Performance Analysis of Three Spectrum Sensing Detection Techniques with Ambient Backscatter Communication in Cognitive Radio Networks

1  Introduction

In wireless CR networks, AmBC supports the detection of the Primary User (PU)/Primary Transmitter (PT) and Secondary User (SU)/Secondary Transmitter (ST) by estimating the spectrum availability for SU data transmission. The throughput performance of the SU is greatly improved compared to that of the PU [1–3]. The promising technology of AmBC can facilitate sustainable communication systems that do not depend on a dedicated carrier emitter. In the present paper, a novel framework is proposed for CR networks, with the support of AmBC. For Harvest-Then-Transmit (HTT) operations, an opportunistic SS framework is migrated with AmBC. The parameters of throughput, energy efficiency and energy consumption are measured to estimate framework performance [2–4]. A hybrid transmitter is developed by combining AmBC with the Internet of Things (IoT) [5–7]. A new spectrum sharing method is proposed by developing AmBC to enable it to estimate the entire RF spectrum. The Direct Link Interface (DLI) means that there are significant errors in the legacy spectrum. To minimize these errors, Conventional Energy Detection (CED) is proposed as a novel, statistical, clustering framework [8–11]. AmBC technology is used to estimate the spectrum slot by detecting signals across the CR spectrum. Two closed form detection thresholds are estimated using detection algorithms. The bit error rate is estimated using the probabilities of 0 or 1 at low and high SNR levels [12–15]. Integrating AmBC into RF-powered CR networks has been shown to be a promising method for achieving energy and spectrum efficient communications, which is very attractive for low-power or no-power communications. In such scenarios, a PU/SU can operate in either transmission mode or backscatter mode. Specifically, an SU can directly transmit data if sufficient energy has been harvested (i.e., transmission mode) or an SU can backscatter ambient signals to transmit data (i.e., backscatter mode). It has been demonstrated that integrating AmBC into RF-powered CRN is a promising strategy for achieving energy and spectrum efficiency in communications, which is very attractive for low-power or no-power communications. An SU can operate in either backscatter mode or transmission mode in these situations. Specifically, if sufficient energy has been harvested, an SU can transmit data directly (i.e., transmission mode) or an SU can backscatter ambient signals to transmit data (i.e., backscatter mode) [16–18] in order to provide security to quasi-passive CR communications in the electromagnetic spectrum and to activate incoming signals using real-time systems. This can be achieved by SS techniques with AmBC low power and efficient spectrum utilization, with the support of the semi-passive adaptive modulation technique. No additional spectral energy is added for quasi-passive opportunistic communication [19–21]. AmBC is a promising technology to support IoT by enabling low power operations like Wi-Fi and TV signals. However, it is more difficult to decode the backscatter signals with AmBC enabled networks and AmBC devices, rather than legacy receivers. Some Machine Learning (ML) based techniques are being developed to assist the AmBC signal detection [22–24].

The present paper focuses on how to estimate user presence more quickly and accurately with high throughput. To estimate channel availability in the CR spectrum, AmBC is used with MFDI, CFDI and HFDI computational modelling techniques. As the channel is not uniform, the radiated power is not uniform across the entire spectrum. We estimate a dynamic threshold for each channel, based on AmBC at a low power level. An analysis of three proposed SS detection techniques involves the Generalized Likelihood Ratio Test (GLRT). The key parameters for sensing the channel and measuring the accuracy of data received by end users are sensing time and throughput. Additive White Gaussian Noise (AWGN) is considered with non-uniform noise in the AmBC environment, to sense the channel and estimate the dynamic threshold. The performance of the three techniques is analysed for the parameters of PD, Pfa, Pmd, sensing time and throughput, and the best detection technique is identified. In wireless communication, environmental noise is not uniform and should vary depending on the parameters of temperature, humidity, pressure, etc. These variations are different at different frequency levels and the propagation loss/attenuation varies considerably in terms of different frequency levels. Therefore, in this paper, the authors considered a non-uniform AWGN to estimate the threshold for each channel. If the threshold value is not estimated properly, errors in relation to Pfa and Pmd arise.

