A radar target recognition method based on sub-band synthesis and multi-feature fusion mechanism

Abstract

The present application belongs to the technical field of radar target recognition, and particularly relates to a radar target recognition method based on sub-band synthesis and multi-feature fusion mechanism. Firstly, the present application fuses sub-band data in phase to generate super wide band data, and then extracts time domain HRRP features and frequency domain GTD features of the target based on the super wide band signal. Compared with the sub-band signal, the super wide band signal can extract high-precision features required for fine identification of the target. Secondly, a multi-feature fusion mechanism is used to calculate the GTD feature with the largest recognition contribution, and the parameter feature is used to assist the training of the feature fusion network model, so as to improve the performance of the target recognition method. Therefore, the method can fully utilize the recognition information provided by the target sample set, extract more robust target features, and realize high-precision identification of similar targets.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of radar target recognition, and particularly relates to a radar target recognition method based on sub-band synthesis and multi-feature fusion mechanism. BACKGROUND

[0002] The high range resolution of the ultra-wideband radar makes it have many new characteristics and advantages in target recognition. Compared with the traditional radar system, the radar echo of the ultra-wideband system carries more abundant electromagnetic scattering information, which is beneficial to extract stable and reliable target features. Limited by the existing production process level, it is a very costly means to make the transmission bandwidth of a single radar very wide by upgrading the radar hardware system. In recent years, many scholars have begun to study the means of signal processing to increase the bandwidth of the transmitted signal to obtain a virtual ultra-wideband signal and then improve the range resolution of the radar system. The multi-sub-band synthesis to form an ultra-wideband technology is an effective way to form an ultra-wideband radar echo with a large bandwidth and high-precision imaging resolution by using the sub-band data of multiple spatially distributed radars working in different frequency bands. This technology improves the range resolution of the radar without upgrading the system hardware of the original radar system, so the sub-band synthesis to generate an ultra-wideband technology has great theoretical research significance and engineering application value.

[0003] With the rapid development of missile manufacturing technology and launch technology, ballistic missiles with high cruise speed, high damage precision and strong penetration ability have become one of the most aggressive tactical weapons in modern warfare. In order to counter the anti-missile system, the ballistic missile will release various types of decoys, fragments and false warheads in the middle of flight due to the penetration requirement. Therefore, how to identify the true warhead from the incoming ballistic target group composed of warheads, missile bodies and various decoys and fragments is the core task of the ballistic missile defense system. In recent years, due to the continuous progress of decoy technology research, modern decoys have similar target characteristics as true warheads, and only a slight structural difference exists between them in appearance, which undoubtedly puts forward higher requirements for the imaging capability of the radar sensor used in the anti-missile system. The S-band phased array radar and X-band dish antenna radar of the wideband system are limited by their own bandwidth, and their range resolution is difficult to meet the fine discrimination requirements of the ballistic target, so it is necessary to study the target feature extraction method based on the ultra-wideband radar echo to obtain high-precision identification features of the target as much as possible, so as to improve the defense performance of the anti-missile system.

[0004] In existing research on automatic radar target identification, High-Resolution Range Profiles (HRRPs) have always been an important feature for radar target identification due to their ease of data acquisition and simple imaging algorithms. HRRPs are the superposition of complex echo signals obtained by radar from the three-dimensional scattering points of a target on the radar line of sight, containing features such as the number, amplitude, and position of the target scattering points. HRRPs can be used directly as identification features, and feature extraction can be performed for identification. Commonly used HRRP feature extraction methods include Fourier transform and subspace feature dimensionality reduction. With the successful application of deep learning algorithms in target identification, Convolutional Neural Networks (CNNs), as an efficient feature extractor, can fully mine data-driven non-parametric features from HRRPs, thereby obtaining good classification results. However, ballistic target identification samples are mostly single-domain samples in the time and frequency domains. Therefore, the features beneficial to identification learned by CNNs from these sample sets are often limited to a single domain, failing to fully utilize the target identification information. This results in a large amount of computational resources being consumed during deep model iteration. To address the above issues, an effective approach is to establish a Geometric Diffraction (GTD) model in the frequency domain that conforms to the electromagnetic scattering characteristics of the target. This model, under conditions of relatively large bandwidth, can accurately estimate parameters such as the relative distance of the target scattering point, the frequency geometric factor, and the scattering intensity. Since the feature dimension of the GTD model parameter set is lower than that of HRRP, it can be used as feature parameters to assist in the training of deep models, significantly improving the target recognition speed under the same recognition method. Summary of the Invention

