CN109188470A - A kind of GNSS cheating interference detection method based on convolutional neural networks - Google Patents

Abstract

The GNSS cheating interference detection method based on convolutional neural networks that the invention discloses a kind of, comprising the following steps: 1) be greater than the relevant peaks quantity N of detection threshold in signal acquisition phase, the two-dimensional matrix A generated when detecting signal capture_peakIf N_peak>=2 are then thought there are curve, if N_peak< 2 then continue step 2)；2) data on two-dimensional matrix A relevant peaks pseudo-code phase axis in ± 2 chip areas are intercepted and obtain detection matrix A_s, by transferring to convolutional neural networks to carry out detection training and classification after data prediction, finally obtain testing result.Detection method detection effect of the invention is preferable, with strong applicability, detects opportunity in signal acquisition phase, moderate complexity；The problem of detecting is difficult to when being able to solve curve and actual signal pseudo-code phase difference Δ T within 2 chips.

Description

A kind of GNSS cheating interference detection method based on convolutional neural networks

Technical field

The invention belongs to the interference detection technique fields in satellite navigation system, in particular to a kind of to be based on convolutional Neural net The GNSS cheating interference detection method of network (Convolutional Neural Network, CNN).

Background technique

Global Satellite Navigation System (Global Navigation Satellite System, GNSS) is a kind of covering Extensively, round-the-clock, real-time and high-precision navigation system.With the continuous development of Satellite Navigation Technique, GNSS is widely used in All kinds of dual-use facilities.With the development of Technique of Satellite Navigation and Positioning, number of users and application scenarios are continuously increased, peace Quan Xing, reliability are also increasingly valued by people.The security threat that present satellites navigation system is faced can mainly divide To be not intended to interference and intentionally interference.Interference refers mainly to artificial malicious interference intentionally, and can be divided into pressing type interference and deception formula Interference and combined interference.Pressing type interference, which refers to, applies high power interference signal to satellite band, causes receiver that can not receive Satellite-signal.And Deceiving interference refers to transmitting and navigation satellite is same or similar however signal that power is more stronger, satellite is led The reception terminal of boat system user may be mistakenly considered this signal to be sent by true navigation satellite, and is captured to it With tracking, leads to the information for receiving terminal generation mistake or exported without information.It is interfered relative to traditional high power pressing type, It is larger that Deceiving interference has advantages, the detection difficulties such as concealment strong, device miniaturization, jamming effectiveness height.

The detection method of existing Deceiving interference is broadly divided into two aspects: first is that based on multiple antennas, passing through the sky of signal Characteristic of field is detected, that is, detects the arrival bearing of the signal of multiple satellites to judge, if there is multiple PRN (Pseudo Random Noise code) signal come from same direction, then it is assumed that there are curves in satellite-signal.Second is that being based on signal Processing, is detected by features such as the time domain of signal, frequency domain, power, including signal absolute power detects, carrier-to-noise ratio detects, Signal quality monitoring, detection based on RAIM (Receiver Autonomous Integrity Monitoring) etc., this is several Kind of method it is with strong applicability, but the accuracy rate of the detection of signal absolute power and carrier-to-noise ratio detection is poor, the inspection based on RAIM Survey needs to resolve signal and complexity is higher, timeliness is not strong.Signal quality monitoring is main to be by signal relevant peaks It is no to be distorted to determine whether the method for existing signal quality monitoring mainly has the detection in signal capture there are curve Relevant peaks quantity and the signal trace time-division analysis correlator output the methods of；Detection relevant peaks quantity when current signal capture Method in curve with actual signal pseudo-code phase difference less than 2 chip when be difficult to differentiate.

Summary of the invention

The GNSS cheating interference detection method based on convolutional neural networks that the purpose of the present invention is to provide a kind of, to solve Above-mentioned technical problem.GNSS cheating interference detection method of the invention is based on convolutional neural networks, by GNSS receiver The two-dimensional matrix generated in signal capture intercepts relevant peaks proximity data and CNN algorithm is transferred to do emphasis inspection as detection data Survey, can more efficient utilize data information, in curve with actual signal pseudo-code phase difference less than 2 chip when can also differentiate, The presence or absence of curve can be detected with higher accuracy rate.

