CN109188470B - GNSS deception jamming detection method based on convolutional neural network - Google Patents

Abstract

The invention discloses a GNSS deception jamming detection method based on a convolutional neural network, which comprises the following steps: 1) in the signal capturing stage, the number N of correlation peaks larger than the capturing threshold in a two-dimensional matrix A generated in the signal capturing process is detected_peakIf N is present_peakNumber greater than 2, the deception signal is considered to exist, if N_peak<2, continuing to carry out the step 2); 2) intercepting data in +/-2 chip regions on a phase axis of a related peak pseudo code of the two-dimensional matrix A to obtain a detection matrix A_sAnd after data preprocessing, carrying out detection training and classification by a convolutional neural network to finally obtain a detection result. The detection method has the advantages of good detection effect, strong applicability, moderate complexity when the detection opportunity is in a signal capture stage; the problem that the phase difference delta T between the false code of the deception signal and the true signal is difficult to detect when the phase difference delta T is within 2 chips can be solved.

Description

GNSS deception jamming detection method based on convolutional neural network

Technical Field

The invention belongs to the technical field of interference detection in a satellite navigation system, and particularly relates to a GNSS deception interference detection method based on a Convolutional Neural Network (CNN).

Background

The Global Navigation Satellite System (GNSS) is a Navigation System with wide coverage, all weather, real time and high precision. With the continuous development of satellite navigation technology, GNSS is widely used in various military and civil facilities. With the development of satellite navigation positioning technology, the number of users and application scenes thereof are continuously increased, and the safety and reliability are more and more emphasized by people. The security threats faced by current satellite navigation systems can be largely classified into unintentional interference and intentional interference. Intentional interference mainly refers to man-made malicious interference, and can be divided into suppressed interference, deceptive interference and combined interference. The pressing type interference means that a strong power interference signal is applied to a satellite frequency band, so that a receiver cannot receive the satellite signal. The deceptive jamming is to transmit a signal which is the same as or similar to the navigation satellite but has stronger power, and a receiving terminal of a satellite navigation system user may misunderstand the signal as being transmitted by a real navigation satellite and acquire and track the signal, so that the receiving terminal generates wrong information or no information is output. Compared with the traditional strong power suppression type interference, the deception type interference has the advantages of strong concealment, equipment miniaturization, high interference efficiency and the like, and the detection difficulty is high.

The existing detection method of the deceptive interference mainly comprises two aspects: firstly, based on multiple antennas, detection is performed through spatial domain characteristics of signals, that is, the incoming wave directions of signals of multiple satellites are detected for judgment, and if multiple prn (pseudo Random Noise code) signals come from the same direction, a spoofed signal is considered to exist in the satellite signals. Secondly, based on signal processing, detection is carried out through characteristics of signals such as time domain, frequency domain, power and the like, including signal absolute power detection, carrier-to-noise ratio detection, signal quality monitoring, detection based on RAIM (receiver Autonomous Integrity monitoring) and the like, the methods are strong in applicability, but the accuracy of signal absolute power detection and carrier-to-noise ratio detection is poor, and detection based on RAIM needs to be carried out on signals, so that the complexity is high, and the timeliness is not strong. The signal quality monitoring mainly judges whether a deceptive signal exists or not by judging whether a signal correlation peak is distorted or not, and the existing signal quality monitoring method mainly comprises the methods of detecting the quantity of the correlation peaks during signal capturing, analyzing the output of a correlator during signal tracking and the like; the current method for detecting the number of correlation peaks in signal acquisition is difficult to distinguish when the phase difference between the false code of the deceptive signal and the true signal is less than 2 chips.

Disclosure of Invention

The invention aims to provide a GNSS deception jamming detection method based on a convolutional neural network, so as to solve the existing technical problems. The GNSS deception jamming detection method is based on the convolutional neural network, a two-dimensional matrix generated by the GNSS receiver during signal capture is used as detection data, related peak adjacent data is intercepted and sent to a CNN algorithm for key detection, data information can be effectively utilized, discrimination can be realized when the phase difference between a deception signal and a true signal pseudo code is less than 2 chips, and the existence of the deception signal can be detected with high accuracy.

In order to achieve the purpose, the invention adopts the following technical scheme:

a GNSS deception jamming detection method based on a convolutional neural network comprises the following steps:

step 2.1, intercepting correlation peak A on two-dimensional search matrix A_peak1Data acquisition detection matrix A in the surrounding area_s(ii) a The two-dimensional search matrix A is a two-dimensional search matrix which is generated by the GNSS receiver in a signal capturing stage and takes Doppler frequency shift and a pseudo code phase as axes; the range of the surrounding area is +/-2 kHz on a Doppler frequency shift axis and +/-2 chips on a pseudo code phase axis; detection matrix A_sIs m in size_s×n_sWherein m is_s＝4/Δf_D+1，n_s＝4/ΔT_C+1，m_sDenotes the length of the matrix on the truncated Doppler frequency shift axis, Δ f_DSearch step size, n, for Doppler shift_sRepresenting the length, Δ T, of the matrix on the phase axis of the truncated pseudo-code_CSearching a step size for the pseudo code phase;

step 2.2, detecting matrix A obtained in step 2.1_sCarrying out pretreatment; the preprocessing includes detecting the matrix A_sMiddle lower than threshold value lambda_PSetting the value of (d) to zero;

step 2.3, detecting matrix A after pretreatment in step 2.2_sInputting a convolutional neural network model to be trained for parameter training and updating to obtain the convolutional neural network model for detecting GNSS deception interference;

step 2.4, obtaining a detection matrix A generated by detecting the GNSS signal to be detected by the convolutional neural network model for detecting the GNSS deception interference through the step 2.3_sAnd obtaining a detection result.