[0005] This invention proposes a radar target recognition method based on subband integration and multi-feature fusion (MFFM) mechanism. Based on the GTD model, it can effectively integrate multiple subband data to generate ultra-wideband data. The obtained ultra-wideband data is then used to extract the time-domain non-parametric features and frequency-domain parametric electromagnetic scattering features required for target recognition. Compared with classical radar target recognition methods, this invention can extract high-precision target recognition features and achieves higher target recognition accuracy.

[0006] The solution of this invention is as follows: First, a geometric diffraction model is used to model the echoes of multiple frequency sub-bands. Correlation processing is then used to coherently register the multiple sub-bands, and full-band data is obtained through sub-band extrapolation. Second, based on the generated ultra-wideband echoes, ultra-wideband HRRP data can be obtained in the time domain, and data-driven features of HRRP are extracted using a neural network. In the frequency domain, robust parametric features of the GTD model are extracted using a sparse reconstruction method and a dual-threshold discrimination method, and these features are then stitched onto a specified time-domain feature map using a multi-feature fusion mechanism. Finally, a parallel network structure is constructed, using the time-domain ultra-wideband HRRP features and the frequency-domain GTD features as inputs to train the network parameters and obtain the final classification result.

[0007] The technical solution of the present invention includes the following steps:

[0008] S1. Based on the GTD model, the radar echoes of the two sub-bands are modeled to obtain the discretized frequency responses Y1 and Y2. Correlation compensation is then used to make Y1 and Y2 coherent, specifically:

[0009] Define the i-th sub-band radar to use the linear frequency modulated signal s i (t) is sent to the target, and the frequency response of the i-th radar target echo is:

[0010]

[0011] Among them, S i (f) is s i Fourier transform of (t), where M is the number of target scattering centers, A m and α m Let f represent the amplitude and frequency dependence factors of the m-th scattering center, respectively. 0i R represents the initial frequency of the i-th sub-band. mi Let m represent the relative distance between the m-th scattering point and the i-th radar, and c be the speed of light;

[0012] Create an overcomplete dictionary Ψ i The element at position (n, l) of the i-th sub-band is represented as:

[0013]

[0014] Among them, f i (n) represents the discretized frequency of the i-th sub-band radar. It is the frequency dependence factor of the scattering point, l1∈{1,...,5}, L2 represents the relative distance between the scattering point and the i-th radar reference point, where L2 ∈ {1,...,L2}, and L2 is set according to the radar's range resolution.

[0015] The discrete frequency response Y of the i-th sub-band lightning i The matrix form is as follows:

[0016] Y i =Ψ i δ i +n i

[0017] Among them, Y i =(S Ri (f i (1)),…,S Ri (f i (N i )))T N i δ is the number of frequency points in the i-th sub-band. i n represents the unknown complex amplitude of the scattering center relative to the i-th sub-band. i For the noise vector, the dictionary Ψ i Contains α m and R mi The value information;

[0018] Using sub-band 1 as a reference, compensate for the phase difference between Y1 and Y2: Assume the range of frequency band 1 is [f 1L ,f 1U The range of frequency band 2 is [f] 2L ,f 2U ], then the full frequency range is [f L ,f U The frequency domain expressions for subband 1 and subband 2 are Y1(f) and Y2(f) respectively, and the phase difference between the two subbands is denoted as λ. Therefore, the cost function is constructed as follows:

[0019]

[0020] Among them, Y1 * (f), Y2 * (f) represents the target echoes of subband 1 and subband 2 extended to the full frequency band, respectively; minimizing the cost function yields λ. * At this point, the simulated full-band echo Y is obtained by arranging the sub-band vectors. * for

[0021] Y * =(Y1) T Y2 T exp(jλ * )) T

[0022] Find Y * Then, a compressed sensing method is used to solve the problem:

