An Improved RF Detection Algorithm Using EMD-based WT

Abstract

More and more problems for public security have occurred due to the limited solutions for drone detection especially for micro-drone in long range conditions. This paper aims at dealing with drones detection using a radar system. The radio frequency (RF) signals emitted by a controller can be acquired using the radar, which are usually too weak to extract. To detect the drone successfully, the static clutters and linear trend terms are suppressed based on the background estimation algorithm and linear trend suppression. The principal component analysis technique is used to classify the noises and effective RF signals. The automatic gain control technique is used to enhance the signal to noise ratios (SNR) of RF signals. Meanwhile, the empirical mode decomposition (EMD) based wavelet transform (WT) is developed to decrease the influences of the Gaussian white noises. Then, both the azimuth information between the drone and radar and the bandwidth of the RF signals are acquired based on the statistical analysis algorithm developed in this paper. Meanwhile, the proposed accumulation algorithm can also provide the bandwidth estimation, which can be used to make a decision accurately whether there are drones or not in the detection environments based on the probability theory. The detection performance is validated with several experiments conducted outdoors with strong interferences.

1. Introduction

Based on the lightweight aircraft, the drones have become one of the widely accepted terms integrated with different devices, which are operated without a human pilot onboard inconjunction with ground control systems [1]. As a useful tool, the drones have are used in military applications initially including search/destroy missions [2], surveillance, aerial photography, and reconnaissance [3], which can provide higher spatial resolution such as up to the sub-centimeter level in small areas due to their low operational altitude [4]. The relatively low cost of different drone platforms is the most important factor, which leads to their widepopularity and recent success. Over the previous decade, drones have been increasing, whichare used for data acquisition by more and more actors and stakeholders such as scientificinstitutions, governmental authorities, individuals and commercial operators [5]. Especially incivil applications, drones are widely applied in land administration [6], documentation of cultural heritage and archaeological sites [7], support for disaster management [8], high-resolution surface reconstruction in the geosciences [9], forest and agriculture change detection [10], surveying and mapping [11], and wildlife observation [12].

Drones can provide more and more convenient services, but numerous problems for the public security come correspondingly. As known, one drone was used to interrupt the US Open tennis match and another crashed at the White House. The drones in illegal flight disturbthe air flight conditions and pose a great threat to the flight safety of passenger aircraft. Somesmall drones are usually used for invading human privacies. More and more accidents have been trigged due to the popularity of drones. Until now, shooting the drones failing to comply with regulatory requirements is considered as the most effective solution. The authors in [13-14] employ the genetic algorithm (GA) to jam the drones. In [15], the GPS signals arejamming, which can make the drone disable. However, all these methods can only work betteronly when these drones are detected, which are impossible to achieve without effective detection solution. Thus, more works are required to deal with the public security. Many detection solutions have been developed in past few years including the acoustic recognition, active radar, visual recognition, infrared spectrum analysis, and RF detection [16-20]. By modulating the RF signals, the active radar was used for drone detection at short ranges [16]. Without considering the flight pattern, the drones are recognized using a calibrated radar crosssection technique in [17] at 2 000 m. The authors employed the complex-log Fourier transformmethod for drone detection, which can acquire the spectrogram features [18]. Based on the log-polar transformations and space-variant resolution, the drones are identified [19], which can only work at short range. The drones are located using the acoustic array in [20] at aneffective range of 300 m. Meanwhile, the detection algorithms based on neural networks (NN) [21], fuzzy systems [22], least squares (LS) [23], frequency based approaches [24], Kalmanfilters [25], and Gauss-Newton [26] etc. have also been developed. However, limited workscan be acquired for drone detection. Thus, some alternative methods have to be developed todeal with the problems of drones in illegal flight. This paper mainly deals with the weak signalextraction solution for the drone detection. The major factors including the static clutter and linear trend term are removed. The principal component analysis is used to acquire the effective radio frequency (RF) signals. And the signal to noise ratios (SNR) of RF signals isimproved based on the automatic gain control algorithm. The empirical mode decomposition (EMD) based wavelet transform (WT) is employed to reconstruct the effective RF signals. Based on the statistic theory, the azimuth information between the radar and drone is acquired. The bandwidth of RF signals is obtained using the developed accumulation algorithm. Using the probability theory, accurate judgements can be made whether there are drones or not in the detection environments.

The following sections are introduced as: the system model is introduced in Sec. 2. Sec. 3 discusses the proposed detection algorithm using passive radar. The detection results arediscussed in Sec. 4. Sec. 5 concludes the whole paper.

