Intelligent Passive Detection of Aerial Target in Space-Air-Ground Integrated Networks

Mingqian Liu,Chunheng Liu,*,Ming Li,Yunfei Chen,Shifei Zheng,Nan Zhao

1 State Key Laboratory of Integrated Service Networks,Xidian University,Shaanxi,Xi’an 710071,China

2 Guilin Changhai Development Co.,Ltd,Guilin 541001,China

3 School of Engineering,University of Warwick,Coventry CV4 7AL,U.K.

4 School of Information and Communication Engineering,Dalian University of Technology,Dalian 116024,China

Abstract: Passive detection of moving target is an important part of intelligent surveillance.Satellite has the potential to play a key role in many applications of space-air-ground integrated networks(SAGIN).In this paper,we propose a novel intelligent passive detection method for aerial target based on reservoir computing networks.Specifically,delayed feedback networks are utilized to refine the direct signals from the satellite in the reference channels.In addition,the satellite direct wave interference in the monitoring channels adopts adaptive interference suppression using the minimum mean square error filter.Furthermore,we employ decoupling echo state networks to predict the clutter interference in the monitoring channels and construct the detection statistics accordingly.Finally, a multilayer perceptron is adopted to detect the echo signal after interference suppression.Extensive simulations is conducted to evaluate the performance of our proposed method.Results show that the detection probability is almost 100% when the signal-to-interference ratio of echo signal is-36dB,which demonstrates that our proposed method achieves efficient passive detection for aerial targets in typical SAGIN scenarios.

Keywords: aerial target detection; decoupling echo state networks;delayed feedback networks;multilayer perceptron; satellite illuminator; space-air-ground integrated networks

I.INTRODUCTION

Integration of space networks is highly achieve ubiquitous coverage for mobile communication networks,which expects to become the one of the enabling technologies for 6G networks [1, 2].At the same time,the use of satellites is found to be of paramount importance in intelligent surveillance.Thus, satellite is viewed as a vital part of space-air-ground integrated networks (SAGIN) to provide broadband access and an illuminator of opportunity to detect moving target[3].In the target passive detection,target detection is achieved by illuminator of opportunity on the reflected echo signal of the target.Hence,the passive detection systems have many advantages including anti-stealth,anti-interference, and an unlikely target for detection by electronic support measures(ESM)systems[4,5].Recently, various radio signals have been applied in passive detection systems, such as FM broadcasting,satellites,digital TV broadcasting,etc.[6,7].Satellite illuminators of opportunity perform better than other illuminators of opportunity regarding global coverage,24-h operation,and easy-to-access signal sources.Therefore, research on target detection using satellite illuminator of opportunity is essential and practically appealing in both industrial and academia.

In satellite illuminator of opportunity enabled target detection, the target reflected signal power is weaker than the direct satellite signal at the receiver.This is because the signal is attenuated and reflected by the target during transmission.Especially when directpath interference(DPI)/multi-path interference(MPI)and other interference occur, which makes it difficult to detect the target.For the target detection of the echo of the global positioning system(GPS)satellite,[8]analyzed the performance of the dual-base passive radar signal with the satellite signal as the irradiation source and detected the air target for the geometry and configuration that may occur in the passive radar system.[9,10]analyzed the detection power of GPS as a radiation source from the perspective of signal energy and fuzzy function.They proposed a method of GPS target echo detection.However, there are no detailed explorations and analyses of interference suppression,detection methods,and detector settings for the direct signal received from the illuminator of opportunity in these studies.Therefore,purifying the GPS direct signal received in the reference channel is significant,especially in the clutter environment.In addition,effectively suppressing the direct wave and other interference in the monitoring channel to detect weak echo signals is important.