[0023] Y * =Ψ * δ * +n

[0024] In the formula, Ψ * and δ * These represent frequency ranges of [f] 1L ,f 1U ]∪[f 2L ,f 2U The dictionary and vectors at time ]; δ * The non-zero elements represent the number of scattering points, which are in Ψ * The corresponding columns are the amplitude, position, and frequency dependence factors of the target scattering point; parameter set It describes the characteristics of the radially distributed scattering centers of the target, reflecting the target's physical properties;

[0025] S2. The scattering center features are screened, and the robust electromagnetic scattering features of the target are obtained through a scattering center feature extraction method based on dual threshold discrimination. Specifically:

[0026] Define the target as having M scattering points relative to the entire frequency band, {A m The complex amplitudes of the scattering centers are represented by the groups m = 1, 2, ..., M, which are then normalized.

[0027]

[0028] Based on the target size, a range window W(r) is set near the target center. Scattering centers outside this range window are discarded. The window function is defined as follows:

[0029]

[0030] Among them, R m R1 and R2 represent the relative positions of the scattering centers, respectively, and the position values ​​corresponding to R1 and R2 are determined by the size of the target to be identified.

[0031] Set an amplitude threshold to filter the amplitude of scattering centers within the distance window:

[0032]

[0033] Among them, M 1 Indicates the order of the scattering center after amplitude filtering, where U represents the unit step function, and A th To determine the threshold, after amplitude-dimensional filtering, the feature parameter set of the scattering center is:

[0034] Based on the obtained first-order scattering centers and their relative positions, perform a first-order forward difference operation on them:

[0035] △R m =|R m+1 -R m |,(m=1,...,M 1 -1)

[0036] Within the distance window, set the distance threshold R. th To eliminate false scattering centers appearing near the target's strong scattering center:

[0037]

[0038] Among them, M *R represents the order of the scattering centers after being filtered by the distance dimension. th This represents the distance threshold; after the above two-dimensional threshold filtering, the precise electromagnetic scattering characteristics of the target are obtained as GTD parameter characteristics;

[0039] S3. Use the filtered GTD parameters to synthesize the ultra-wideband frequency domain echo, specifically:

[0040] The constructed frequency range is [f L ,f U The full-band dictionary Ψ U The columns of the dictionary correspond to the estimated values ​​of the relative distance and frequency dependence factor of the scattering points;

[0041] Using dictionary Ψ U and amplitude vector δ * Reconstructing ultrawideband frequency domain echo Y U for:

[0042] Y U =Ψ U δ *

[0043] S4. Perform inverse Fourier transform on the acquired echo signal to obtain its HRRP as the training set and test set;

[0044] S5. Construct a multi-feature fusion recognition network consisting of two processing branches: a CNN Block and a Transformer Block. The CNN Block comprises three convolutional pooling blocks and a fully connected block. Each convolutional pooling block consists of a pooling layer after two convolutional layers. The fully connected block consists of interconnected neurons. The convolutional kernels in the three convolutional pooling blocks are all 1×3 in size with a stride of 1, and the number of kernels is 32, 64, and 128 respectively. The feature map output from the final convolutional pooling block is batch normalized, concatenated with the GTD parameter features, and then input into the fully connected block. The fully connected block contains two fully connected layers. The first fully connected layer is followed by a Dropout operation to randomly discard some neurons. The activation function of the second fully connected layer is set to softmax, yielding the final object classification result of the CNN Block branch. The Transformer Block... The Block consists of two multi-encoder blocks and one fully connected block. Each multi-encoder block consists of six encoder layers stacked together, with an embedding layer placed before the first encoder layer to perform an embedding operation on the input data. After the input data is feature-encoded by the first multi-encoder block, the output feature map is concatenated with the GTD parameter features. The concatenated feature map is then input into the second multi-encoder block for feature encoding again before being input into the fully connected block. The fully connected block is used to connect the second multi-encoder block and the softmax classifier. The classification results of the two parallel branches, CNN Block and Transformer Block, are weighted through a linear layer to obtain the final classification result of the multi-feature fusion recognition network.

