The main tool in our measurement setup is the hand-held FieldFox Spectrum Analyser (SA) characterized with a low average noise level around -150 dBm/Hz (when the internal pre- amplifier is active). The spectrum analyser is connected to a custom made discone antenna [171] mounted on top of an SUV car. The setup is interfaced via Ethernet port to a laptop equipment with MATLAB®. A capturing script configures the parameters of the SA, then periodically requests spectrum sweeps for the range 400 to 6000 MHz with 1.0 MHz reso- lution bandwidth. In order to capture the geographic effect on the measurements, we utilize a standard USB-GPS receiver, and stamp each sweep with the corresponding location and time. In Fig. 2.25, the diagram of the setup is illustrated.
We summarize the measurement parameters in Table2.13, noticing that the resolution bandwidth is selected as 1.0 MHz in order to speed up the sweep cycle, that is to allow more location-dependent analysis, and to capture any rapid changes in the environment [72]. This wide resolution bandwidth comes at the cost of higher noise power. The total travel route as obtained from the measurement log was 248 km, performed over 3.8 hours.
Table 2.13 Measurement Parameters
Parameter Value
Frequency Span 400 - 6000 MHz
Resolution Bandwidth 1.00 MHz
Sweep Time (Average) 3,950 ms
Noise floor (Average) -149.55 dBm/Hz Antenna System Gain (Average) -7.35 dBi
Antenna Polarization Vertical
Total Route Length 248 km
Total Measurement Duration 3.8 hours
Preamplifier Gain 20 dB
Measurement Regions
The main scope of our experiment is to measure the spectrum occupancy from the per- spective of a mobile vehicle experienced in different classes of urban environments. Ac- cordingly, we first accurately describe the sensed environment in terms of the surrounding population density, utilizing the Australian population geographic maps, published by the Australian Bureau of Statistics [2], then we segregate the raw measurement data into three different groups corresponding to high, medium and low population densities. The overlay of the travelled route is depicted in Fig. 2.26 on top of the population density geographic map of greater Melbourne area.
As it can be noticed from the figure, the travel route is divided into three distinct groups, according to the degree (the class) of urbanization:
• Class 1: High population density class (over than 8,000 persons/km2), corresponds to Melbourne Central Business District (CBD). We refer to this class as dense urban environment.
• Class 2: Medium population density class (from 500 to 5,000 persons/km2), corre- sponds to a typical suburban environment.
• Class 3: Low population density class (less than 500 persons/km2), corresponds to a typical rural environment.
Fig. 2.26 The travel route of the experiment in greater Melbourne is divided into 3 classes or regions. The left bar indicates the population density as per the Australian Bureau of Statistics. [2]
Antenna Calibration
In this experiment we have utilized a wideband discone antenna [171]. In order to guarantee a homogeneous detection of the sensed signal, we perform gain calibration inside an ane- choic chamber with the aid of a network analyser and a calibrated horn antenna acting as a reference gain antenna, where the gain of the discone antenna can be calculated using the following expression:
GDis.= S12Case1− S12Case2+ GRef., (2.27)
where S12Case1 and S12Case2 are the insertion loss measured by the network analyser cor-
responding to the following cases: (Case 1) one port is connected to the reference antenna, and the second port to the discone antenna, (Case 2) both ports are connected to identical reference horn antennas. The term GRef.represents the gain of the reference horn antenna as
per the factory calibration report. This calculation is performed of each measurement point corresponding to a certain frequency. The resulting gain-frequency plot of the antenna is depicted in Fig. 2.27, noticing that the measurement includes the effect of the cable loss.
Fig. 2.27 The discone antenna gain, with/without the effect of the attached cable loss.
Noise Analysis and Detection Threshold
From the perspective of cognitive radios, spectrum sensing is the key element in understand- ing the surrounding radio environment, where diverse sensing methods are proposed in the literature [172]. However, when utilizing a spectrum analyser, the only feasible choice is to use the energy detector method, which requires no prior knowledge about the signal being measured.
In energy detector method, the collected signal energy is compared to a predefined threshold, we call it θ , where the channel occupancy is simply identified when the sig- nal energy is above this threshold. In a cognitive radio, determining the optimum value of the detection threshold depends on both internal receiver’s noise and the monitored signal’s strength. However, in our particular case we can only obtain knowledge about the internal noise of the spectrum analyser.
In order to understand the impairing noise statistics, we record 1,000 samples for each frequency point (400 to 6000 MHz) while connecting a calibrated matched load (50Ω). Then we obtain the histogram distribution for these samples, as depicted in the Fig. 2.28, where the colour indicates the normalized intensity of the histogram, equivalent to the Probability Density Function PDF of the noise power.
greater than a threshold θ ( f ) is equivalent to the Complimentary Cumulative Distribution Function (CCDF) obtained at θ ( f ), expressed as:
CCDFN( f )(θ ( f )) = P[N( f ) > θ ( f )]. (2.28) Accordingly we can obtain the threshold that corresponds to a certain level of a false alarm value. The latter is defined as the probability of judging a channel as occupied, while in fact it is not. That is due to the contribution of the noise power in raising the sensed signal level. If x is the actual received signal and n is the noise signal, we can write the total apparent sensed signal energy Esseen by the spectrum analyser during the integration interval ∆Tsas
the following:
Es=
∑
i∈∆Ts
[x(i) + n(i)]2, (2.29)
where i is merely the collected samples index within the integration interval ∆Ts (sensing
interval). Then the false alarm probability can be expressed as:
pf = P[Es> θ × ∆Ts|Ho], (2.30)
where Ho is the hypothesis of having an unoccupied channel. We can notice that in the
case of Ho, the false alarm probability is simply equal to the noise CCDF, because the total
apparent sensed signal energy will be Es= ∑i∈∆Ts[0 + n(i)]
2
= N∆Ts.
In our measurement we select a false alarm value at pf = 5%, to minimize the errors
in the occupancy evaluation, where a moving average is employed to smooth the inherited fluctuations in the threshold level. The window of the moving average is selected to cover 20 frequency samples, i.e. 20 MHz. The resulting detection threshold (the fifth-percentile) is depicted in Fig.2.28.