On the other hand, the difficulty of target detection with the extremely low echo power has been talked before.[11]first used a wide range of cancellation algorithms to eliminate direct waves and multi-path interference.Then,it employed the anti-interference characteristics of the cyclic mutual ambiguity function to construct detection statistics for target detection.For the weak satellite signals using multiple satellites, in[12], the detection quantity was obtained based on a mutual fuzzy function processing of the fourth-order cyclic cumulant.In[13], based on fractional Fourier,the authors proposed a multi-frame technique to detect moving space targets in passive radar systems.This technique improved the system detection capability by the mixed coherent/incoherent integration processing of the fractional Fourier domain using a signal.However, the clutter characteristics have a particular impact on the target detection performance, and the above method did not consider complicated clutter environments.In complicated clutter environments, the weak echo signal is quickly submerged in substantial clutter interference, making target detection difficult.Therefore,detecting an aerial target in the background of clutter is crucial.

In existing works, the target detection problem in the background of clutter was studied from various aspects.In [14], authors used the built-in clutter subspace to suppress the clutter by orthogonally projecting the signal’s unit under test to the orthogonal clutter subspace.Based on the unit average constant false alarm detector for detection, [15] modeled the complex working environment faced by external radiation source radar, investigated the characteristics of clutter,and provided a foundation for the design of clutter processing.Meanwhile,authors in[16,17]studied the detection of external radiation sources with clutter.In[18], authors derived the detection statistics based on the maximum eigenvalues for the detection of constant false alarms in the background of K-distribution clutter.[19]applied a radial basis neural network to predict and cancel the time series of clutter and performed Fourier transform on the canceled signal for target detection.However, this method did not consider the effect of different signal-to-noise ratios on detection performance.In[20],authors calculated the mean and variance of the fourth power of the backscatter prediction error based on linear prediction to detect the target in the clutter with some prior knowledge.However,the above methods have a low signal-to-noise ratio.As such,the detection performance is affected by the clutter, which could not be directly applied to the target detection using the satellite illuminator of opportunity.

In this paper,an intelligent passive detection method for aerial target detection is proposed, by using satellite illuminator of opportunity in clutter environment.The main contributions of the paper are summarized as follows:

• To suppress interference for the direct wave in the monitoring channel,we utilize the delay feedback networks (DFNs) to improve the satellite direct wave signal in the reference channel containing clutter.Moreover, the minimum mean square error (MMSE) filter is used to adaptively suppress the satellite direct wave interference in the monitoring channel.

• After the direct wave suppression, the signal of the monitoring channel is input into the decoupling echo state networks (DESNs).The difference between the predicted value and the DESNs’input value is extracted as the detection statistics.

• The multilayer perceptron(MLP)is used to detect the target based on the extracted detection statistics.

The remainder of this paper is organized as follows.In Section II, the system model is presented.The direct wave improvement in the reference channel and its suppression in the monitoring channel are illustrated in Section III.Detection statistics construction and the detector design based on DESNs are proposed in Section IV.Section V shows the numerical results to verify the detection performance.Finally, Section VI concludes the whole paper.

II.SYSTEM MODEL

An illustration of system architecture for passive aerial target detection with satellite illuminator of opportunity for SAGIN is shown in Figure 1.Figure 1 shows that the reference channel receives the direct wave signal irradiated by the satellite and the clutter interference signal.The monitoring channel receives the echo signal reflected by the target, the direct wave signal,and the clutter as interference.Because the power of the direct wave signal and clutter signal in the monitoring channel is often higher than that of the echo signal,the echo signal is submerged in the direct wave and clutter interference.As such,the signal processing module processes and detects the signals received by the two channels.The signal processing includes purifying the direct wave in the reference channel, suppressing the direct wave signal and detecting the echo signal in the monitoring channel.

Figure 1. Passive detection system model in space-airground integrated networks.

In Figure 1,the signalxr(t)in the reference channel can be expressed as[21]

whereηstands for the amplitude of the direct wave signal,s(t)represents the direct wave signal,c(t)denotes the clutter withKdistribution,n(t) is a statistically independent stationary Gaussian white noise with zero mean.The echo signalxs(t)in the monitoring channel is described as

whereη′is the amplitude of the echo signal,τis the time delay of the echo signal relative to the direct wave signal,fdrepresents the Doppler frequency shift of the echo signal relative to the direct wave signal, and Ω denotes the amplitude of the direct wave signal in the monitoring channel,n′(t)is a statistically independent stationary Gaussian white noise with zero mean.