2  Proposed Methods

To estimate whether a channel in the spectrum is free or busy, two hypotheses can be written to assume that only noise is present (1) or that both user and noise are present (2) [7]. These hypotheses are represented as:

H0=y(n)=w(n)(1)

H1=y(n)=s(n)+w(n)(2)

where H0 denotes the condition in which the spectrum slot is vacant and only AWGN (represented by w(n)) is exhibited. User presence in the spectrum slot is shown in the scenario, H1, where the original signal is represented by s(n) and the noise signal by w(n).

In wireless communications, various noise levels are considered to exist in each channel. Therefore, the threshold for each power/SNR level must be estimated and the threshold changes dynamically, depending on the received SNR. To estimate the threshold for each channel, ‘N’, a jointly Gaussian random variable [7] is given as:

P(y;Hi)=1(2πσ2)N/2det(C)1/2exp⁡[−12(y−Hi)TC−1(y−Hi)](3)

where C−1 is the inverse covariance, T is the transpose of the signal and Hi is H0 or H1. The 5G spectrum bands are the closest to the satellite frequency bands, i.e., they have frequency values in the tens of GHz. To estimate the threshold value in the 5G spectrum for user presence and absence, GLRT detection criteria are used. The GLRT threshold estimation condition [8] is:

L(y)=P(y;H1)P(y;H0)=γ(4)

2.1 MFDI

In the MFDI method, the threshold value in the decision-making device determines how the Matched Filter (MF) will be used to make decisions. The cellular radio ecosystem contains near-end and far-end consumers. High SNR signals are received by near-end users, whereas low SNR signals are received by far-end users. It is vital to estimate user presence below the cellular spectrum threshold. If the threshold value is static and high, at a low SNR, no user can be detected, i.e., Pfa. Alternatively, if the threshold value is low, then powerful noise signals cross the threshold and indicate a user’s presence, i.e., Pmd. Therefore, threshold estimate is the crucial factor in effectively identifying user presence. We developed a dynamic threshold estimation equation and the best threshold value is determined using the GLRT detection criteria. Different users have different fading conditions and propagation path losses. To estimate the dynamic threshold, a non-uniform AWGN is considered. The received signal is fed through an analogue-to-digital converter and the digital signal, x(n), is used as the input to the MF (Fig. 1). The MF maximizes the signal and suppresses the noise. The output of the MF, y(n), is analysed via GLRT, using the Least Square (LS) method to estimate the threshold. The threshold condition for MFDI with GLRT (MFDIG) is [9]:

L(Y)=(H1−H0)TC−1y≤>[ln⁡(γ)+12((H1−H0)TC−1(H0+H1))]∗[hST+W](5)

where W is the M × N noise matrix, h is an LS estimator for the desired channel and S is the spectrum slots/channels [s1, s2,…, sN ]T.

Figure 1: MFDI spectrum sensing technique

2.2 CFDI

In the CFDI method, the received signal, y(n), serves as the input to the N-point Fast Fourier Transform (FFT) (Fig. 2). The FFT is used for discrete spectral analysis to determine the frequency content of the signal. FFT has been used to achieve computational efficiency by adopting a ‘divide-and-conquer’ approach. A correlator is used to compare the threshold value of the current sample with that of the preceding sample, using the FFT signal output. If the values match, the current sample is used. If they do not match, the current sample is sent for further analysis. The threshold condition for CFDI with GLRT (CFDIG), using the LS method, is [10]:

Ryy∗∝=[ln⁡(γ)+12((H1−H0)TC−1(H0+H1))exp(−j2παnfs)]1/2∗[hST+W](6)

Figure 2: CFDI spectrum sensing technique

2.3 HFDI

In CFDI, the unfiltered input sample of the received signal is applied directly to the N-point FFT [11]. Due to the unfiltered input signal, PD might be low at the detection level. The MF simultaneously increases the signal component and suppresses the noise amplitude. We attempted to combine the two detection algorithms of MFDI and CFDI into HFDI. The proposed HFDI block diagram is shown in Fig. 3.