Further, a threshold value λ in step 2.2_PThe values of (A) are as follows: lambda [ alpha ]_P＝2A_mean(ii) a In the formula, A_meanThe mean of the matrix a is searched for two dimensions.

Further, the method also comprises the following steps:

step 1, traversing a two-dimensional search matrix A, and detecting the number N of correlation peaks larger than a capture threshold_peak(ii) a If N is present_peakMore than or equal to 2, considering that 2 or more correlation peaks exist in the signal, and obtaining a detection result of the existence of the deception jamming signal; if N is present_peak1, jumping to step 2.1; the two-dimensional search matrix A is a two-dimensional matrix which takes Doppler frequency shift and pseudo code phase as axes and is generated by the GNSS receiver in a signal acquisition stage.

Further, step 1 is carried out with A_peak/A_mean＞λ_acqFor the capture conditions, A_peakFor two-dimensional search of the peak value of the matrix A, A_meanFor two-dimensional search of the mean value, λ, of the matrix A_acqIn order to capture the threshold, the method specifically comprises the following steps:

step 1.1, calculating the mean value A of the two-dimensional search matrix A_meanThe calculation formula is

N_AIs the number of elements in A, A_ijSearching the ith row and the jth column of the element in the matrix A in two dimensions;

step 1.2, traversing the two-dimensional search matrix A to obtain a correlation peak A_peak1If A is_peak1/A_mean＞λ_acqThen there is a correlation peak number N _peak1, jumping to step 1.3; if N is present_peakWhen the received signal is not the signal of the current pseudo-random code, the detection is finished;

step 1.3, data processing is carried out on the two-dimensional search matrix A: a is to be_peak1Data in a +/-1 kHz chip area on a Doppler frequency shift axis and a +/-1 chip area on a pseudo code phase axis are set to be zero; searching a two-dimensional matrix A 'obtained after data zero setting to obtain a second correlation peak A'_peak(ii) a If A'_peak/A_mean＞λ_acqThen N is_peakConsidering that 2 correlation peaks exist in the received signal, and obtaining a detection result of the existence of the deception jamming signal; whether or notThen N is_peakStep 2.1 is skipped to 1.

Further, in the convolutional neural network model detection training of the step 2.3, the input layer converts the input data X into a convolutional layer C through convolution operation; the convolution operation process is as follows:

in the formula, W_KAs a convolution kernel, b_KFor the bias parameter, f (.) is the activation function, the convolution kernel W_KParameter (b) and bias parameter (b)_KIs a trainable parameter; the superscript k represents the bias parameters of the kth group of data; the subscript K represents W_KAnd b_KTraining parameters for convolution operation;

the convolution layer C is converted into a characteristic P through pooling operation;

the full connection layer integrates the characteristics output by the convolution layer and the pooling layer, and pulls the characteristics P into a column vector Fv; fv obtains an output result O after being calculated by a softmax function, wherein the O represents the probability that the result is each classification label; the calculation process is as follows:

O＝softmax(O_o),O_o＝f(W_o ^TFv+b_o)

wherein O is_oFor network output, f (.) is an output layer activation function; the calculation process of the softmax function is as follows:

r_χdenotes the x variable; w_oAs parameters of the matrix between the fully connected layer and the output layer, b_oIs a bias parameter; w_oAnd b_oIs a trainable parameter; the subscript o denotes W_oAnd b_oFor calculating O_oThe training parameters of (1).

Further, average pooling operation is adopted in the process that the convolutional layer C is changed into the characteristic P through pooling operation;

the average pooling operation process is as follows:

wherein S is a pooling window of size S₀@S₁×S₂，S₀、S₁、S₂Number, length and width of pooling windows, respectively, @ for separating the number, length x width; the subscripts τ and ν denote the τ -th row of P and the ν -th column of elements.

Further, Δ T_CLess than or equal to 1 chip.

Compared with the prior art, the invention has the following beneficial effects:

the invention relates to a GNSS deception jamming detection method based on a convolutional neural network, which applies the convolutional neural network to GNSS deception jamming detection and is a deception jamming signal detection method for detecting the number of related peaks during signal capture, compared with the existing detection method, the method intercepts the adjacent data of the related peaks and passes the adjacent data of the related peaks to a CNN algorithm for secondary key detection, can utilize data information more effectively, can distinguish more accurately when the phase difference between the deception signals and the true signals and pseudo codes is less than 2 chips, and can detect the existence of the deception signals with higher accuracy; the detection precision and accuracy can be adjusted by adjusting the pseudo code phase search step length, and the smaller the step length value is, the better the detection effect is; after the parameters of the CNN model are trained by using training data, the model can be directly used for classification calculation, and the CNN algorithm is adopted at a higher speed; the detection accuracy of the invention is higher than that of the direct absolute power detection and the method of using MLP for classification by using the signal tracking correlator output.