[0045] The obtained HRRP training data of each frequency band is input into the multi-feature fusion recognition network, and the model is pre-trained for 20 rounds based on the backpropagation algorithm, and its network weights are saved.

[0046] Select a feature from the GTD parameter feature set obtained from data of each frequency band, stitch it onto the output feature map of a specific position of the two parallel processing branches in the multi-feature fusion recognition network, load the saved network weights, input the HRRP test set and the selected model feature parameter set, and continue to train the model parameters for 20 rounds.

[0047] Based on the average classification results obtained from the training and test sample sets, calculate the recognition contribution F of the selected parameter feature class. c :

[0048]

[0049] Where loss1 and loss2 are the loss values ​​of the multi-feature fusion recognition network before and after GTD parameter feature concatenation, D1 and D2 represent the average feature map error matrices before and after GTD parameter feature concatenation, and ||D1-D2||2 represents the average error change of the output feature map before and after feature fusion; D1 and D2 are calculated by the following formula:

[0050]

[0051]

[0052] Where, δ l denoted as the error matrix of layer l, z represents the value of the current neuron, * represents convolution, σ′ represents the derivative with respect to the activation function, and ⊙ represents the Hadamard product;

[0053] Calculate the F-values ​​of all class parameter features sequentially. c Value, select F c The parameter with the highest value is connected to the feature map of the channel to assist in network training;

[0054] S6. Use the backpropagation algorithm to train the model parameters of the multi-feature fusion recognition network. The model is updated 50 times to obtain the trained multi-feature fusion recognition network.

[0055] S7. Use the trained multi-feature fusion recognition network to perform target recognition and obtain classification results.

[0056] The beneficial effects of this invention are as follows: This invention is a radar target recognition method based on subband integration and multi-feature fusion mechanisms. First, subband data is coherently fused to generate ultra-wideband data, and then the time-domain HRRP features and frequency-domain GTD features of the target are extracted based on the ultra-wideband signal. Compared to subband signals, high-precision features required for refined target recognition can be extracted based on the ultra-wideband signal. Second, the GTD feature with the greatest contribution to recognition is calculated through a multi-feature fusion mechanism, and this parameter feature is used to assist in the training of the feature fusion network model to improve the performance of the target recognition method. Therefore, this method can fully utilize the recognition information provided by the target sample set, extract more robust target features, and achieve high-precision recognition of similar targets. Attached Figure Description

[0057] Figure 1 This is a schematic diagram of the top-level structure of the method of the present invention;

[0058] Figure 2 This is a schematic diagram of the multi-feature fusion mechanism proposed in the method of the present invention;

[0059] Figure 3 Five similar radar targets;

[0060] Figure 4 The HRRP obtained by the method of the present invention using data from different frequency bands;

[0061] Figure 5 The results show the estimated parameters of the GTD model, where subplot (a) shows the relative distance distribution of the scattering points; subplot (b) shows the frequency dependence factor distribution of the scattering points; and subplot (c) shows the normalized amplitude distribution of the scattering points.

[0062] Figure 6 A schematic diagram of the ultrawideband HRRP corresponding to the five target categories.

[0063] Figure 7 The images show the training curves for different recognition methods, with sub-image (a) showing the CNN training curve, sub-image (b) showing the MFN training curve, and sub-image (c) showing the MFN training curve after incorporating model features. Detailed Implementation

[0064] The technical solutions in the embodiments of the present invention will now be clearly and completely described in conjunction with the accompanying drawings.

[0065] like Figure 1 The diagram shows the process of this invention, which specifically includes:

[0066] Step 1: Model the radar echoes of the two sub-bands based on the GTD model to obtain discretized frequency responses Y1 and Y2. Use correlation compensation to make Y1 and Y2 coherent. Use compressed sensing methods to solve the frequency domain electromagnetic characteristics of the correlated sub-bands.