Figure 3: HFDI spectrum sensing technique

The samples received by the HFDI receiver are applied to MF, as y(n). At some point in time, the MF enhances the signal component and decreases the noise amplitude. The N-point FFT input is combined with the MF output. The frequency domain signal from the FFT is applied to a correlator, which compares previous samples to the results of the N-point FFT. If the values match, the current sample is used. If the new samples do not correlate with the previous samples, they are further analysed via a threshold detector. The dynamic threshold is computed for the proposed HFDI using GLRT detection criteria (HFDIG) with the LS method [12]:

L(Y)=(H1−H0)σ2∗[ln⁡(γ)+12((H1−H0)TC−1(H0+H1))exp(−j2παnfs)]∗[hST+W](7)

3  Performance Evaluation Metrics

To analyse the performance of the detection system, the following parameters are considered [11]: During SS, the channel/sample energy was detected and given to the comparator to check the threshold level. If the detected sample energy was above the threshold level, the user was presented in the spectrum slot. If the sample energy was less than the threshold level, the spectrum was identified as vacant. Threshold level played a vital role in identifying user presence in the spectrum. Accurate information was estimated for the spectrum by the parameter PD. Hence, PD was estimated using Eq. (8).

Pfa=Q(γ−(H1−H0)TC−1H0(H1−H0)TC−1(H1−H0))(8)

If the user was located far away from the cell tower or in the overlay region, a weak signal was received and the SNR value recorded almost negative decibels. At a low SNR, noise domination was increased, or the noise signal energy was higher than the original signal power. In this case, the threshold value was low, the strongly noise dominated signal crossed the threshold and the decision was made that the user was present, however, there was no user. This type of wrong detection is referred to as Pfa. These false alarms are estimated using Eq. (9):

PD=Q[Q−1(Pfa)−(H1−H0)TC−1(H1−H0)(9)

If the user was located far away from the cell tower or in the overlay region, a weak signal was received, and the SNR value recorded almost negative decibels. At a low SNR, the noise domination was increased, or the noise signal energy was higher than the original signal power. In this case, the threshold value was high, the signal did not cross the threshold and the decision was made that the user was absent, however, there was no user. These inaccurate detections are known as Pmd and are estimated using Eq. (10):

Pmd=Q(γ−(H1−H0)TC−1H1(H1−H0)TC−1(H1−H0))(10)

Throughput is an important parameter, the improvement of which entirely depends on quicker sensing of the spectrum to allow more time for data transmission. In this paper, we considered both the PU and the SU. In Fig. 4, TF is the frame duration, TS is the time to sense the channel, Tc is the time taken for the sensed information to be given to the fusion centre and T is the total sensing and decision time of the spectrum. The data transmission duration is, therefore, TF − (Ts + Tc). The throughput estimation with respect to a false alarm is:

Thf=TF−TT(1−Pfa)(11)

Thd=TF−TT(1−PD)(12)

Figure 4: Spectrum sensing time representation

The data are not continuously transmitted in the spectrum; when the spectrum is scanning, data transmission should be in the ideal mode. Therefore, the time interval in data transmission is represented as β. The length of the frame transmitted in the spectrum is represented as TRF:

Thf1=TF−TT(1−Pfa)e(−TRFβ)(13)

Thd1=TF−TT(1−PD)e(−TRFβ)(14)

Sensing time is measured for the CR spectrum is:

T_s=tN(15)

where Ts is the sensing time and N is the number of samples required for the CR system to achieve the target PD of 100%.

4  Results

To test and validate the results, the authors have considered the assumptions with respect to the MATLab Simulator, which is shown in Table 1.