In general,GPS satellite signals use C/A codeC(t)or P codeC(t) to perform spread spectrum processing on dataD(t).This paper considers GPS satellite signals with C/A code for spread spectrum processing.Therefore,the signal received by the reference channel can be rewritten as[22]:

wherePrepresents the signal power,C(t)is the C/A code of the satellite signal,D(t) is the satellite navigation data,fcdenotes the carrier frequency of the received signal.

TheKdistributed clutter signal contains two main undulating components, namely the speckle component and the modulation component.Among them,the speckle component has an exponential probability density with an average value ofh,which is expressed as[23,24]

and the modulation componenthis described by the gamma distribution as

where Γ(v)is gamma function.Thus, the probability density functionP(r) of theKdistribution clutter is given by

wherebrepresents a scale parameter,vdenotes a shape parameter,andKv(.)is a second-class modified Bessel function of orderv.

This paper designs a target detection method based on reservoir computing networks (RCNs) to detect aerial target with the clutter using satellite illuminator of opportunity.The framework of the proposed method is shown in Figure 2.The target detection framework mainly includes two sections, which are direct wave interference suppression and target detection.The direct wave interference suppression section includes direct wave signal purification and direct wave interference suppression, and the target detection includes detection statistics construction based on DESNs and MLP-based detection decision.

Figure 2. Structure of target detection framework based on reservoir computing networks.

III.DIRECT WAVE INTERFERENCE ADAPTIVE SUPPRESSION

Given that the monitoring channel has a substantial direct wave signal, the direct wave signal power would be tens of dB higher than the echo signal power[25].Therefore, to effectively the target, it is essential to suppress the direct wave interference.

3.1 Direct Wave Signal Purification Based on DFNs

To complete the direct wave signal purification from the clutter,we utilize the DFNs to effectively suppress the clutter in the reference channel.The signal after the clutter suppression is demodulated to obtain the correct navigation data using the GPS characteristics,and then the reconstruction of the direct wave signal is completed.The above steps can effectively achieve direct wave signal purification.The specific steps of the purification method are illustrated as follows:

Step 1: Use the GPS acquisition algorithm to obtain the spreading code informationC(t), phase offset informationτl,and frequency offset informationflcorresponding to the reference channel signal.

Step 2: Use the DFNs to suppress the clutter in the reference channel effectively.The network structure of the DFNs is shown in Figure 3, and the training algorithm is described as[26,27]:

Figure 3. Network structure of DFNs.

wherexr(k)denotes the input state vector andkis discrete time variable,(k) is the estimated output state vector,the weightωiis the state weight,Nrepresents the number of neurons in the DFNs,τ=Nθis the delay length,andθis the equal delay interval.

Step 3: Demodulate the GPS direct wave signal to obtain navigation data.Firstly normalize the signalx′(k),then use the frequency offset informationfl,the spreading code informationC(t)and the phase offset valueτlto recover the navigation data, then calculate the recovery data envelope,and use thesign(·)function to complete the navigation data decision.

Step 4: Obtain a pure direct wave signal, modulate the recovered navigation data onto a phasesynchronized local C/A code,and up-convert to obtain a reconstructed direct wave signal.

3.2 Direct Wave Interference Suppression Based on Minimum Mean Square Error

The adaptive filter is used to suppress the direct wave in the monitoring channel.The structure of the adaptive filter is shown in Figure 4.The principle of the adaptive filter is based on the MMSE algorithm [28],and the iteration formula can be expressed as

Figure 4. Structure of the adaptive filter.

and

wherexs(k)stands for the received signal of the monitoring channel,xr(k)is the signal received by the reference channel,w(k)denotes the weight vector of the filter,e(k)is the error signal,λis a small normal number,µrepresents the step size factor, and 0< µ <2.