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

A kind of GNSS cheating interference detection method based on convolutional neural networks, comprising the following steps:

Step 2.1, relevant peaks A is intercepted in two-dimensional search matrix A_peak1Data in peripheral region obtain detection matrix A_s；Matrix A be GNSS receiver signal acquisition phase generate using Doppler frequency shift and pseudo-code phase as the two-dimensional search of axis Matrix；The range of peripheral region is ± 2kHz on Doppler frequency shift axis, ± 2 chips on pseudo-code phase axis；Detect matrix A_sIt is big Small is m_s×n_s, wherein m_s=4/ Δ f_D+ 1, n_s=4/ Δ T_C+ 1, m_sIndicate the length of matrix on Doppler frequency shift axis after intercepting, Δf_DFor Doppler frequency shift step-size in search, n_sIndicate the length of matrix on pseudo-code phase axis after intercepting, Δ T_CFor pseudo-code phase search Step-length；

Step 2.2, detection matrix A step 2.1 obtained_sIt is pre-processed；Pretreatment includes that will test matrix A_sIn it is low In threshold value λ_PValue zero setting；

Step 2.3, by the pretreated detection matrix A of step 2.2_sConvolutional neural networks model to be trained is inputted to carry out Parameter training updates, and obtains the convolutional neural networks model for detecting GNSS cheating interference；

Step 2.4, it is obtained by step 2.3 to be checked for detecting the convolutional neural networks model inspection of GNSS cheating interference Survey the detection matrix A that GNSS signal generates_s, obtain testing result.

Further, threshold value λ in step 2.2_PValue are as follows: λ_P=2A_mean；In formula, A_meanFor the equal of two-dimensional matrix A Value.

Further, further includes:

Step 1, searching matrix A, detection are greater than the relevant peaks quantity N of detection threshold_peak；If N_peak>=2, it is believed that signal In have 2 or more relevant peaks, obtain that there are the testing results of cheating interference signal；If N_peak=1, go to step 2.1； Two-dimensional matrix A be GNSS receiver signal acquisition phase generate using Doppler frequency shift and pseudo-code phase as the Two-Dimensional Moment of axis Battle array.

Further, with A in step 1_peak/A_mean>λ_acqFor contact conditions, A_peakFor two-dimensional matrix peak value, A_meanIt is two Tie up matrix mean value, λ_acqFor detection threshold, specifically includes the following steps:

Step 1.1, the mean value A of two-dimensional matrix A is calculated_mean, calculation formula isN_AFor of element in A Number, A_ijFor the element in two-dimensional matrix；

Step 1.2, two-dimensional matrix A is searched for, maximum value A is obtained_peak1If A_peak1/A_mean>λ_acq, then relevant peaks quantity N_peak =1, go to step 1.3；If N_peak=0, it is believed that receive the signal that current PRN is not present in signal, terminate detection；

Step 1.3, data processing is carried out to two-dimensional matrix A: by A_peak1± 1kHz, puppet on the Doppler frequency shift axis of position Data zero setting on code phase axis in ± 1 chip area；Two-dimensional matrix A after searching for data processing obtains maximum value A_peak2；If A_peak2/A_mean>λ_acq, then N_peak=2, it is believed that receiving has 2 relevant peaks in signal, there are the detections of cheating interference signal for acquisition As a result；Otherwise N_peak=1, go to step 2.1.

Further, in the convolutional neural networks model inspection training of step 2.3, input layer is by input data X through pulleying Product operation becomes convolutional layer C；Convolution algorithm process are as follows:

In formula, W_KFor convolution kernel, b_KFor offset parameter, f () is activation primitive, convolution kernel W_KIn parameter and biasing ginseng Number b_KFor can training parameter；

Convolutional layer C becomes P by pond operation；

Full articulamentum integrates the feature in the characteristics of image figure by convolutional layer and pond layer, by characteristics of image P Pull into a column vector Fv；Fv obtains output result O after the calculating of softmax function, and O indicates that result is each tag along sort Probability；Calculating process are as follows:

O=softmax (O_o),O_o=f (W_o ^TFv+b_o)

Wherein O_oFor network output, f () is output layer activation primitive；The calculating process of softmax function are as follows:W_oFor the matrix parameter between full articulamentum and output layer, b_oFor offset parameter；W_oAnd b_oIt is that can instruct Practice parameter.

Further, convolutional layer C uses average pond operation during becoming P by pond operation；

Average pond calculating process are as follows:

Wherein, S is pond window, size S₀@S₁×S₂, S₀、S₁、S₂The respectively quantity, length and width of pond window.

Further, Δ T_CLess than or equal to 1 chip.

Compared with prior art, the invention has the following advantages:

GNSS cheating interference detection method based on convolutional neural networks of the invention, convolutional neural networks are applied to GNSS cheating interference detection, be it is a kind of based on signal capture when detect relevant peaks quantity cheating interference signal detecting method, phase Than in existing detection method, the present invention transfers to CNN algorithm to do secondary emphasis by intercepting relevant peaks proximity data to detect, can be compared with Data information is efficiently utilized, it also can be compared with fine-resolution, energy when curve and actual signal pseudo-code phase difference are less than 2 chip Enough presence or absence with higher accuracy rate detection curve；By adjusting pseudo-code phase step-size in search, detection essence is adjusted Degree and accuracy, step value is smaller, and detection effect is better；The parameter of CNN model is using training data after training, Ke Yizhi It connects and carries out classified calculating with model, it is very fast using CNN algorithm speed；Detection accuracy of the invention is examined higher than direct absolute power Survey and export using signal trace correlator the method classified using MLP.