Drawings

FIG. 1 is a schematic diagram of a conventional satellite navigation system repeater spoofing interference model;

FIG. 2 is a schematic diagram of a two-dimensional search for signal acquisition of a conventional navigation satellite receiver;

FIG. 3 is a schematic block diagram of a signal acquisition principle of a conventional navigation satellite receiver based on an FFT algorithm;

FIG. 4 is a diagram illustrating an acquisition result of a satellite navigation signal only in a conventional received signal;

FIG. 5 is a diagram illustrating the results of capturing satellite navigation signals and spoofed signals simultaneously present in a received signal according to the prior art;

FIG. 6 is a schematic block diagram of a flow chart of a GNSS spoofing interference detection method based on a convolutional neural network according to the present invention;

FIG. 7 shows a convolutional neural network at A in an embodiment of the present invention_sThe size is 9 multiplied by 9;

FIG. 8 shows a GNSS deception jamming detection method based on convolutional neural network according to the present invention, which adopts different Δ T_CComparing the detection results with a schematic diagram;

FIG. 9 is a schematic diagram illustrating comparison of detection effects of different models adopted by a GNSS deception jamming detection method based on a convolutional neural network according to the present invention;

FIG. 10 is a schematic diagram showing comparison of detection results of image classification using the CNN algorithm and the KNN algorithm of the present invention;

fig. 11 is a comparison of the CNN detection method of the present invention and a spoofed signal verification method based on signal processing.

Detailed Description

The invention is described in further detail below with reference to the figures and specific examples.

Referring to fig. 1 to fig. 6, a GNSS spoofing interference detection method based on a convolutional neural network according to the present invention is based on a system model of a satellite navigation network: satellite navigation signals exist all the time, deception jamming signals may exist, namely the system has two conditions; h₀: the GNSS receiver receives signals, and only satellite navigation real signals exist in the signals; h₁: the received signal contains a satellite navigation real signal and a deception signal at the same time, the deception signal simulates parameters such as a pseudo code phase, a Doppler frequency shift and the like of the real signal, and the power of the deception signal is slightly higher than that of the real signal, so that the deception signal can be captured with higher probability when a GNSS receiver captures the deception signal.

After receiving satellite signals, the GNSS receiver converts the signals into intermediate frequency signals through down-conversion:

wherein s is_LS(t) is the direct satellite navigation signal, s_MP(t) is a multipath signal, n (t) is a noise signal, s_Spoof(t) is a spoof signal.

Wherein P is signal power, C (t) is pseudo code (CA code), D (t) is navigation data, f_IFIs a theoretical intermediate frequency, f_DopplerFor doppler shift, Φ is the initial pseudo-code phase,

for initial carrier phase, the corner mark m_kRepresenting the kth multipath signal and M representing the total of M multipath signals.

The GNSS receiver captures the intermediate frequency signal, and the signal capturing methods include a time domain correlator-based method, a matched filter-based method, an FFT-based method and the like, and all generate a two-dimensional matrix for searching a correlation peak and roughly estimating Doppler frequency shift and a pseudo code phase of the satellite navigation signal. In the signal acquisition stage, the receiver searches all the PRNs in sequence to generate a two-dimensional array with Doppler frequency shift and pseudo code phase as axes, namely a two-dimensional search matrix A with the size of mxn, wherein m represents the data length of the Doppler frequency shift axis, and m ═ f (f)_Dmax-f_Dmin)/Δf_D+1，f_{Doppler_range}Search range for Doppler shift, f_DmaxAnd f_DminAre respectively f_{Doppler_range}Upper and lower limits of, Δ f_DFor the doppler shift search step, n represents the length of the pseudo code phase axis,

T_{CAcode_range}for the search range of the pseudo-code phase,

is T_{CAcode_range}Length of (1), Δ T_CThe step size is searched for the pseudo code phase. As shown in FIG. 2, T is the GPS signal_{CAcode_range}Is [1,1023 ]]Search range of Doppler shift is f_{Doppler_range}＝[-7kHz,7kHz]。

In the simulation of the present invention, an FFT-based acquisition method is adopted, and the signal acquisition process is shown in fig. 3:

(1) setting local carrier frequency to f_w＝f_IF+f_Dmin+(i-1)Δf_D，i＝1。f_IFThe theoretical intermediate frequency after down-conversion of the satellite signal.

(2) To receive the intermediate frequency signal S_R(t) is multiplied by the in-phase and quadrature signals output by the local carrier generator, and a complex signal s (t) of the baseband is obtained by a low-pass filter, wherein,

I. q is the in-phase and quadrature signals, respectively, and T is the capture period.

1) FFT is performed on s (t) to obtain S (f).

2) The local pseudo-code generator generates a pseudo-code signal C (t) according to the current PRN, performs FFT and obtains C by conjugation^*(f)。

3) Mixing S (f) and C^*(f) Multiplying and performing IFFT to obtain a one-dimensional array

The size is n × 1.