2. System Model

As shown in Fig. 1, a drone, a receiver, and a controller are considered in the developed passive radar system. The controller can emit RF signals, which can meet the requirements of communicating with the drone. The receiver can acquire RF signals emitted by the controller. As known, RF signals are usually within 2.4-2.4835 GHz, which can control the most drones for flight and image transmission based on the frequency hop spread spectrum technique [27]. Meanwhile, the RF signals have to meet the IEEE 802.11 standard [28].

Fig. 1. The system model.

Fig. 2. The spectrums of RF signals.

The used drone and the corresponding controller are developed by the DJ-Innovations. Using the receiver, the data can be acquired for each given distance such as 2500 m. Assuming that the power of the received RF signals is given by y(f ) in dBm, where f denotes the frequency samples within 2.4-2.5 GHz. All 5120 samples can be acquired with the interval of fw = 19.53 kHz as shown in Fig. 2. The results indicate that the RF signal emitted by the controller has a larger bandwidth, which is approximately 9.8 MHz. In real detectionenvironment, due to the larger range between the receiver and the drone, the RF signals are usually too weak to find especially in complex multipath environments. This paper aims at the drone detection outdoors.

3. Proposed Algorithm

Fig. 3 shows the proposed algorithm based on the acquired RF signals AMN at 500 m using theradar, which is an azimuth-frequency matrix as show in Fig. 4.

Fig. 3. The proposed detection algorithm.

Fig. 4. The acquired RF signals at 500 m.

Fig. 5. The results removing static clutter.

3.1. Static Clutter Removal

The static clutter in AMN [29] can be estimated as

(1)

Subtracting (1) for each value in A, the results C are acquired as shown in Fig. 5.

3.2. Linear Trend Suppression

This subsection employs linear trend suppression to eliminate the potential terms with lineartrend. As shown in Fig. 6, using the linear trend suppression, the results are [30]

(2)

where

(3)

Fig. 6. The results using linear trend suppression.

Fig. 7. The results using principal component analysis.

3.3. Principal Component Analysis

The principal component analysis is used to acquire RF signals. Usually, (2) can be divided into several complementary subspaces as [31]

D=USVT (4)

where S denotes a diagonal matrix, UMM denotes an unitary matrix, and VNN is an unitary matrix. The singular values i in S meet 1 ≥ 2 ≥3 ≥ ... ≥r ≥ 0. The columns in U, i.e. the eigenvectors of WWT, are composed of the left singular vectors. The columns in V, i.e. the eigenvectors of WTW, are composed of the right singular vectors. Equation (4) can beredefined as

(5)

where Mk is the kth intrinsic image with the same dimensions as E. As usual, (5) can be divided into two parts as

D=Mdrone + MNoise (6)

where Mdrone is the RF signals, MNoise is noises, which can be removed. Thus, the effective RF signals can be acquired as & nbsp;

(7)

Fig. 7 shows the results based on the principal component analysis algorithm.

3.4. Automatic Gain Control

As an adaptive algorithm, the automatic gain control algorithm is extensively used in many applications [32]. The gain values of the inputs in a given range usually can be adjusted to appropriate levels using the fed levels of the averaged outputs.

Fig. 8. The results using automatic gain control.

An effective application of automatic gain control method is to enhance the amplitudes of the weak signals in radar systems [33]. Via adjusting the gain values based on the calculated power of the signals in a window w, the key is to compare the gain values and gmax as

(8)

where 0,2, , 1 n N = −  , 0,1, , i M w = −  . The normalized gain values gnorm[i, n] are

(9)

where gmin[i, n] is the acquired minimal gain, which can be expressed as

(10)

where e[i, n] is the calculated power of the signals in a window w, which can be expressed as

(11)

Using the automatic gain control method, gmax can be determined based on the calculated gain values. In this paper, gnorm[i, n] is set to be 0.6 based on the empirical evidences [34].

We can acquire the following results given by

(12)

The results using the automatic gain control algorithm are shown in Fig. 8. We can see RF signals are improved significantly, which can make it easy to detect.

3.5. Noise Suppression

As an adaptive algorithm, EMD is widely used to deal with the non-static signals [35], which can make it divided into various intrinsic mode functions (IMFs) and a residual trend.

Fig. 9. The acquired IMFs using EMD algorithm.