Detailed description of the invention

Fig. 1 is existing satellite navigation system deceiving jamming model schematic；

Fig. 2 is existing navigational satellite receiver signal capture two-dimensional search schematic diagram；

Fig. 3 is signal capture schematic block diagram of the existing navigational satellite receiver based on fft algorithm；

Fig. 4 is that there is only the schematic diagrames of the capture result of satellite navigation signals in existing reception signal；

Fig. 5 is the signal that the capture result of satellite navigation signals and curve is existed simultaneously in existing reception signal Figure；

Fig. 6 is a kind of process schematic block of GNSS cheating interference detection method based on convolutional neural networks of the invention Figure；

Fig. 7 is a kind of GNSS cheating interference detection method based on convolutional neural networks of the invention in Δ T_C=0.5 yard Detection process schematic diagram when piece,；

Fig. 8 is that a kind of GNSS cheating interference detection method based on convolutional neural networks of the invention uses different Δ T_C's Testing result contrast schematic diagram；

Fig. 9 is a kind of GNSS cheating interference detection method based on convolutional neural networks of the invention using different models Detection effect contrast schematic diagram；

Figure 10 is the testing result contrast schematic diagram that image classification uses CNN algorithm and KNN algorithm of the invention；

Figure 11 is detection method and the curve method of inspection comparison diagram based on signal processing of the invention.

Specific embodiment

Invention is further described in detail in the following with reference to the drawings and specific embodiments.

Please refer to Fig. 1 to Fig. 7, a kind of GNSS cheating interference detection method based on convolutional neural networks of the invention, base In system model be satellite navigation network: satellite navigation signals exist always, and cheating interference signal there may be that is, deposit by system In two kinds of situations；H₀: there is only satellite navigation actual signals in GNSS receiver reception signal；H₁: it receives and is existed simultaneously in signal Satellite navigation actual signal and curve, curve simulate the parameters such as pseudo-code phase, the Doppler frequency shift of actual signal, function Rate is slightly above actual signal, it is enable to have bigger probability captured in GNSS receiver capture.

After GNSS receiver receives satellite-signal, intermediate-freuqncy signal is converted by down coversion by signal:

Wherein, s_LSIt (t) is satellite navigation direct signal, s_MPIt (t) is multipath signal, n (t) is noise signal, s_Spoof(t) For curve.

Wherein, P is signal power, and C (t) is pseudo-code, that is, CA code, and D (t) is navigation data, f_IFFor theoretical intermediate frequency, f_Doppler For Doppler frequency shift, Φ is initial pseudo-code phase,For original carrier phase, footmark m_kIndicate that kth road multipath signal, M indicate altogether There is the road M multipath signal.

GNSS receiver captures intermediate-freuqncy signal, and the method for signal capture has method based on time domain correlator, base Method in matched filter and the method based on FFT etc., can generate two-dimensional matrix, for searching for relevant peaks and rough estimate The Doppler frequency shift and pseudo-code phase of satellite navigation signals.Successively all PRN are scanned in signal acquisition phase receiver, It generates using Doppler frequency shift and pseudo-code phase as the two-dimensional array of axis, i.e. matrix A, size is m × n, wherein m indicates Doppler The data length of frequency displacement axis, m=(f_Dmax-f_Dmin)/Δf_D+ 1, f_{Doppler_range}For the search range of Doppler frequency shift, f_DmaxWith f_DminRespectively f_{Doppler_range}Bound, Δ f_DFor Doppler frequency shift step-size in search, n indicates the length of pseudo-code phase axis,T_{CAcode_range}For the search range of pseudo-code phase,For T_{CAcode_range}Length, Δ T_CFor pseudo-code phase Step-size in search.As shown in Fig. 2, when signal is GPS signal, T_{CAcode_range}For [1,1023], Doppler frequency shift search range is f_{Doppler_range}=[- 7kHz, 7kHz].

The catching method based on FFT is used in present invention emulation, signal capture process is as shown in Figure 3:

(1) setting local carrier frequency is f_w=f_IF+f_Dmin+(i-1)Δf_D, i=1.f_IFFor satellite-signal down coversion it Theoretical IF frequency afterwards.

(2) by received intermediate-freuqncy signal S_R(t) it is multiplied, passes through with orthogonal signalling with the same phase of local carrier generator output Low-pass filter obtains complex signal s (t)=I+jQ of base band, whereinI, Q is respectively to be with phase and orthogonal road signal, T Capture the period.

1) FFT is done to s (t) and obtains S (f).

2) local pseudo-code generator generates pseudo-code signal c (t) according to current PRN, is FFT and conjugation is taken to obtain C^*(f)。

3) by S (f) and C^*(f) it is multiplied and is IFFT and obtain one-dimension arraySize For n × 1.

If 4) i < m, i=i+1, f_w=f_IF+f_Dmin+(i-1)Δf_D, continue step 2)；If i=m, by's Collection is combined into two-dimensional search matrix A,Size is m × n.