4) If i<m, then i ═ i +1, f_w＝f_IF+f_Dmin+(i-1)Δf_DContinuing with step 2); if i is m, then

Two-dimensional search of set compositionThe matrix a is a matrix of a,

the size is m × n.

When the satellite navigation signal of the current PRN does not exist in the received signal, no correlation peak meeting the acquisition condition exists in the A; when only the satellite navigation signal of the current PRN is present in the received signal, there are only 1 correlation peaks in the two-dimensional search matrix a that are greater than the acquisition threshold, as shown in fig. 4. When the satellite navigation signal of the current PRN and the spoofed signal exist in the received signal at the same time, the two-dimensional matrix generated by signal acquisition will have 2 or more correlation peaks greater than the detection threshold, as shown in fig. 5, and when the spoofed signal is shifted less from the true signal, the correlation peaks may completely overlap or partially overlap, which makes detection of spoofed interference difficult.

The detection method provided by the invention mainly aims at the problem that the phase difference delta T between the spurious signals and the true signals is difficult to detect when the phase difference delta T is within 2 chips, and detects whether the situation that correlation peaks are overlapped exists or not by intercepting data in a plus or minus 2 chip area on a maximum correlation peak spurious code phase axis, namely whether the correlation peaks generated by the spurious interference signals exist or not. The invention discloses a GNSS deception jamming detection method based on a convolutional neural network, which specifically comprises the following steps:

step 1, in a signal capturing stage, a GNSS receiver generates a two-dimensional search matrix a with doppler frequency shift and a pseudo code phase as axes. Traversing two-dimensional search matrix A to detect the number N of related peaks larger than the capture threshold_peakIf N is present_peak2 or more, considering that 2 or more correlation peaks exist in the signal and a deception jamming signal exists; if N is present_peakIf no spoofed interference signal is detected, step 2 is skipped.

The step 1 specifically includes:

firstly, detecting the number N of correlation peaks larger than the capture threshold_peak

With A_peak/A_mean＞λ_acqFor the capture conditions, A_peakFor two-dimensional search of the peak value of the matrix A, A_meanFor two-dimensional search of the mean value, λ, of the matrix A_acqFor capturingAnd (4) threshold. The simulation method comprises the following steps:

1.1) determining the mean value of A

N_AIs the number of elements in A, A_ijThe ith row and jth column elements in the matrix A are searched for in two dimensions.

1.2) traversing the two-dimensional search matrix A to obtain the maximum value A_peak1If A is_peak1/A_mean＞λ_acqThen the number of correlation peaks N_peakStep 1.3) is continued, else N_peakThe signal is considered to have no current PRN present in the signal at 0.

1.3) mixing A_peak1And data in a +/-1 chip area on a Doppler frequency shift axis and a +/-1 chip area on a pseudo code phase axis are set to be zero. Searching A to obtain the maximum value A_peak2If A is_peak2/A_mean＞λ_acqThen N is_peakConsidering 2 correlation peaks in the signal, a deceptive jamming signal exists, otherwise N_peakNo spoofed jamming signal is detected, continuing with step 2).

Step 2, intercepting data in a +/-2 chip area on a phase axis of a related peak pseudo code of the two-dimensional search matrix A to obtain a detection matrix A_sAfter data preprocessing, carrying out detection training and classification by a Convolutional Neural Network (CNN); and detecting training until the data is used up or a preset convergence condition is reached.

The implementation method of the step 2 is as follows:

step 2.1, intercepting correlation peak A on two-dimensional search matrix A_peak1Data in a +/-2 kHz area on a Doppler frequency shift axis and a +/-2 chip area on a pseudo code phase axis around the position are obtained to obtain a new detection matrix A_sSize is m_s×n_sWherein m is_s＝4/Δf_D+1，n_s＝4/ΔT_C+1。

Step 2.2, data is preprocessed, since in matrix A_sOnly data near the correlation peak are correlated with the detection, so a is_sMiddle lower than threshold value lambda_PThe value of (d) is set to zero.

Step 2.3, detecting the matrix A_sAnd (5) carrying out detection training and verification by the CNN. Structure of CNN model according to A_sIs preset. Firstly, training the parameters of a CNN model with a set structure by using training data of known scene classification, wherein the training cutoff condition is that the iteration times are finished or the error is not changed or the error meets the requirement, and after the training is finished, inputting test data to test the detection effect of the CNN model.

The step 2 specifically comprises the following steps:

2) and intercepting the two-dimensional search matrix A, and submitting the two-dimensional search matrix A as an image to CNN for detection training and classification.

2.1) truncating A on the two-dimensional search matrix A_peak1Data in a +/-2 kHz area on a Doppler frequency shift axis and a +/-2 chip area on a pseudo code phase axis around the position are obtained to obtain a new matrix A_sSize is m_s×n_sWherein m is_s＝4/Δf_D+1，n_s＝4/ΔT_C+1. In general, Δ f is used in GNSS receivers_DSet at 0.5kHz, Δ T_CSet to 0.5 chips.