The signal of the monitoring channel after the interference suppression of the satellite direct wave signal contains the echo signal, clutter and Gaussian white noise,which is expressed as

IV.AERIAL TARGET DETECTION BASED ON DESNS

In this section, the echo signal is detected based on the DESNs.This detection method extracts the echo signal’s characteristics through the DESNs and uses the MLP to complete the data decision.

4.1 Decoupling Echo State Networks

The echo state networks (ESNs) are recurrent neural network architecture, the core of which are the reservoir with tens or hundreds of neurons[29].These neurons are randomly and sparsely connected o each other and can be used to predict clutter signals.DESNs are a particular case of ESNs,and its reservoir consists of multiple reservoirs,and the number of neurons in each reservoir can be set to be equal or unequal [30].Set DESNs of four reservoirs with repeated weight matrices.Each reservoir has the same number of neurons.The structure of the four-reservoir DESNs is shown in Figure 5.

Figure 5. Basic structure of the four reservoir DESNs.

DESNs have a cyclic weight matrix in the form of the following block matrix,

and the reservoir state update is given by[31]

wheref(.)stands for theS-shaped activation function,u(k)represents the input state,xss(k)denotes the internal state of the reservoir,d(k -1) is the feedback signal,andγ(k)is the current artificial noise vector inserted into the state update equation.If the number of input layer nodes isKand the number of output layer nodes isL,the cyclic weight matrixWis a matrix ofN ×N, the input weight matrixWinis a matrix ofN ×K,and the feedback weight matrixWbis a matrix ofN ×L.The DESNs used in this paper has four reservoirs,and each reservoir has the same number of neurons,where the input weight matrixWin,the feedback weight matrixWb,and the reservoir state matrixx(k)are given by

and

and the output of DESNs is expressed as

whereWoutrepresents the output weight matrix, andfoutdenotes a linear function.

4.2 Detection Statistics Construction Based on DESNs

In this section, we input the monitoring channel signal after the direct wave suppression into the trained DESNs to obtain the detection statistics.The process is detailed as follows.

Step 1: Firstly,complete the training of the DESNs by(7).Secondly,add the clutter sequencec(k)to the input end of the DESNs.Thirdly, calculate the reservoir statexc(k).Through the obtained state,the input clutterc(k), and outputy(k) =c(k), the state matrixMand the output sequenceZare obtained,whereM=(uT,,yT),Z=yT.Finally, the output matrixWoutis obtained by calculating the linear regression,whereMWout=Z.

Step 2:After the suppression of satellite direct wave interference, we input the monitoring channel signalxss(k) into the DESNs to obtain a predicted clutter sequencexss(k).

Step 3: Calculate the cancellation of the input signalxss(k) and the output signal(k), i.e.,z′′(k) =xss(k)-(k).Usez′′(k)as the decision variable for echo signal detection[32,33].The absolute values of the cancellation signals are shown in Figure 6 when the target exists or does not exist,respectively.

Figure 6. The absolute value of detection statistics based on DESNs.

4.3 Detector Design

A MLP is used to detect the signals in the monitoring channel, where the binary hypothesis is described as:H1denotes the event that the monitoring channel contains the echo signalη′s(k-τ)e-j2πfdk.H0is the event that the monitoring channel does not contain the echo signalη′s(k-τ)e-j2πfdk, which is expressed as

The MLP is pre-trained by multilayer neurons,which uses the MMSE as the cost function and the backpropagation algorithm to fine-tune the network’s weights and biases.The hidden layer of MLP is described as[34]

whereσ(·) denotes the sigmoid function,xmrepresents the input signal,hmis the hidden layer,bi1andbi2are the network offsets,andWistands for the network weight.Based on the binary assumption of(17),the decision rule of the output result of the MLP is:

where the output “01” indicates that the target exists,the output“10”denotes that the target does not exist.The MLP is pre-trained by multilayer neurons, uses the minimum mean square error as the cost function,and uses a backpropagation algorithm to fine-tune the weights and biases in the network.By inputting the detection statistics into the trained MLP, and the output result of the MLP is“01”or“10”.When outputuof the MLP is“01”,target is detected,andH1is true.When outputuof the sensor is “10”, no target is detected, andH0is true.The performance of the MLP can be denoted asP(u/Hj′),j′=0,1.