Fig. 9 shows the first 12 IMFs of the RF signals and a residual trend. The residual trend is considered as noises. For the IMFs, the boundary between RF signals and noises is & nbsp;

(13)

where

(14)

(15)

(16)

and Nv is the number of values in IMFv, Cv is the maximum value in IMFv, and Ov is the number of local extrema in IMFv. For AWGN, (14) is approximately a constant. As a result, IMFv can be regarded as noises when (13) < 1. It was determined by simulation that IMFs 1 to 6 can beconsidered as noises. Using EMD, the results are shown in Fig. 10, which indicate the RF signals are improved obviously.

Fig. 10. The results using EMD

Fig. 11. Eight levels wavelet packet decomposition tree.

3.6. Signal Reconstruction

This subsection employs WT to reduce influences of noises and reconstruct RF signals [36 -37]. In wavelet analysis, RF signals for each given azimuth Gm(n), are used to compare with the wavelet function. The acquired coefficients can be used for showing how RF signals and the wavelet functions match. Due to RF signals are stored in digital form, the coefficients obtained from the discrete WT are

(17)

where k denotes the samples of RF signals, * denotes the complex conjugate, a is the scaling parameters, b is the translation parameters of the mother wavelet Ψ(n), and ( ) ( ) k b a ψ − is theson wavelets.

Fig. 12. The decomposed signals using WT technique.

In this paper, the Haar wavelet is considered as mother wavelet due to its simplenumerical implementation and simplified computation. By employing a wavelet packet decomposition tree with eight levels as shown in Fig. 11, the decomposed RF signals areshown in Fig. 12. Based on the spectrums of Welch power of RF signals, they can bereconstructed from the decomposed signals d4 as shown in Fig. 12. Fig. 13 shows thereconstructed signals H.

Fig. 13. The results using WT.

Fig. 14. The results using averaging filter.

3.7. Averaging Filter

This subsection proposes an improved filter to improve SNR. The following results can be obtained when the filter is performed on H in azimuth direction as

(18)

where 1, , 12 k M =      . 12 M    denotes the maximum integer value less than M/12.

Fig. 14 shows the results based on the filter, which show RF signals are improvedeffectively and the better performance on noises suppression.

3.8. Frequency Estimation

This subsection proposes three methods for frequency estimations of RF signals.

Firstly, the Hilbert transform is used to improve SNR. For each azimuth signal Im(n), the Hilbert transform is applied [37], which is given by

(19)

where ω denotes the frequency samples.

To acquire the frequency estimations, the characteristics including the skewness [38], kurtosis, standard deviation, and maximum slope are analyzed. For each azimuthal signal Jm(n), the standard deviation is given by

(20)

where μm denotes the mean value of Jm(n). And the maximum slope is expressed as

(21) Using RF signals at 500 m, the analyzed results are shown in Fig. 15. Thus, the frequency value corresponding to the maximum (20) is considered as the estimation τ1, which is

(22)

where

(23)

And the frequency value corresponding to maximum (21) is considered as the estimation & tau; 2, which is

(24)

where

(25)

ω denotes the frequency samples.

Further, the accumulation method in azimuth direction is proposed to estimate the frequency, which is given by

(26)

Fig. 15. The analyzed results of the RF signals.

Fig. 16. The results using the accumulation method.

As shown in Fig. 16, the maximum (22) is referred as the frequency estimation τ3, whichis given by

(27)

where

(28)

3.9. Target Determination

To determine whether there are drones in the detection environment or not, the errorscalculated from the three frequency estimates are employed as

(29)

Due to the bandwidth of RF signals is about 10 MHz, the calculated errors can be used as a threshold to determine whether there are drones in the detection environments. Based on the three frequency estimates, three errors can be acquired. There are drones when λ ≥ 2, while there are not drones when λ < 2. λ denotes the number of the error which is larger than 10 MHz.

3.10. Azimuth Estimation

To acquire the azimuth estimation, the standard deviation of each frequency signal is analyzed. For each frequency signal Jm(n), the standard deviation is

(30)

where μn denotes the mean value of Jm(n).

Fig. 17. The analyzed results of RF signals.

Using RF signals obtained at 500 m, the analyzed results are shown in Fig. 17, whichindicate the area with the normalized values larger than 0.4 are considered as the azimuthrange.

4. Results and Discussion

Experimental results are compared with the minimum mean distance (MMD) classifier. Using the passive radar system, several experiments are conducted to validate the developed algorithm. (1) In the conducted first experiment, the used drone in the air is 100 m in height. The distance from the radar to drone is 500 m, 1500 m, 2400 m, 2500 m, and 2800 m athorizontal direction. The conducted experiments are used to validate the better detection capability in long ranges. (2) The second experiment is carried out with the drone at a distance of 2400 m away from the radar. There are several interference sources in the detectionenvironments. The conducted experiment is used for testing the performance of noisesuppression using the developed algorithm. (3) In the third experiment, two drones are used astargets at different distances. The capability of multiple drone detection is validated.