When receiving the satellite navigation signals that current PRN is not present in signal, without the relevant peaks for meeting contact conditions in A； When receiving in signal, there is only only have 1 phase for being greater than detection threshold when the satellite navigation signals of current PRN, in two-dimensional matrix A Guan Feng, as shown in Figure 4.When existing simultaneously the satellite navigation signals and curve of current PRN in reception signal, signal capture The two-dimensional matrix of generation will have 2 or more the relevant peaks for being greater than detection threshold, as shown in figure 5, and in curve and really When the offset of signal is smaller, these relevant peaks may be completely overlapped or partly overlap, this is just that the detection of cheating interference is brought Difficulty.

Detection method proposed by the invention is mainly for curve and actual signal pseudo-code phase difference Δ T in 2 chips Within when be difficult to the problem of detecting, by the data on interception maximum correlation peak pseudo-code phase axis in ± 2 chip areas, with detection It whether there is the case where related overlap of peaks, i.e., the relevant peaks generated with the presence or absence of cheating interference signal.A kind of base of the invention In the GNSS cheating interference detection method of convolutional neural networks, specifically includes the following steps:

Step 1, GNSS receiver can be generated in signal acquisition phase using Doppler frequency shift and pseudo-code phase as the two dimension of axis Matrix A.Relevant peaks quantity N of the search two-dimensional matrix A detection greater than detection threshold first_peakIf N_peak>=2, it is believed that in signal There are 2 or more relevant peaks, there are cheating interference signals；If N_peak=1, cheating interference signal is not detected, jumps to step Rapid 2.

Above-mentioned steps 1 specifically include:

Relevant peaks quantity N of the detection greater than detection threshold first_peak

With A_peak/A_mean>λ_acqFor contact conditions, A_peakFor two-dimensional matrix peak value, A_meanFor two-dimensional matrix mean value, λ_acqFor Detection threshold.Present invention emulation uses method and step are as follows:

1.1) mean value of A is found outN_AFor the number of element in A.

1.2) A is searched for, maximum value A is obtained_peak1If A_peak1/A_mean>λ_acq, then relevant peaks quantity N_peak=1, continue to walk It is rapid 1.3), otherwise N_peak=0, it is believed that the signal of current PRN is not present in signal.

1.3) by A_peak1± 1kHz on the Doppler frequency shift axis of position, the number on pseudo-code phase axis in ± 1 chip area According to zero setting.A is searched for, maximum value A is obtained_peak2If A_peak2/A_mean>λ_acq, then N_peak=2, it is believed that have 2 correlations in signal Peak, there are cheating interference signal, otherwise N_peak=1, cheating interference signal is not detected, continues step 2).

Step 2, the data on two-dimensional matrix A relevant peaks pseudo-code phase axis in ± 2 chip areas are intercepted and obtain detection matrix A_s, by transferring to convolutional neural networks (CNN) to carry out detection training and classification after data prediction；Detection training to data are used It is complete or reach the default condition of convergence.

The implementation method of step 2 is as follows:

Step 2.1, relevant peaks A is intercepted on two-dimensional matrix A_peak1± 2kHz on Doppler frequency shift axis around position, Data on pseudo-code phase axis in ± 2 chip areas obtain new detection matrix A_s, size m_s×n_s, wherein m_s=4/ Δ f_D + 1, n_s=4/ Δ T_C+1。

Step 2.2, data are pre-processed, due in matrix A_sIn data only near relevant peaks it is related to detection, So by A_sIn be lower than threshold value λ_PValue zero setting.

Step 2.3, it will test matrix A_sCNN is transferred to carry out detection training and verifying.The structure of CNN model is according to A_sIt is big It is small to preset.The CNN model parameter for setting structure is trained using the training data of known scene classification first, is instructed Practicing cut-off condition is that the number of iterations is completed or error no longer changes or error reaches requirement, after training finishes, input test number It is tested according to the detection effect to CNN model.

Step 2 specifically includes the following steps:

2) matrix A is intercepted, transfers to CNN to carry out detection training and classification as image.

2.1) A is intercepted in matrix A_peak1± 2kHz on Doppler frequency shift axis around position, ± 2 on pseudo-code phase axis Data in chip area obtain new matrix A_s, size m_s×n_s, wherein m_s=4/ Δ f_D+ 1, n_s=4/ Δ T_C+1。 Generally by Δ f in GNSS receiver_DIt is set as 0.5kHz, Δ T_CIt is set as 0.5 chip.