2.2) preprocessing the data, since in matrix A_sOnly data near the correlation peak are correlated with the detection, so a is_sMedium below lambda_PSetting the value of (d) to zero; wherein in the simulation_P＝2A_mean。

2.3) will detect matrix A_sAnd (5) carrying out detection training and verification by the CNN. Structure of CNN model according to A_sIs preset. The structure of the CNN model needs to be considered as mainly setting the number and arrangement of convolutional layers and pooling layers and the sizes of convolutional cores and pooling windows, and the setting principle is generally as follows: 1. the size of the output image is close to the regular image with the same length and width through the current layer operation; 2. the size of the output layer is not decimal through the operation of each layer, wherein the operation of the convolution layer is subtraction, M_C＝M_I-K₁+1，N_C＝N_I-K₂+1, the operation of the pooling layer is division, M_P＝M_I/S₁，N_P＝N_I/S₂M, N respectivelyIndicates the length and width of the current layer, subscript I indicates the input layer, C indicates the convolutional layer, P indicates the pooling layer, K₁、K₂Respectively, the length and width of the convolution kernel. S₁、S₂Respectively, the length and width of the pooling window. For example, when A is_sThe structure of the CNN model is shown in fig. 7 when the size is 9 × 9, and the processing procedure is as follows:

2.3.1) input layer X, size F_x@M_x×N_x1@9 × 9, i.e., 1 image of 9 × 9 per input sample.

2.3.2) are convolved into convolution layer C with size F_C@M_C×N_C4@4 × 4, wherein F_C＝K₀，M_C＝M_I-K₁+1，N_C＝N_I-K₂+1. The process is as follows:

C＝f((∑W_K*X)+b_K)

wherein, W_KIs a convolution kernel of size K₀@K₁×K₂＝4@6×6，K₀、K₁、K₂The number, length and width of the convolution kernels, respectively. b_KIs a bias parameter with a size of K₀X 1, f (.) is the activation function, with sigmoid, tanh, Relu, etc. Convolution kernel W_KParameter (b) and bias parameter (b)_KAre trainable parameters.

2.3.3) the pooling has the methods of maximum pooling, average pooling and the like, and the method of average pooling is adopted at this time. After pooling operation, the obtained product is changed into a pooling layer P with a size of F_P@M_P×N_P4@2 × 2, wherein F_C＝S₀，M_P＝M_C/S₁，N_P＝N_C/S₂. The process is as follows:

wherein S is a pooling window of size S₀@S₁×S₂＝4@2×2，S₀、S₁、S₂Number, length and width of pooling windows, respectively。

2.3.4) full connectivity layer integrates features in the image feature map through the convolutional and pooling layers, pulling the image features into a column vector Fv of size F_Fv@M_Fv×N_Fv1@1 × 16, wherein N_Fv＝F_PM_PN_P。

2.3.5) Fv is calculated by a softmax function to obtain an output result O with the size of F_O@M_O×N_O1@1 × 2, wherein N_OEqual to the number of classification labels, and O represents the probability that the result is each classification label. The process is as follows:

O＝softmax(O_o),O_o＝f(W_o ^TFv+b_o)

wherein O is_oFor network output, f () is the output layer activation function. The calculation process of the softmax function is as follows:

the objective is to calculate the probability that x is labeled j and make the sum of the probabilities 1. W_oAs parameters of the matrix between the fully connected layer and the output layer, b_oIs a bias parameter. W_oAnd b_oAre trainable parameters.

2.3.6) CNN updates parameters using BP (error Back propagation) Back-propagation algorithm. MSE (mean Square error) is used as a transmission error E, and the transmission error E is propagated reversely. Using a squared error cost function:

where Y is the actual classification label of the sample data.

The main formula for updating the parameters is as follows:

wherein W is the updated connection parameter, W_iThe initial connection parameter is η, and the learning efficiency is η. To W_K、b_K、W_oAnd b_oAnd solving the partial derivatives and updating the parameters.

Returning to the full connection layer from the output layer, and performing error transmission and parameter updating:

O_d＝E⊙f′(O_o)，

wherein O is_dResidual error of F size, representing output layer return_o@M_o×N_o1@2 × 1, f' () denotes a derivative of f (, an |), indicates an almar inner product, i.e., a vector multiplication of corresponding positions, Fv_dResidual error of size F, representing full link layer return_Fv@M_Fv×N_Fv＝1@16×1。

Returning Fv from full link layer to pooling layer for error propagation_dChange to P_dAnd the error is reversed, the size is the size of the upper pooling layer, F_P@M_P×N_P＝4@2×2

Returning the convolution layer from the pooling layer, and performing error transfer and parameter updating:

C_d＝f′(C)⊙up(P_d)，

C_dresidual error of convolution layer return, size F_C@M_C×N_C＝4@4×4，

Representing upsampling.

The process of forward propagation, loss calculation, backward propagation and parameter update is one iteration (epoch) every time a set of data is input.

And after the parameters of the CNN model are updated through training parameters, test data are used as input to the CNN model for detection and classification, the detection effect is verified, and the CNN model for detecting GNSS deception interference is obtained.