The false alarm probability of the MLP fully connected layer can be denoted asPi(1/H0) =Pfi, the detection probability of the MLP fully connected layer can be denoted asPi(1/H1) =Pdi, and the final detection probabilityPdis expressed as

whereS1={i|ui=1,∀i=1,··· ,h},S0={i|ui=0,∀i=1,··· ,h}.The overall false alarm probabilityPfis expressed as

V.NUMERICAL RESULTS AND DISCUSSION

In this section, simulation results are provided to validate the effectiveness of our proposed method.We use MATLAB (9.5.0.944444 (R2018b), MathWorks Company, Natick, MA, USA) to simulate the target detection method proposed in this paper.The type of satellite signal in the simulation experiment is a GPS satellite signal.In the following, we present the parameters settings.The sampling frequency is set tofs=15.7542GHz.The sampling duration is 100ms.The carrier frequency of the GPS satellite direct wave signal is 1.57542GHz.The time delay of the echo signal relative to the direct wave isτ=50ms.The echo signal’s frequency deviation relative to the direct wave isfd=600Hz.The MLP has five layers in which the number of hidden layer neurons is: 569, 156, 45, the number of input layer neurons is 1500,and the number of output layer neurons is 2.There are 12800 training samples and 6400 test samples.After 500 iterations,the trained networks are obtained,and the weights are saved,which will be applied to the test data.

Experiment 1: We evaluate the impact of network parameters on the direct wave suppression in the DFNs.Assume that the signal-to-noise ratio range of the echo signal and the clutter signal is -50dB～-20dB.The direct wave power of the GPS satellite signal is -100dBm.The corresponding echo power difference of the direct wave is 40dB.We analyze the network by changing the signal-to-interference ratio(SIR)S/Jis given by

whereprdenotes the received echo signal power,Pcis the power of clutter signal,Pdindicates the satellite direct wave signal power, andPnis the power of Gaussian white noise.

Figure 7 shows the impacts of SIR onS/Jbeforeand after the suppression of direct wave in the monitoring channel,where the number of neurons in DFNs is 50, the delay intervals are 1 sampling duration, 5 sampling duration,25 sampling duration,respectively.Figure 8 shows the change of the integrated SIR before and after the suppression of direct wave in the monitoring channel when the delay interval is 5 sampling duration,and the number of neurons is 20,50,and 75,respectively.

Figure 7. The delay interval of DFNs influence on direct wave suppression.

From Figure 7 and Figure 8, it is observed that the network parameters in the DFNs,such as the delay interval and the number of neurons,have little effect on the direct wave suppression.This is because DFNs has good stability, and it can well model the clutter.Meanwhile,the proposed method can effectively suppress direct waves for different SIRs.

Figure 8. The number of neurons of DFNs influence on direct wave suppression.

Experiment 2: To evaluate the impacts of the number of reservoir neurons on the detection performance,we conduct target detection under the condition that the SIRs of echo signal and clutter signal are within the range of-50～-20dB.Suppose that the direct GPS satellite signal power is -100dBm, and the difference between the echo power and the direct wave power is-40dB.The number of DESNs reservoirs is 4.The sparse density of the reservoir weight matrix is 0.3.The spectral radius is 0.7.The number of neurons in each reservoir is 100,250,400,and 550,respectively.The simulation results are shown in Figure 9.

Figure 9. Relationship between DESNs reservoir neuron number and detection performance.