2.2) data are pre-processed, due in matrix A_sIn data only near relevant peaks it is related to detection, so By A_sIn be lower than λ_PValue zero setting；Wherein λ in this emulation_P=2A_mean。

2.3) it will test matrix A_sCNN is transferred to carry out detection training and verifying.The structure of CNN model is according to A_sSize it is pre- First set.The structure of CNN model needs to think the quantity and arrangement and convolution kernel of predominantly convolutional layer and the pond layer of setting With the size of pond window, setting principle is general are as follows: 1, make the size for exporting image equal to length and width by current layer operation Regular image it is close；2, the size of output layer is made decimal do not occur by every layer of operation, wherein the operation of convolutional layer For subtraction, M_C=M_I-K₁+ 1, N_C=N_I-K₂+ 1, the operation of pond layer is division, M_P=M_I/S₁, N_P=N_I/S₂, M, N are respectively indicated The length and width of current layer, subscript I indicate that input layer, C indicate that convolutional layer, P indicate pond layer, K₁、K₂Respectively the length of convolution kernel and It is wide.S₁、S₂The respectively length and width of pond window.For example, working as A_sWhen size is 9 × 9, the structure of CNN model such as Fig. 6 institute Show, treatment process is as follows:

2.3.1) input layer X, size F_x@M_x×N_x=1@9 × 9, i.e., the image that each input sample is 19 × 9.

2.3.2) pass through convolution algorithm, become convolutional layer C, size F_C@M_C×N_C=4@4 × 4, wherein F_C=K₀, M_C= M_I-K₁+ 1, N_C=N_I-K₂+1.Process are as follows:

C=f ((∑ W_K*X)+b_K)

Wherein, W_KFor convolution kernel, size K₀@K₁×K₂=4@6 × 6, K₀、K₁、K₂Respectively the quantity, length of convolution kernel and It is wide.b_KFor offset parameter, size K₀× 1, f () are activation primitive, there is sigmoid, tanh, Relu etc..Convolution kernel W_KIn Parameter and offset parameter b_KBeing can training parameter.

2.3.3) there are the methods of maximum pond, average pond in pond, this uses the method in average pond.By Chi Huayun It calculates, becomes pond layer P, size F_P@M_P×N_P=4@2 × 2, wherein F_C=S₀, M_P=M_C/S₁, N_P=N_C/S₂.Process are as follows:

Wherein, S is pond window, size S₀@S₁×S₂=4@2 × 2, S₀、S₁、S₂Respectively the quantity of pond window, It is long and wide.

2.3.4) full articulamentum integrates the feature in the characteristics of image figure by convolutional layer and pond layer, by image Feature pulls into a column vector Fv, size F_Fv@M_Fv×N_Fv=1@1 × 16, wherein N_Fv=F_PM_PN_P。

2.3.5) Fv obtains output result O, size F after the calculating of softmax function_O@M_O×N_O=1@1 × 2, In, N_OEqual to tag along sort number, O indicates that result is the probability of each tag along sort.Process are as follows:

O=softmax (O_o),O_o=f (W_o ^TFv+b_o)

Wherein O_oFor network output, f () is output layer activation primitive.The calculating process of softmax function are as follows:Purpose is to calculate the probability that the label of x is j, and the sum of probability is made to be 1.W_oFor full articulamentum with Matrix parameter between output layer, b_oFor offset parameter.W_oAnd b_oBeing can training parameter.

2.3.6) CNN uses BP (Error Back Propagation) back-propagation algorithm undated parameter.With MSE (Mean Square Error) is transmission error E, backpropagation.Using square error cost function:

Wherein Y is the actual classification label of sample data.

The main formulas that parameter updates are as follows:Wherein W is updated connection ginseng Number, W_iFor initial Connecting quantity, η is learning efficiency.To W_K、b_K、W_oAnd b_oEtc. asking local derviation, and undated parameter.

Full articulamentum is returned from output layer, error propagation is carried out and parameter updates:

O_d=E ⊙ f'(O_o),

Wherein O_dIndicate the residual error that output layer returns, size F_o@M_o×N_o=1@2 × 1, f'(.) indicate leading for f () Number, ⊙ indicate alma inner product, the i.e. multiplication of vectors of corresponding position, Fv_dIndicate the residual error that full articulamentum returns, size F_Fv@ M_Fv×N_Fv=1@16 × 1.

Pond layer is returned from full articulamentum, error propagation is carried out, by Fv_dVariation is P_d, and error is inverted, size is upper layer The size of pond layer, F_P@M_P×N_P=4@2 × 2

Convolutional layer is returned from pond layer, error propagation is carried out and parameter updates:

C_d=f'(C) ⊙ up (P_d),

C_dIndicate the residual error that convolutional layer returns, size F_C@M_C×N_C=4@4 × 4,Indicate up-sampling.

One group of data of every input, the process that propagated forward, costing bio disturbance, backpropagation, parameter update are an iteration (epoch)。

After the parameter of CNN model is updated by training parameter, CNN model inspection point is given using test data as input Class verifies detection effect, obtains the CNN model for detecting GNSS cheating interference.