4.1. Detection Performance at Different Distances

This subsection validates the performances of the developed detection algorithm using RF signals obtained at different distances including 500 m, 1500 m, and 2400 m, which are shown in Figure 18. We can see RF signals can only be extracted obviously in short range. However, RF signals are too weak to identify in long range such as 2400 m especially with stronginterferences.

Fig. 18. The acquired RF signals at (a) 500 m; (b) 1500 m; and (c) 2400 m. 

Fig. 19. The results after removing noises at (a) 500 m; (b) 1500 m; and (c) 2400 m.

Based on the developed detection algorithm, the results after removing clutters are shown in Fig. 19, which indicate RF signals are improved significantly compared with the results in Fig. 18.

4.2. Detection Performance Under Strong Interference

The detection capability of the developed algorithm under strong interference environments is validated based on RF signals acquired at different distances including 2500 m and 2800 m. In the detection environment, there are several interference sources as shown in Fig. 20. Thereare several interference sources i.e. the signals in the elliptical region with larger amplitudesthan RF signals, which make drone more challenging to identify. The results after removingnoises based on the developed algorithm are shown in Fig. 21. Based on the acquired results, the major interferences are suppressed effectively and RF signals are improved.

Fig. 20. The received RF signals at (a) 2500 m; and (b) 2800 m.

Fig. 21. The results after removing noises at (a) 2500 m; and (b) 2800 m

4.3. Frequency Estimation

The frequency estimates of RF signals at different distances based on the characteristics analysis algorithm are shown in Fig. 22.

Fig. 22. Frequency estimates using the characteristics analysis algorithm at (a) 500 m; (b) 1500 m; (c) 2400 m; (d) 2500 m; and (e) 2800 m

Fig. 23. Frequency estimates using the accumulation method at (a) 500 m; (b) 1500 m; (c) 2400 m; (d) 2500 m; and (e) 2800 m

Using the results, the developed algorithm can provide accurate frequency estimatescompared with the analyzed skewness and kurtosis. The frequency estimates of RF signals are within 2.42-2.43 GHz when the distances are 500 m, 1500 m, 2500 m, and 2800 m. Further, when the distance is 2400 m, the frequency estimate is in the range of 2.46-2.47 GHz. All the estimated bandwidths of RF signals are about 10 MHz.

4.4. Azimuth Estimation

Fig. 24. Azimuth estimates using the developed method at (a) 500 m; (b) 1500 m; (c) 2400 m; (d) 2500 m; and (e) 2800 m

The azimuth estimates at different distances using the developed algorithm are shown in Fig. 24. From the results, the azimuth estimates are within 90°-120° when the distances are 500 m, 1500 m, 2500 m, and 2800 m. When the range is 2400 m, the azimuth estimate is within 50 & deg;-80& deg; . All results fit well with the acquired RF signals as shown in Figs. 18 and 20.

4.5. Multiple Targets Detection

Fig. 25. The results from two targets (a) received RF signals; (b) results removing noises; (c) frequency estimates using the characteristics analysis method; (d) frequency estimates using the accumulation method; and (e) azimuth estimate using the developed algorithm.

The detection capability of multiple targets is validated using the radar system. Two drones are used as targets in the detection environments one is at a distance of 2400 m, and another is at a distance of 2500 m. The received RF signals are shown in Fig. 25(a), Fig. 25(b) shows the results after removing noises. Based on the results, two target areas are obvious to acquire, which provide the basis for the drone detection. The frequency estimates of the two drones are shown in Fig. 25(c) using the developed characteristics analysis algorithm. Fig. 25(d) provides the two frequency estimates acquired from the proposed accumulation method. Finally, the two azimuth estimates at different distances based on the developed algorithm areshown in Fig. 25(e). Based on the results, the azimuth estimate of one drone is within 90°-120°, while another is within 50°-80°.

5. Conclusion

This paper aims at the detection of aerial targets based on the passive radar. Using the passive radar, an improved algorithm is developed to identify drones. The noises affecting drone detection are suppressed effectively, and RF signals are improved based on the principal component analysis and automatic gain control algorithms. The EMD-based WT is used toreconstruct RF signals. The drones can be detected precisely. The azimuth information can beacquired based on the statistical theory. The monitoring system is validated and tested with different experiments conducted in stronger interference condition.