting band and3G downlink band (licensed to StarcommsNigeria) respectively. The low prediction errors are the lowest recorded in the works wehave considered so far. Further analysis on energy utilization reveals bands with higherutilization level record higher sensing energy reduction as compared to bands withlower utilization that record lower energy reduction.Cooperative spectrum sensing has already been shown to improve the reliability ofspectrum sensing by exploiting the spatial dimension through cooperation. This hasbeen shown to increase the detection probability because the probability that all userswill experience deep fading has been reduced. Successful deployment of cognitive radiowill depend on the development of highly reliable spectrum sensing techniques ofwhich cooperative spectrum sensing fits the bill. The concept of cooperative spectrumsensing has already been extended to spectrum occupancy measurements. Mostspectrum occupancy measurements were done using a single device, the data captured isthus from a single device and can be viewed as unreliable especially in harsh channelconditions. All in all, the problem of weight selection was resolved using geneticalgorithm, thereby improving the overall accuracy, secondly, the data used wasobtained from a cooperative spectrum measurement which involved two devices unlikethe traditional spectrum occupancy measurement that involves a single device thisimproves the reliability of the model.