The Convolutional Neural Network (CNN) is one kind of artificial neural network, is widely applied to the fields of image recognition, voice recognition and the like in recent years, and has a simple structureSingle, less training parameters, strong adaptability and the like. The method introduces the thought of deep learning into the neural network, extracts the characteristics of different levels of the image from shallow to deep through convolution operation, and enables the whole network to automatically adjust the parameters of a convolution kernel by utilizing the training process of the neural network. The convolutional neural network comprises a characteristic extractor consisting of a convolutional layer and a sub-sampling layer, and the method of local connection, weight sharing and sub-sampling is adopted, so that the model complexity is simplified, the parameter quantity is reduced, and the risk of overfitting is reduced. The invention adopts a CNN algorithm to realize deception jamming detection. At present, the multi-antenna based method mainly aims at the situation that a single deception jamming source transmits a plurality of prn (pseudo Random Noise code) signals, the application of the method has limitations, and the use of multiple antennas increases the cost of a civil satellite navigation receiver; the method for analyzing the output of the correlator during signal tracking is equivalent to the method for manually extracting signal characteristics, cannot fully utilize signal information, and has lower detection accuracy when the signal receiving power is unstable and the signal-to-noise ratio is small. The method for detecting the number of correlation peaks in signal acquisition is difficult to distinguish when the phase difference between the false code of the deceptive signal and the true signal is less than 2 chips. The method of the invention utilizes a two-dimensional matrix A to detect whether deception signals exist in the received signals, and the detection is mainly based on whether two or more signals exist in the same PRN mark, namely whether two or more correlation peaks exist in the two-dimensional matrix A. Firstly, detecting the number N of correlation peaks larger than an acquisition threshold lambda_peakIf N is present_peakIf the number of the detected signals is more than or equal to 2, considering that deception signals exist, and obtaining a detection result; if N is present_peak<2, continuing to execute the subsequent steps; the problem that the phase difference (delta T for short) between the deception signal and the true signal pseudo code is difficult to detect when the phase difference is within 2 chips is solved by intercepting the two-dimensional matrix A as an image and submitting the image to CNN for detection training and classification. The detection method provided by the invention is a method based on signal quality monitoring, and during signal capture, a two-dimensional array generated by a GNSS receiver is used as detection data, and related peak adjacent data is intercepted and sent to a CNN algorithm for key detection. Here, we regard the two-dimensional array as an equivalent image, and use the advantage of CNN in image recognition to improve the efficiency of detecting the spoofed interference signals.The detection method has strong applicability, forward opportunity and moderate complexity; the common use of software receivers offers the possibility of application of the present detection method.

In order to verify the performance of the GNSS deception jamming detection method based on the convolutional neural network, the following simulation experiment is carried out:

the intermediate frequency signal of the analog GNSS receiver is a GPS satellite navigation signal with the sampling frequency of 16.3676MHz and the theoretical intermediate frequency of 4.1304MHz, does not concern the message, and randomly generates message data D (t). The simulated satellite navigation signal contains direct signals and multipath signals, one path of multipath signals is simulated, the fading is more than 3dB, and the SNR of the receiving signal to noise ratio of a receiver is-20 to-15 dB. The simulated deception signal and the real signal only have different Doppler frequency shift, pseudo code phase and power, the Doppler frequency shift difference delta f randomly changes in a +/-1 kHz interval, the pseudo code phase difference delta T changes in a +/-2 chip interval, and the power is 1.1-1.5 dB greater than that of the direct signal. The simulation data totals 415800 groups of data and is divided into H₀: the GNSS receiver receives signals, wherein only real satellite navigation signals exist in the signals; h₁: the GNSS receiver receives 207900 groups of signals with two scenes of satellite navigation true signals and deception signals simultaneously. Wherein H₁The data of the scene is distinguished according to the pseudo code phase difference delta T of the deception signal and the real signal, the delta T value interval is 0 to 2 chips, the step length is 0.1 chip, 21 types are counted, and 9900 groups of data of each type are grouped. 207900 groups of the two scenes and 1/2 corresponding to various types of data are taken as training data; of the remaining 1/2 data, H₁The data of the scene are divided into 21 classes according to the delta T, each class has 4950 groups of data, H₀The data of the scene is divided into 21 parts, and each type of 4950 groups of data is combined with the 21 types of data to obtain 21 types of test data and 9900 groups of data.

The receiver uses an FFT-based acquisition method, f_{Doppler_range}Taking [ -7kHz,7kHz]，Δf_DTaking 0.5kHz, T_{CAcode_range}1023. During the detection,. DELTA.T_CFor the pseudo code phase search step, 1 chip, 0.5 chip, 0.25 chip and 0.1 chip are taken respectively for comparison.

