A fast recognition method for SAR imaging patterns based on neural network

Abstract

The invention discloses a SAR imaging mode fast recognition method based on a neural network, which comprises the following steps: S1: constructing a sample data set of a SAR signal power graph; S2: creating and initializing a BP neural network model; S3: using the sample data set to pre-train the neural network model to obtain a SAR imaging mode recognition network model; S4: using the pre-trained network model to predict the power graph sample of the received signal to obtain its corresponding imaging mode. The recognition method is based on sidelobe reconnaissance to model the jammer intercepted signal, and the obtained signal model is more universal. The neural network is used to adaptively extract the power graph features of each mode, which can provide higher accuracy and faster recognition speed, and has certain robustness to Gaussian white noise, and has important application value in the field of radar electronic countermeasures.

Description

Technical Field

[0001] The present invention relates to the fields of electronic reconnaissance and deep learning, and more specifically, to a method for rapid recognition of SAR imaging patterns based on a neural network. Background Art

[0002] Synthetic aperture radar (SAR), with its all-day, all-weather, wide-range, and high-resolution observation capabilities, plays a crucial role in electronic warfare intelligence acquisition. Various jamming techniques targeting SAR have become a research hotspot in recent years. Among these, SAR deception jamming techniques can produce highly realistic jamming signals with characteristics consistent with true echoes in both range and azimuth. However, to achieve a realistic deception jamming effect, parameter reconnaissance of the target SAR is generally required to obtain information such as signal parameters and satellite motion parameters. Combined with the desired false image, the jamming signal is generated using echo simulation. Recognizing SAR signal imaging patterns is a crucial step in parameter reconnaissance, facilitating targeted deception jamming.

[0003] Traditional SAR imaging pattern recognition methods mainly perform manual comprehensive judgment by analyzing signal characteristics such as the received signal power diagram and Doppler history, which has disadvantages such as low recognition efficiency and poor anti-noise performance. Summary of the Invention

[0004] To address these issues, the present invention proposes a neural network-based rapid SAR imaging pattern recognition method. Compared to traditional SAR imaging pattern recognition techniques, this method provides a neural network model that can accurately identify the original imaging pattern of intercepted signals. This significantly improves recognition efficiency while ensuring robustness to low signal-to-noise ratios and insensitivity to jammer locations.

[0005] The present invention discloses a method for rapid recognition of SAR imaging patterns based on a neural network, comprising the following steps:

[0006] S1: Construct a sample dataset of SAR signal power maps;

[0007] S2: Create and initialize the BP neural network model;

[0008] S3: Use the sample data set to pre-train the neural network model to obtain the SAR imaging pattern recognition network model;

[0009] S4: Use the pre-trained network model to predict the power map sample of the received signal to obtain its corresponding imaging mode.

[0010] S1: Construct a sample dataset of SAR signal power maps;

[0011] Model the SAR signal intercepted by the ground jammer, and then simulate to obtain the power map samples of the SAR signal, and obtain a sufficient amount of sample data sets under different parameters;

[0012] S2: Create and initialize the BP neural network model;

[0013] Select the BP neural network model, create and initialize the model;

[0014] S3: Use the sample data set to pre-train the BP neural network model;

[0015] The generated sample data set is divided into a training set and a test set. The training set is input into the BP neural network model for pre-training, and the test set is used to calculate the recognition accuracy. According to the quality of the recognition results, the network parameters and hyperparameters of the BP neural network model are appropriately adjusted to obtain a better BP neural network model.

[0016] S4: Use the pre-trained BP neural network model to predict the power map samples of the SAR signal and obtain the corresponding imaging mode;

[0017] The ground jammer receives the SAR signal echo in a fixed imaging mode, processes the SAR signal echo to obtain the SAR signal power map sample, and inputs the SAR signal power map sample into the pre-trained BP neural network model to output the corresponding imaging mode category.