Figure 9 shows the detection performance of DESNs for different numbers of reservoir neurons.When the SIR is -40dB, the detection performance with different number of reservoir neuron numbers is shown in Table 1.From Figure 9 and Table 1,it is observed that the probability of detection does not always increase as the number of neurons decreases in the low SIR regimes.This is because too many neurons cause over-learning, which reduces detection performance.When the number of reservoir neurons is 250,the detection performance is better in the low SIR regimes,the reservoir results of the DESNs can be optimized to improve the detection performance.After the parameters of the DESNs are fixed,the detection probability of the echo also gradually increases as the echo SIR increases.Moreover,the detection accuracy tend to be stable when the SIR more than-35dB.Simulation results show that the proposed method is practical and feasible for passive detection using satellite signals of opportunity in the presence of clutter.

Table 1. Detection performance of DESNs with different number of reservoir neuron numbers.

Experiment 3: We evaluate the impact of sampling points of the input signal of the MLP on the detection performance of the satellite echo signal.For the MLP detection method,we consider that the sampling points are 300 points, 600 points, 1200 points, and 1500 points, the GPS satellite signal’s direct wave power is -100dBm, and the difference between the echo power and the direct wave power is -40dB.The detection results are shown in Figure 10.

In Figure 10, under the same SIR, the detection probability improves as the number of sampling points increases.Since the detection performance with 1500 sampling points is better than that with other selected sampling points,we use 1500 sampling points for target detection.

Figure 10. Relationship between the number of detection points and the detection probability.

Experiment 4: We evaluate the impacts of different detection methods on the detection performance with satellite echo when the number of sampling points is 1500.In this case, the SIRs of the echo signal and the clutter signal range from-50dB to-20dB.The proposed method in this paper is compared with the one using a threshold for data decision in [35].The detection performance of ESNs is also compared.The direct wave power of the GPS satellite signal is -100dBm.The difference between the echo power and the direct wave power is -40dB.The number of neurons in the DESNs is 250,and the number of neurons in the ESNs are 1500 and 900, respectively.Simulation results are shown in Figure 11.

Figure 11. Performance comparison with different detection methods.

In Figure 11, it is observed that, under the same SIR,the proposed method is superior to the other two methods.The four reservoirs of DESNs with repeated weight matrices reduce the number of reservoir neurons that need to be trained and use MLP to avoid setting thresholds, which improves the detection performance.Meanwhile, the detection probability of the echo signal reaches almost 100%when the SIR is-36 dB.The results demonstrate that the proposed method has an excellent detection performance for passive aerial target detection under low SIR.

The experiments are conducted using an Intel Core i9-9920X CPU and two GIGABYTE GeForce RTX2080 graphics PC.The training samples used for detection are 12,800, and the test samples are 6,400.The off-line training time is 240666.20606 seconds,and the detection time is 120217.87297 seconds.In comparison,the ESNs+MLP method’s off-line training time is 275964.50101 seconds, and the time used for detection is 137878.51781 seconds.The number of samples used in[33]is 6400,and the detection time is 120220.67283 seconds.The online time complexity of the proposed method is lower than the method using ESNs+MLP,and it is slightly lower than the time complexity of the method proposed in[33].

VI.CONCLUSIONS

In order to enhance the accuracy and reliability of moving aerial target passive detection,a novel framework of the passive detection system in space-airground integrated networks has been designed and corresponding passive detection method for aerial target based on reservoir computing networks has been proposed in this paper.Simulation results have shown that the delayed feedback network’s parameters have little effect on the direct wave suppression, and that the network is relatively stable.The proposed method in this paper is better than the existing methods.Meanwhile,the detection performance of the proposed method reaches 100%when the signal-to-noise ratio is-36dB.Therefore,the proposed method can effectively detect air targets in typical space-air-ground integrated networks scenarios.

ACKNOWLEDGEMENT

This work was supported by the National Natural Science Foundation of China under Grant 62071364, in part by the Aeronautical Science Foundation of China under Grant 2020Z073081001, in part by the Fundamental Research Funds for the Central Universities under Grant JB210104,in part by the Shaanxi Provincial Key Research and Development Program under Grant 2019GY-043,and in part by the 111 Project under Grant B08038.