We chose different options for the process of the inventionPseudo code phase search step Δ T_CComparing the detection effect of the time, and detecting the matrix A_sThe sizes of (a) and the CNN model structure used in the simulation are shown in table 1:

TABLE 1 intercept sampling frequency and CNN model structure comparison table in simulation

As shown in FIG. 8, the simulation respectively simulates that the method of the present invention takes different pseudo code phase search step lengths Δ T_CAnd detecting the deception signals with different pseudo code phase differences. The phase difference of the pseudo code of the deception signal compared with the real signal is 0-2 chips, and the Doppler frequency shift difference takes a value randomly in a +/-1 kHz area. It can be seen that the detection effect is poor when the phase offset between the spoofed signal and the real signal is small, and the detection effect is better and better as the phase offset of the pseudo code is increased and the difference between the spoofed signal and the real signal is larger and larger. Overall, the step size Δ T_CThe smaller the value, the better the detection effect. Delta T_CAt 1 chip, the detection probability is low when the pseudo code phase offset difference is around 0.5 chip and 1.5 chip. This is because the offset of 0.5 chip cannot be recognized efficiently with the recognition accuracy of 1 chip, and the distance N · Δ T_CThe farther N ∈ R, the worse the detection effect. As shown in tables 1 and 2,. DELTA.T_CThe smaller the value, the more complex and time the detection algorithm. In practical application, the detection effect and the speed requirement can be selected in a trade-off mode.

Table 2 simulation single data detection speed comparison table

In FIG. 9, we take Δ T_C0.5 chips, when the detection matrix a is present_sSize of [9,9 ]]Comparing the detection effects when the CNN model adopts different structures, the method comprises the following steps:

structure 1: { input, C1-2 @6 × 6, S2-2 @2 × 2, fullconnection, output };

structure 2: { input, C1-2 @4 × 4, S2-2 @2 × 2, fullconnection, output };

structure 3: { input, C1-2 @2 × 2, S2-2 @2 × 2, fullconnection, output };

structure 4: { input, C1-4 @6 × 6, S2-4 @2 × 2, fullconnection, output }.

As can be seen from fig. 9, the detection probability of the spoof signal is substantially the same with different models. When the phase difference delta T of the pseudo code is smaller than 1 chip, the difference is slightly different, the basic rule is that the larger the convolution kernel is, the more training parameters are, the better the detection effect is, but when the trainable parameters are increased, the training time and the detection time are also increased, and the detection effect and the complexity need to be balanced.

The invention uses a two-dimensional matrix for classification in a signal capturing stage, and in fig. 10, the influence of different image classification algorithms on the detection effect is compared. The KNN (k-nearest neighbor) algorithm is also a commonly used image recognition algorithm, and calculates images to be recognized and all training images, finds k training images closest to the images to be recognized, and takes the label with the most corresponding labels as a final result. As can be seen from FIG. 10, the classification by CNN algorithm is more accurate than the classification by KNN algorithm when the Δ T is larger_CMeanwhile, the detection probability of the CNN is higher than that of the KNN. As can be seen from table 2, the CNN algorithm is faster than the KNN algorithm, because the parameters of the CNN model can be directly classified and calculated by the model after being trained by using the training data, and the KNN algorithm is faster and more accurate when comparing all training data with the data to be identified during each classification and calculation.

In fig. 11, we have done several signal processing based methodsIn comparison, direct absolute power Detection, classification by using MLP (Detection of spherical impedance using Machine Learning based on Multi-Layer Neural Network in Single-Frequency GPS Receivers, Shafiee E, Mosavi M R, Moazedi M.2017) and classification by using CNN by using a two-dimensional matrix in a signal capturing stage are respectively adopted. The method for classifying by using MLP and using signal tracking correlator output mainly comprises the steps of using in-phase and quadrature early, timely and late correlation values I output by the signal tracking correlator_E,I_P,I_L,Q_E,Q_P,Q_LObtaining Delta characteristic x through the calculation of formulas (1), (2) and (3)₁Early-late feature x₂Signal level characteristic x₃And training and judging by an MLP (Multi-Layer Perceptron) neural network.

As can be seen from FIG. 11, the detection accuracy of the method of the present invention is higher than that of the two comparison schemes. Although the method of the present invention is longer than the two comparison schemes in detection time, compared with the method of using the signal tracking correlator output to be classified by MLP recognition, the method of the present invention can perform detection after signal acquisition, and the method of using the signal tracking correlator output needs to perform detection after the signal tracking correlator output is generated, so the method of the present invention does not fall behind the method of using the signal tracking correlator output in detection result time.

While the invention has been described in further detail with reference to specific preferred embodiments thereof, it will be understood by those skilled in the art that various changes in form and details may be made therein without departing from the spirit and scope of the invention as defined by the appended claims.

Claims (7)

1. A GNSS deception jamming detection method based on a convolutional neural network is characterized by comprising the following steps:

step 2.1, intercepting correlation peak A on two-dimensional search matrix A_peak1Data acquisition detection matrix A in the surrounding area_s(ii) a The two-dimensional search matrix A is a two-dimensional search matrix which is generated by the GNSS receiver in a signal capturing stage and takes Doppler frequency shift and a pseudo code phase as axes; the range of the surrounding area is +/-2 kHz on a Doppler frequency shift axis and +/-2 chips on a pseudo code phase axis; detection matrix A_sIs m in size_s×n_sWherein m is_s＝4/Δf_D+1，n_s＝4/ΔT_C+1，m_sDenotes the length of the matrix on the truncated Doppler frequency shift axis, Δ f_DSearch step size, n, for Doppler shift_sRepresenting the length, Δ T, of the matrix on the phase axis of the truncated pseudo-code_CSearching a step size for the pseudo code phase;

step 2.2, detecting matrix A obtained in step 2.1_sCarrying out pretreatment; the preprocessing includes detecting the matrix A_sMiddle lower than threshold value lambda_PSetting the value of (d) to zero;

step 2.3, detecting matrix A after pretreatment in step 2.2_sInputting a convolutional neural network model to be trained for parameter training and updating to obtain the convolutional neural network model for detecting GNSS deception interference;

step 2.4, obtaining a detection matrix A generated by detecting the GNSS signal to be detected by the convolutional neural network model for detecting the GNSS deception interference through the step 2.3_sAnd obtaining a detection result.