[0018] Compared with the existing technology, the invention of a neural network-based SAR imaging pattern rapid recognition method has the following beneficial effects:

[0019] The proposed SAR imaging pattern recognition method balances computational speed and accuracy, while also being robust to low signal-to-noise ratios. Furthermore, the proposed scheme can also improve the received signal-to-noise ratio by combining parameter reconnaissance with pulse compression, thereby enhancing system stability. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 It is the geometric relationship model of strip SAR;

[0021] Figure 2 It is a scanning SAR geometric relationship model;

[0022] Figure 3 It is the geometric relationship model of spotlight SAR;

[0023] Figure 4 (a) is a sample of the stripe pattern generated by a simulation experiment;

[0024] Figure 4 (b) is a sample of the scanning pattern generated by a simulation experiment;

[0025] Figure 4 (c) is a sample of the bunching pattern generated by a simulation experiment;

[0026] Figure 5 This is the network model structure diagram used in the experiment;

[0027] Figure 6 is the azimuth antenna pattern in the center of the SAR observation area; DETAILED DESCRIPTION

[0028] The following describes in detail a neural network-based SAR imaging pattern rapid recognition method proposed by the present invention in conjunction with the accompanying drawings.

[0029] Without loss of generality, here we consider the spaceborne SAR using a common microstrip planar array antenna, the imaging modes adopted are strip, scanning and beamforming, and the transmitted signal adopts a linear frequency modulation signal system.

[0030] The spaceborne SAR signal intercepted by the ground-based reconnaissance receiver is a periodic pulse train. In free space, the power density S of the SAR signal at the reconnaissance receiver is e Expressed as:

[0031]

[0032] Among them, P t is the transmit power of the spaceborne SAR; is the transmission gain of the spaceborne SAR antenna in the direction of the reconnaissance receiver, is the azimuth angle of the antenna pattern, is the range angle of the antenna pattern; R is the distance between the spaceborne SAR and the reconnaissance receiver; L t is the transmission loss of the spaceborne SAR, including the feeder loss between the SAR transmitter and the transmitting antenna;

[0033] The effective area A of the reconnaissance receiver antenna r for:

[0034]

[0035] Among them, G r In order to detect the receiving gain of the receiver antenna in the direction of the spaceborne SAR, G r Set to a fixed value; λ is the wavelength of the SAR signal;

[0036] In free space, ignoring the transmission loss of electromagnetic waves in the atmosphere, the signal power P at the location of the reconnaissance receiver is rf for:

[0037]

[0038] Furthermore, the signal power at the input of the reconnaissance receiver is:

[0039]

[0040] Among them, L r is the receiving loss of the reconnaissance receiver, including the feeder loss from the reconnaissance receiver antenna to the reconnaissance receiver input terminal and the polarization mismatch loss between the reconnaissance receiver antenna and the SAR signal;

[0041] Correspondingly, the signal amplitude PA at the output of the reconnaissance receiver is expressed as:

[0042]

[0043] Where Z is the load impedance of the reconnaissance receiver;

[0044] According to the plane array theory, assuming that the SAR antenna has uniform aperture illumination, when the SAR antenna size is much larger than the wavelength of the SAR transmitted signal, the range and azimuth planes of the antenna pattern are approximately expressed as sinc functions. The two-dimensional antenna pattern in amplitude form is expressed as:

[0045]

[0046] Among them, D A is the equivalent length of the antenna in azimuth, D R is the equivalent length of the antenna distance; represents the normalized sinc function;

[0047] Therefore, the transmission gain of the spaceborne SAR antenna in the direction of the reconnaissance receiver is expressed as:

[0048]

[0049] Among them, G t (0,0) is and The transmission gain at this time is the peak main lobe gain.

[0050] The above is a general model of the two-dimensional antenna pattern applicable to various imaging modes of spaceborne SAR. However, for different imaging modes, the azimuth angle of the antenna pattern and distance angle The way they change over time is different. Based on the assumptions of uniform linear motion and linear geometry of the spaceborne SAR platform, the azimuth and range angles in strip and spotlight modes are derived below.

[0051] The geometric relationship of spaceborne SAR in strip mode is as follows: Figure 1 As shown. Assume that the coordinates of the spaceborne SAR are R(0,Vr η,H), the ground point target coordinates are A(R G ,0,0), the coordinates of the ground reconnaissance receiver are P(x0,y0,0), and after derivation, for the geometric relationship of the spaceborne SAR in strip mode, the azimuth angle and range angle of the antenna pattern are expressed as:

[0052]

[0053]

[0054] Among them, V r is the flight speed of the SAR platform (including the satellite carrying SAR); η represents the azimuth slow time, η c is the beam center crossing time of the point target; H is the orbit height; R G The ground distance from the point target to the radar track. The radar track refers to the moving track of the vertical projection of the SAR on the earth's surface.