4.1 PD

A comparison of the three proposed methods for parameter PD is shown in Fig. 5. When the power level increases from −10 to 0 dB, the detection probability also increases. The best method shows a high PD at low SNR values. At −10 dB, the detection probabilities are 0.45, 0.66 and 0.69 for CFDIG, MDIG and HFDIG, respectively, showing that the proposed HFDIG is superior. HFDIG and MDIG provide a high detection probability of 1 from −4 dB onwards, but the CFDIG only shows this performance at −1 dB.

Figure 5: Comparison of PD vs. input SNR

HFDIG and MDIG perform better than CFDIG at low SNR values. From −10 to −4 dB, HFDIG provides a higher detection probability than MDIG at all SNR values. The proposed HFDIG is, therefore, superior. The rate of increase of detection probability between the proposed three methods is shown in Fig. 6.

Figure 6: Comparison of performance on PD using dynamic thresholds

If the threshold value is identified accurately, then there is a low detection probability. If the threshold level is changed according to the input SNR, there is a better chance of improving the detection probability than using the fixed threshold approach. The dynamic threshold approach provides more accurate detection than the fixed threshold approach for the parameter PD. MDIG, CFDIG and HFDIG perform better than the existing methods. Fig. 6 shows the rate of increase of detection probability from −10 to −5 dB. At −10 dB, the detection probabilities of HFDIG, MDIG and CFDIG are 0.69, 0.66 and 0.45, respectively. At −5 dB, the detection probabilities of HFDIG, MDIG and CFDIG are 0.98, 0.94 and 0.83, respectively. Therefore, when the power level increases from −10 to −5 dB, the rate of increase of the detection probabilities for HFDIG, MDIG and CFDIG are 0.29, 0.28 and 0.38, respectively. CFDIG has the highest rate of increase in detection probability but performs less well at low SNR levels. HFDIG has higher detection probabilities overall, therefore, at a low SNR level, this is the best method when evaluated against the parameter PD.

4.2 Pfa

A comparison of the three proposed methods for the parameter Pfa is shown in Fig. 7. When the power level increases from −10 to 0 dB, Pfa decreases. The best method gives a low Pfa at low SNR values. At −10 dB, the Pfa is 0.48, 0.33 and 0.30 for CFDIG, MDIG and HFDIG, respectively, meaning that HFDIG offers the lowest probability of showing a false alarm. HFDIG and MDIG show almost zero false alarms from −4 dB onwards, but CFDIG only reaches its lowest probability (of 0.02) at 0 dB. HFDIG and MDIG show fewer false alarms than CFDIG at low SNR values and HFDIG indicates fewer false alarms than MDIG at all SNR values.

Figure 7: Comparison of Pfa vs. input SNR

The rate of decrease in false alarm probability for the three methods is shown in Fig. 8. If the fixed threshold value is low, then there is a high probability of generating false alarms. If the threshold level is changed according to the input SNR, then the false alarm performance will also change. In terms of the dynamic threshold, MDIG, CFDIG and HFDIG provide better detection than existing methods. The performance analysis for Pfa is shown in Fig. 8. At −10 dB, the false alarm probabilities of HFDIG, MDIG and CFDIG are 0.30, 0.33 and 0.48, respectively; at −5 dB, the probabilities are 0.01, 0.05 and 0.22, respectively. Therefore, when the power level is increased from −10 to −5 dB, the decreases in false alarm probability for HFDIG, MDIG and CFDIG are 0.29, 0.28 and 0.26, respectively. HFDIG has the highest rate of decrease of Pfa and shows fewer false alarm probabilities at low SNR levels. It is, therefore, the superior detection method at lower SNR values, when evaluated against the Pfa parameter.