2. The GNSS deception jamming detection method based on convolutional neural network as claimed in claim 1Method, characterized by a threshold value λ in step 2.2_PThe values of (A) are as follows: lambda [ alpha ]_P＝2A_mean(ii) a In the formula, A_meanThe mean of the matrix a is searched for two dimensions.

3. The GNSS deception jamming detection method based on convolutional neural network as claimed in claim 1, further comprising:

step 1, traversing a two-dimensional search matrix A, and detecting the number N of correlation peaks larger than a capture threshold_peak(ii) a If N is present_peakMore than or equal to 2, considering that 2 or more correlation peaks exist in the signal, and obtaining a detection result of the existence of the deception jamming signal; if N is present_peak1, jumping to step 2.1; the two-dimensional search matrix A is a two-dimensional matrix which takes Doppler frequency shift and pseudo code phase as axes and is generated by the GNSS receiver in a signal acquisition stage.

4. The GNSS deception jamming detection method based on convolutional neural network as claimed in claim 3, wherein A is used in step 1_peak/A_mean＞λ_acqFor the capture conditions, A_peakFor two-dimensional search of the peak value of the matrix A, A_meanFor two-dimensional search of the mean value, λ, of the matrix A_acqIn order to capture the threshold, the method specifically comprises the following steps:

step 1.1, calculating the mean value A of the two-dimensional search matrix A_meanThe calculation formula is

N_AFor two-dimensional search of the number of elements in the matrix A, A_ijSearching the ith row and the jth column of the element in the matrix A in two dimensions;

step 1.2, traversing the two-dimensional search matrix A to obtain a correlation peak A_peak1If A is_peak1/A_mean＞λ_acqThen there is a correlation peak number N_peak1, jumping to step 1.3; if N is present_peakWhen the received signal is not the signal of the current pseudo-random code, the detection is finished;

step 1.3, data processing is carried out on the two-dimensional search matrix AProcessing: a is to be_peak1Data in a +/-1 kHz chip area on a Doppler frequency shift axis and a +/-1 chip area on a pseudo code phase axis are set to be zero; searching a two-dimensional matrix A 'obtained after data zero setting to obtain a second correlation peak A'_peak(ii) a If A'_peak/A_mean＞λ_acqThen N is_peakConsidering that 2 correlation peaks exist in the received signal, and obtaining a detection result of the existence of the deception jamming signal; otherwise N_peakStep 2.1 is skipped to 1.

5. The GNSS deception interference detection method based on the convolutional neural network as claimed in claim 1, wherein in the convolutional neural network model detection training of step 2.3, the input layer converts the input data X into convolutional layer C through convolution operation; the convolution operation process is as follows:

in the formula, W_KAs a convolution kernel, b_KFor the bias parameter, f (.) is the activation function, the convolution kernel W_KParameter (b) and bias parameter (b)_KIs a trainable parameter; the superscript k represents the bias parameters of the kth group of data; the subscript K represents W_KAnd b_KTraining parameters for convolution operation;

the convolution layer C is converted into a characteristic P through pooling operation;

the full connection layer integrates the characteristics output by the convolution layer and the pooling layer, and pulls the characteristics P into a column vector Fv; fv obtains an output result O after being calculated by a softmax function, wherein the O represents the probability that the result is each classification label; the calculation process is as follows:

O＝softmax(O_o),O_o＝f(W_o ^TFv+b_o)

wherein O is_oFor network output, f (.) is an output layer activation function; the calculation process of the softmax function is as follows:

r_χdenotes the x variable; w_oAs parameters of the matrix between the fully connected layer and the output layer, b_oIs a bias parameter; w_oAnd b_oIs a trainable parameter; the subscript o denotes W_oAnd b_oFor calculating O_oThe training parameters of (1).

6. The GNSS deception interference detection method based on the convolutional neural network as claimed in claim 5, wherein the average pooling operation is adopted in the process that the convolutional layer C is changed into the feature P through the pooling operation;

the average pooling operation process is as follows:

wherein S is a pooling window of size S₀@S₁×S₂，S₀、S₁、S₂Number, length and width of pooling windows, respectively, @ for separating the number, length x width; the subscripts τ and ν denote the τ -th row of P and the ν -th column of elements.

7. The GNSS deception interference detection method based on convolutional neural network as claimed in any of claims 1 to 6, wherein Δ T is_CLess than or equal to 1 chip.

Images