[0055] The geometric relationship of the spaceborne SAR in scanning mode is shown in Figure 2. Assume that the coordinates of the spaceborne SAR, the coordinates of the ground point target, and the coordinates of the ground reconnaissance receiver are the same as those in strip mode. Since the observation area in scanning mode consists of multiple sub-strip areas, the center of the beam coverage area will be scanned along the centerline of each sub-strip in turn. Therefore, the range and direction angles of the antenna pattern of the spaceborne SAR in scanning mode are expressed as:

[0056]

[0057]

[0058] Among them, R G (i),i=1,2,3,…,N s Indicates the ground distance of each sub-strip centerline, N s The number of sub-stripes.

[0059] The geometric relationship of spaceborne SAR in spotlight mode is as follows: Figure 3 As shown. Assume that the coordinates of the spaceborne SAR, the coordinates of the ground point target, and the coordinates of the ground reconnaissance receiver are the same as those in the strip mode. After derivation, for the spaceborne SAR in the beamforming mode, the azimuth angle and range angle of the antenna pattern are expressed as:

[0060]

[0061]

[0062] Through the derivation of the azimuth and range angles of the antenna patterns of the three imaging modes, it can be seen that the antenna scanning methods of different modes are different, which causes the power diagram of the signal intercepted by the jammer to have fixed characteristics corresponding to the SAR imaging mode. Therefore, subsequent experiments will describe the signal power diagram by extracting the pulse peak value, and use this as the main basis for judging the imaging mode.

[0063] Considering that spaceborne SAR generally transmits linear frequency modulation (LFM) signals, which are easy to intercept, the model is based on the LFM signal:

[0064] The spaceborne SAR signal intercepted by the reconnaissance receiver is expressed as:

[0065]

[0066] Where τ represents the distance to the fast time; T r is the pulse width of the LFM signal; f c is the carrier frequency of the transmitted signal; K r is the frequency modulation of the LFM signal; R(η) is the distance from the spaceborne SAR to the reconnaissance receiver, which is expressed as follows based on the straight line geometry assumption:

[0067]

[0068] The spaceborne SAR signal demodulated by the reconnaissance receiver is expressed as:

[0069]

[0070] Using multiple sets of simulation parameters, a large number of training samples are generated for the three imaging modes. The sample production process is as follows: Assuming that the jammer continuously intercepts 1000 pulses, the amplitude peaks of all pulses are extracted to obtain a sample in the form of a matrix of size 500×1. By setting different signal-to-noise ratios, signal parameters, and geometric parameters, a sufficient number of training set samples and test set samples are generated. The samples of strip, scan, and beam modes are as follows: Figure 4 During the simulation experiment, the parameter settings adopted are shown in Table 1.

[0071] Table 1 Main parameters involved in simulation data

[0072]

[0073] After the data set is prepared, it is input into the network model for training. In the experiment, a three-hidden-layer BP neural network model is selected. The number of features in the hidden layer is set to 100, 25, and 3 respectively; the learning rate is set to 0.00001, and the batch size is set to 32. The structure of the three-hidden-layer BP neural network model is as follows: Figure 5 shown.

[0074] First, Table 2 shows the impact of the jammer's received signal-to-noise ratio (SNR) on the recognition accuracy.

[0075] Table 2 Experimental results of signal-to-noise ratio analysis of the initial solution without matched filtering

[0076]

[0077] Experimental results show that when the SNR is above 0dB, recognition accuracy can reach 85%. However, as the SNR decreases, recognition performance degrades significantly. To address this issue, this method proposes an improved scheme based on matched filtering: First, the received signal's frequency modulation rate is estimated using a fractional Fourier transform (FrFT). Then, a matched filter is applied to the received signal, effectively improving the received signal-to-noise ratio (SNR). Simulations of this improved scheme using matched filtering analyze the impact of SNR on recognition accuracy. The results are shown in Table 3.