Convolutional neural networks (CNN) are one kind of artificial neural network, are widely used in image recognition in recent years, voice is known Not Deng fields, have the characteristics that structure is simple, training parameter is few, adaptable.The thought of deep learning has been introduced into mind by it In network, by convolution algorithm come the feature of the different levels of extraction image from the superficial to the deep, the instruction of neural network is utilized Practice the parameter that process allows whole network to automatically adjust convolution kernel.Convolutional neural networks include one by convolutional layer and sub-sampling layer structure At feature extractor, using local connection, weight is shared and the method for sub-sampling simplifies model complexity, reduces ginseng Number quantity, reduces the risk of over-fitting.Present invention employs CNN algorithms to realize that cheating interference detects.Currently, being based on more days The method of line emits the feelings of multiple PRN (Pseudo Random Noise code) signals mainly for single cheating interference source Condition, there are limitations for application, and will increase the cost of commercial satellite navigation neceiver using multiple antennas；The signal trace time-division The method of analysis correlator output is equivalent to artificial extraction signal characteristic, can not make full use of to signal message, receives function in signal Detection accuracy is lower when rate is unstable, noise is smaller.The method of relevant peaks quantity is detected when signal capture in curve It is difficult to differentiate when with actual signal pseudo-code phase difference less than 2 chip.Method of the invention utilizes two-dimensional matrix A, to reception signal In detected with the presence or absence of curve, predominantly detect according to being that identical PRN identifies whether that there are two or more signals, i.e., Two-dimensional matrix A whether there is two or more relevant peaks.Relevant peaks quantity N of the detection greater than detection threshold λ first_peakIf N_peakThen think for >=2 to obtain testing result there are curve；If N_peak< 2 continue to execute subsequent step；Pass through interception Two-dimensional matrix A transfers to CNN to carry out detection training and classification as image, and then solves curve and actual signal pseudo-code phase The problem of detecting is difficult to when poor (later abbreviation Δ T) is within 2 chips.Detection method provided by the invention is a kind of based on signal The method of quality-monitoring, in signal capture, the two-dimensional array that GNSS receiver is generated intercepts relevant peaks as detection data Proximity data transfers to CNN algorithm to do emphasis detection.Here, we regard the two-dimensional array as equivalent image, using CNN in image It identifies the advantage of aspect, improves the efficiency of cheating interference signal detection.Detection method strong applicability of the invention, opportunity be forward, Moderate complexity；And software receiver generally using for this detection method application provide possibility.

In order to verify the performance of the GNSS cheating interference detection method proposed by the present invention based on convolutional neural networks, we Carry out following emulation experiment:

The intermediate-freuqncy signal of simulation GNSS receiver be sample frequency be 16.3676MHz, theoretical intermediate frequency is 4.1304MHz's GPS satellite navigation signal, is indifferent to text, random to generate telegraph text data D (t).Direct projection is had in the satellite navigation signals of simulation Signal and multipath signal, multipath signal, decline are greater than 3dB all the way for simulation, and receiver received signal to noise ratio SNR is -20~-15dB. There is only the difference of Doppler frequency shift, pseudo-code phase and power, Doppler frequency shift difference Δ f for the curve and actual signal of simulation Change at random in the section ± 1kHz, pseudo-code phase difference Δ T variation in ± 2 chip intervals, power be greater than direct signal 1.1~ 1.5dB.It emulates data and amounts to 415800 groups of data, be divided into H₀: there is only true satellites to lead in GNSS receiver reception signal Navigate signal；H₁: satellite navigation actual signal is existed simultaneously in GNSS receiver reception signal and two kinds of scenes of curve are each 207900 groups.Wherein H₁The data of scene press curve again and the pseudo-code phase difference Δ T of actual signal is distinguished, Δ T value area Between be 0 to 2 chip, 0.1 chip of step-length, amount to 21 classes, 9900 groups of data of every class.Take the 1/ of two kinds of scenes and corresponding Various types of data 2, which amount to 207900 groups, is used as training data；In remaining 1/2 data, H₁The data of scene are 21 classes, every class further according to Δ T points 4950 groups of data, H₀The data of scene are divided into 21 parts, and 4950 groups of data of every class merge to obtain 21 class testings with above-mentioned 21 class data Data, 9900 groups of data of every class.

Receiver uses the catching method based on FFT, f_{Doppler_range}It takes [- 7kHz, 7kHz], Δ f_D0.5kHz is taken, T_{CAcode_range}=1023.In detection process, Δ T_CFor pseudo-code phase step-size in search, 1 chip is taken respectively, 0.5 chip, 0.25 yard Piece and 0.1 chip compare.