Figure 8: Comparison of performance on Pfa using dynamic thresholds

4.3 Pmd

A comparison of the three proposed methods for the parameter, Pmd, is shown in Fig. 9. As the power level increases from −10 to 0 dB, the Pmd decreases. The best detection method is the one that offers the lowest Pmd at low SNR values. At −10 dB, the Pmd is 0.54, 0.40 and 0.37 for CFDIG, MDIG and HFDIG, respectively. The proposed HFDIG shows superior performance. HFDIG and MDIG show almost zero missed detections from −4 dB onwards, but CFDIG only shows zero from −1 dB onwards. HFDIG and MDIG, therefore, they provide superior missed detection over CFDIG at low SNR values. Between −10 and −4 dB, HFDIG performs better than MDIG across all SNR values.

Figure 9: Comparison of Pmd vs. input SNR

The rate of decrease of Pmd in all three methods is shown in Fig. 10. If the fixed threshold value is high, then there is a high Pmd. If the threshold level changes according to the input SNR, missed detection performance will also change. The dynamic threshold leads to more accurate detection, as MDIG, CFDIG and HFDIG offer less missed detection than the existing method (Fig. 11).

Figure 10: Comparison of performance on Pmd using dynamic thresholds

Figure 11: Local sensing time vs. probability of false alarm at SNR = −10 dB

At −10 dB, the Pmd of HFDIG, MDIG and CFDIG is 0.37, 0.40 and 0.54, respectively; at −5 dB, the probabilities are 0.01, 0.02 and 0.16, respectively. Therefore, when the power level increases from −10 to −5 dB, the rates of decrease of Pmd for HFDIG, MDIG and CFDIG are 0.36, 0.38 and 0.38, respectively. MDIG and CFDIG show equal rates of decrease in missed detection in relation to HFDIG. At a low SNR level, HFDIG has fewer missed detection probability levels and shows greater improvement. Hence, HFDIG is the superior method when evaluated against the missed detection parameter.

Fig. 11 represents the sensing time in milliseconds (ms) vs. the probability of a false alarm at an SNR of −10 dB. HFDIG is quicker than the other two techniques, saving 0.01 and 0.02 ms by comparison with MFDIG and CFDIG, respectively, when Pfa = 0.1.

Fig. 12 represents the sensing time (ms) vs. the PD at an SNR of −10 dB. HFDIG is quicker than the other two techniques, saving 0.1 and 0.14 ms in relation to MFDIG and CFDIG, respectively, when PD = 1.

Figure 12: Local sensing time vs. probability of detection at SNR = −10 dB

HFDIG shows superior performance according to PD, Pfa, Pmd and sensing time. Therefore, throughput is estimated for HFDIG at various lambda values.

Fig. 13 shows the normalized throughput vs. the Pfa at various data transmission times, shown as beta (β) values. The Pfa increases the β value exponentially, as per Eq. (13). Fig. 13 shows that β = 500 ms gives better throughput than β = 400 ms and β = 450 ms. Therefore, when β is increased, throughput also increases.

Figure 13: Normalized throughput vs. probability of false alarm

5  Conclusion

The present research has compared the performance of AmBC technology with HFDIG, MFDIG and CFDIG computing modelling techniques by setting against the parameters of PD, Pfa, Pmd, local sensing time and throughput. HFDIG exhibits superior performance across all parameters and there is an improvement in detection, false alarms and missed detection of 3%, 3% and 4%, respectively. Compared with HFDIG, the local sensing time of CFDIG and MFDIG is improved by 3% and 2% at Pfa = 0.5, and the local sensing time of CFDIG and MFDIG is improved by 2% and 4 % at PD = 0.5. Hence, empty slots in the RF spectrum are sensed and identified more quickly and accurately with the proposed HFDIG.

Funding Statement: The authors would like to express their gratitude to the Ministry of Higher Education Malaysia for funding this research project through Fundamental Research Grant Scheme (FRGS) with Project Code: FRGS/1/2022/TK02/UCSI/02/1 and also to UCSI University, Malaysia and to the Vice Presidency for Graduate Studies, Business and Scientific Research (GBR) at Dar AI Hekma University, Jeddah, Saudi Arabia.

Conflicts of Interest: The authors declare that they have no conflicts of interest to report regarding the present study.