[0078] Table 3 Experimental results of signal-to-noise ratio analysis using the improved scheme of matched filtering

[0079]

[0080] Experimental results show that the use of matched filtering can greatly improve the robustness of the proposed method to low signal-to-noise ratios, making the recognition accuracy stable at around 90% and not deteriorating rapidly with the decrease of the signal-to-noise ratio.

[0081] In addition, considering that the relative position of the jammer to the center of the SAR observation area is unpredictable and uncontrollable in engineering when the jammer is reconnaissance on the SAR satellite, the following experiment verifies the robustness of the proposed method to the uncertainty of the jammer's relative position. Figure 6 As shown, three typical interferer positions are selected for the first sidelobe before the main lobe illumination: one at the peak of the first sidelobe, one at the trough between the main sidelobe and the first sidelobe, and one at the midpoint of the first two positions.

[0082] The signal-to-noise ratio is set to 12 dB, and the other simulation parameters are consistent with Table 1. The experimental results are shown in Table 4.

[0083] Table 4 Experimental results of the analysis of the relative position of the jammer

[0084]

[0085] For the initial scheme without pulse compression, the accuracy is maintained above 86% when the jammers are deployed at points A, B, and C. After pulse compression, the accuracy reaches 92%. Therefore, the proposed scheme is robust to the uncertainty of the relative position of the jammers.

Claims

1. A method for rapid recognition of SAR imaging patterns based on neural networks, characterized in that: The following steps are involved: S1: Construct a sample dataset of SAR signal power graphs; Model the SAR signal intercepted by the ground jammer, and then simulate to obtain the power map samples of the SAR signal, and obtain a sufficient amount of sample data sets under different parameters; S2: Create and initialize the BP neural network model; Select the BP neural network model, create and initialize the model; S3: Use sample data sets to pre-train the BP neural network model; The generated sample data set is divided into a training set and a test set. The training set is input into the BP neural network model for pre-training, and the test set is used to calculate the recognition accuracy. According to the recognition results, the network parameters and hyperparameters of the BP neural network model are appropriately adjusted to obtain a better BP neural network model. S4: using the pre-trained BP neural network model to predict the power map samples of the SAR signal and obtain the corresponding imaging mode; The ground jammer receives the SAR signal echo in a fixed imaging mode, processes the SAR signal echo to obtain the SAR signal power map sample, and inputs the SAR signal power map sample into the pre-trained BP neural network model to output the corresponding imaging mode category; The modulation frequency of the SAR signal intercepted by the ground jammer is estimated by fractional Fourier transform, and then the SAR signal intercepted by the ground jammer is matched filtered, thereby equivalently improving the received signal-to-noise ratio. The modeling of the SAR signal intercepted by the ground jammer specifically includes: First, let the coordinates of the satellite-borne SAR be R(0,V r η,H), the ground point target coordinates are A(R G ,0,0), the coordinates of the ground reconnaissance receiver are P(x0,y0,0), and after derivation, the azimuth angle and range angle of the antenna pattern of the spaceborne SAR in the strip mode are respectively expressed as: The range and direction angles of the antenna pattern of the spaceborne SAR in scanning mode are expressed as: For spaceborne SAR in spotlight mode, the azimuth and range angles of the antenna pattern are expressed as: Among them, V r is the flight speed of the SAR platform, η represents the azimuth slow time, η c is the beam center crossing time of the point target, H is the orbit height, N s Indicates the ground distance of the center line of each sub-strip, N s is the number of sub-strips, R G is the ground distance from the point target to the radar track; Then, by using the azimuth angle and range angle of the antenna pattern of the spaceborne SAR in strip mode, scanning mode and spotlight mode, combined with the two-dimensional antenna pattern and modeled based on the LFM signal, the spaceborne SAR signal intercepted by the reconnaissance receiver is expressed as: Where τ represents the distance acceleration time, T r is the pulse width of the LFM signal, f c is the carrier frequency of the transmitted signal, K r is the modulation frequency of the LFM signal, R(η) is the distance from the spaceborne SAR to the reconnaissance receiver, and according to the straight line geometry assumption, it can be expressed as:

Images