We choose different pseudo-code phase step-size in search Δ T to method of the invention_CWhen detection effect be compared, Detect matrix A_sSize and emulation employed in CNN model structure it is as shown in table 1:

Interception sample frequency and the CNN model structure table of comparisons when table 1 emulates

As shown in fig. 7, simulating the method for the present invention respectively in emulation takes different pseudo-code phase step-size in search Δ T_CWhen, it is right The detection case of the curve of different pseudo-code phase differences.Curve poor compared with the pseudo-code phase of actual signal is 0 chip~2 yard Piece, Doppler frequency shift difference random value in the region ± 1kHz.As can be seen that when the phase between curve and actual signal Detection effect is poor when deviating smaller, and the difference of the increase, curve and the actual signal that deviate with pseudo-code phase is increasingly It is big and detection effect is become better and better.Generally, step delta T_CIt is worth smaller, detection effect is better.ΔT_CWhen for 1 chip, in pseudo-code phase Position offset deviation detection probability in 0.5 chip and 1.5 chips or so is lower.This is because the accuracy of identification of 1 chip can not be effective Identify the offset of 0.5 chip, distance N Δ T_C, N ∈ R is remoter, and detection effect is poorer.As shown in Table 1 and Table 2, Δ T_CBe worth it is smaller, Detection algorithm is more complicated, and the time is longer.In practical application, it can be weighed according to the requirement to detection effect and rate and be selected.

Table 2 emulates single Data Detection velocity contrast's table

In fig. 8, we take Δ T_C=0.5 chip, detects matrix A at this time_sSize is [9,9], compared CNN model and adopts Detection effect when different structure is taken, is respectively as follows:

Structure 1:{ input, C1~2@6 × 6, the@of S2~2 2 × 2, fullconnection, output }；

Structure 2:{ input, C1~2@4 × 4, the@of S2~2 2 × 2, fullconnection, output }；

Structure 3:{ input, C1~2@2 × 2, the@of S2~2 2 × 2, fullconnection, output }；

Structure 4:{ input, C1~4@6 × 6, the@of S2~4 2 × 2, fullconnection, output }.

From figure 8, it is seen that in the case where using different models, it is essentially identical to the detection probability of curve.In puppet It is slightly different when code phase difference Δ T is less than 1 chip, basic law be convolution kernel it is bigger, can training parameter it is more, detection effect It is better, but can training parameter when increasing, training duration and detection duration also will increase, and need to weigh detection effect and complexity.

The present invention is classified using signal acquisition phase two-dimensional matrix, and in Fig. 9, we be compared using different figures Influence as sorting algorithm to detection effect.KNN (k-NearestNeighbor) algorithm is also a kind of common image recognition calculation Method, it calculates images to be recognized and all training images, finds the k training closest with images to be recognized and schemes Picture, the label for taking corresponding label most are final result.From fig. 9, it can be seen that carrying out identification classification than using using CNN algorithm It is more accurate that KNN algorithm is classified, as Δ T_CWhen identical, the detection probability of CNN is above KNN.From table 2 it can be seen that adopting KNN algorithm speed is used faster with CNN algorithm ratio, this is because the parameter of CNN model is using training data after training, it can Directly to carry out classified calculating with model, and KNN be required to when carrying out each classified calculating by all training datas with wait know Other data are compared, so more quick and precisely using CNN algorithm.

In Figure 10, we compare several methods based on signal processing, be respectively the detection of direct absolute power, Classified (Detection of Spoofing Attack using using the output of signal trace correlator using MLP Machine Learning based on Multi-Layer Neural Network in Single-Frequency GPS Receivers, Shafiee E, Mosavi M R, Moazedi are M.2017) and it is proposed by the present invention utilize signal acquisition phase Two-dimensional matrix is classified using CNN.It is exported using signal trace correlator using the method main process that MLP classifies and is Same phase and orthogonal early, timely, late correlation I using the output of signal trace correlator_E,I_P,I_L,Q_E,Q_P,Q_L, by formula (1), (2), (3) are calculated Delta feature x₁, early stage-iate feature x₂, signal level feature x₃, transfer to MLP (Multi- Layer Perceptron) neural network is trained and adjudicates.

From fig. 10 it can be seen that the method for the present invention Detection accuracy is higher than two kinds of comparison schemes.Although method of the invention To be longer than two kinds of comparison schemes between when detecting, but the method classified is identified by MLP with using the output of signal trace correlator Compare, method of the invention can be carried out detecting after signal capture, and the method for utilizing the output of signal trace correlator Need signal trace generate correlator output after can just be detected, with the inventive method testing result when Between it is upper and keep abreast with the method exported using signal trace correlator.

The above content is a further detailed description of the present invention in conjunction with specific preferred embodiments, and it cannot be said that A specific embodiment of the invention is only limitted to this, for those of ordinary skill in the art to which the present invention belongs, is not taking off Under the premise of from present inventive concept, several simple deduction or replace can also be made, all shall be regarded as belonging to the present invention by institute Claims of submission determine scope of patent protection.

Claims (7)

1. a kind of GNSS cheating interference detection method based on convolutional neural networks, which comprises the following steps:

Step 2.1, relevant peaks A is intercepted in two-dimensional search matrix A_peak1Data in peripheral region obtain detection matrix A_s；Square Battle array A be GNSS receiver signal acquisition phase generate using Doppler frequency shift and pseudo-code phase as the two-dimensional search matrix of axis； The range of peripheral region is ± 2kHz on Doppler frequency shift axis, ± 2 chips on pseudo-code phase axis；Detect matrix A_sSize be m_s ×n_s, wherein m_s=4/ Δ f_D+ 1, n_s=4/ Δ T_C+ 1, m_sIndicate the length of matrix on Doppler frequency shift axis after intercepting, Δ f_DFor Doppler frequency shift step-size in search, n_sIndicate the length of matrix on pseudo-code phase axis after intercepting, Δ T_CFor pseudo-code phase step-size in search；

Step 2.2, detection matrix A step 2.1 obtained_sIt is pre-processed；Pretreatment includes that will test matrix A_sIn be lower than door Limit value λ_PValue zero setting；

Step 2.3, by the pretreated detection matrix A of step 2.2_sIt inputs convolutional neural networks model to be trained and carries out parameter Training updates, and obtains the convolutional neural networks model for detecting GNSS cheating interference；

Step 2.4, it is obtained by step 2.3 to be detected for detecting the convolutional neural networks model inspection of GNSS cheating interference The detection matrix A that GNSS signal generates_s, obtain testing result.

2. a kind of GNSS cheating interference detection method based on convolutional neural networks according to claim 1, feature exist In threshold value λ in step 2.2_PValue are as follows: λ_P=2A_mean；In formula, A_meanFor the mean value of two-dimensional matrix A.

3. a kind of GNSS cheating interference detection method based on convolutional neural networks according to claim 1, feature exist In, further includes:

Step 1, searching matrix A, detection are greater than the relevant peaks quantity N of detection threshold_peak；If N_peak>=2, it is believed that have 2 in signal A or above relevant peaks, there are the testing results of cheating interference signal for acquisition；If N_peak=1, go to step 2.1；Two dimension Matrix A be GNSS receiver signal acquisition phase generate using Doppler frequency shift and pseudo-code phase as the two-dimensional matrix of axis.

4. a kind of GNSS cheating interference detection method based on convolutional neural networks according to claim 3, feature exist In with A in step 1_peak/A_mean>λ_acqFor contact conditions, A_peakFor two-dimensional matrix peak value, A_meanFor two-dimensional matrix mean value, λ_acq For detection threshold, specifically includes the following steps:

Step 1.1, the mean value A of two-dimensional matrix A is calculated_mean, calculation formula isN_AFor the number of element in A, A_ij For the element in two-dimensional matrix；

Step 1.2, two-dimensional matrix A is searched for, maximum value A is obtained_peak1If A_peak1/A_mean>λ_acq, then relevant peaks quantity N_peak=1, Go to step 1.3；If N_peak=0, it is believed that receive the signal that current PRN is not present in signal, terminate detection；

Step 1.3, data processing is carried out to two-dimensional matrix A: by A_peak1± 1kHz, pseudo-code phase on the Doppler frequency shift axis of position Data zero setting on the axis of position in ± 1 chip area；Two-dimensional matrix A after searching for data processing obtains maximum value A_peak2；If A_peak2/A_mean>λ_acq, then N_peak=2, it is believed that receiving has 2 relevant peaks in signal, there are the detections of cheating interference signal for acquisition As a result；Otherwise N_peak=1, go to step 2.1.

5. a kind of GNSS cheating interference detection method based on convolutional neural networks according to claim 1, feature exist In in the convolutional neural networks model inspection training of step 2.3, input data X is become convolution by convolution algorithm by input layer Layer C；Convolution algorithm process are as follows:

In formula, W_KFor convolution kernel, b_KFor offset parameter, f () is activation primitive, convolution kernel W_KIn parameter and offset parameter b_K For can training parameter；

Convolutional layer C becomes P by pond operation；

Full articulamentum integrates the feature in the characteristics of image figure by convolutional layer and pond layer, and characteristics of image P is pulled into One column vector Fv；Fv obtains output result O after the calculating of softmax function, and O indicates that result is the general of each tag along sort Rate；Calculating process are as follows:

O=softmax (O_o),O_o=f (W_o ^TFv+b_o)

Wherein O_oFor network output, f () is output layer activation primitive；The calculating process of softmax function are as follows:W_oFor the matrix parameter between full articulamentum and output layer, b_oFor offset parameter；W_oAnd b_oIt is that can instruct Practice parameter.

6. a kind of GNSS cheating interference detection method based on convolutional neural networks according to claim 5, feature exist In convolutional layer C uses average pond operation during becoming P by pond operation；

Average pond calculating process are as follows:

Wherein, S is pond window, size S₀@S₁×S₂, S₀、S₁、S₂The respectively quantity, length and width of pond window.

7. a kind of GNSS cheating interference detection side based on convolutional neural networks according to any one of claim 1 to 6 Method, which is characterized in that Δ T_CLess than or equal to 1 chip.