[0067] Step 1-1: The i-th (i=1,2) sub-band radar will transmit the linear frequency modulated signal s i (t) is sent to the target. According to the geometric diffraction theory, the frequency response of the i-th radar target echo is:

[0068]

[0069] Where S i (f) is s i Fourier transform of (t), where M is the number of target scattering centers, A m and α m Let A represent the amplitude and frequency dependence factors of the m-th scattering center, respectively. m It is related to the shape and material of the target, α m Different values ​​of f correspond to different geometric structure types. 0i R represents the initial frequency of the i-th sub-band. mi Let m represent the relative distance between the m-th scattering point and the i-th radar, and c be the speed of light.

[0070] Step 1-2: Create an overcomplete dictionary Ψ iThe element at position (n, l) of the i-th (i = 1, 2) sub-band is represented as

[0071]

[0072] Where f i (n) represents the discretized frequency of the i-th sub-band radar, α l1 It is the frequency dependence factor of the scattering point, l1∈{1,...,5}, L2 represents the relative distance between the scattering point and the i-th radar reference point, where L2 ∈ {1,...,L2}, and L2 is set according to the radar's range resolution.

[0073] Steps 1-3: Discrete frequency response Y of the i-th (i=1,2) sub-band mine i The matrix form is

[0074] Y i =Ψ i δ i +n i ,

[0075] Where Y i =(S Ri (f i (1)),…,S Ri (f i (N i ))) T Assume the i-th sub-carrier has N i One frequency point. δ i n represents the unknown complex amplitude of the scattering center relative to the i-th sub-band. i For the noise vector, the dictionary Ψ i Contains α m and R mi The value information.

[0076] Steps 1-4: Using sub-band 1 as a reference, compensate for the phase difference between Y1 and Y2. Assume the range of frequency band 1 is [f 1L ,f 1U The range of frequency band 2 is [f] 2L ,f 2U ], then the full frequency range is [f L ,f U The frequency domain expressions for sub-band 1 and sub-band 2 are Y1(f) and Y2(f) respectively, and the phase difference between the two sub-bands is denoted by λ. Therefore, the cost function can be constructed as follows:

[0077]

[0078] Among them, Y1 * (f), Y2 *(f) represents the target echoes of subband 1 and subband 2 extended to the full frequency band, respectively. Minimizing the cost function yields λ. * At this point, the simulated full-band echo Y can be obtained by arranging the sub-band vectors. * for

[0079]

[0080] Find Y * Then, a compressed sensing method is used to solve the problem.

[0081] Y * =Ψ * δ * +n,

[0082] In the formula, Ψ * and δ * These represent frequency ranges of [f] 1L ,f 1U ]∪[f 2L ,f 2U The dictionary and vector at that time. δ * The non-zero elements represent the number of scattering points, which are in Ψ * The corresponding columns are the amplitude, position, and frequency dependence factors of the target scattering point. Parameter set It describes the characteristics of the radially distributed scattering centers of the target, reflecting the target's physical properties.

[0083] Step 2: Screen the scattering centers and obtain robust electromagnetic scattering characteristics of the target by using a scattering center feature extraction method based on dual threshold discrimination.

[0084] Step 2-1: Assume the target has M scattering points relative to the entire frequency band, {A m The complex amplitudes of the scattering centers are represented by the groups m = 1, 2, ..., M, which are then normalized.

[0085]

[0086] Step 2-2: Based on the target size, set a range window W(r) near the target center, and discard scattering centers located outside this range window. Define the window function as follows:

[0087]

[0088] Among them, R m R1 and R2 represent the relative positions of the scattering centers, respectively, and the position values ​​corresponding to R1 and R2 are determined by the size of the target to be identified.

[0089] Steps 2-3: Set an amplitude threshold and filter the amplitude of scattering centers within the distance window. The calculation process is as follows:

[0090]

[0091] Where M 1 Indicates the order of the scattering center after amplitude filtering, where U represents the unit step function, and A th To determine the threshold, after amplitude-dimensional filtering, the feature parameter set of the scattering center is:

[0092] Step 2-4: Based on the order scattering centers obtained in Step 2-3, and their relative positions, perform a first-order forward difference operation:

[0093] △R m =|R m+1 -R m |,(m=1,...,M 1 -1),

[0094] Within the distance window, set the distance threshold R. th To eliminate false scattering centers near the target's strong scattering center, the calculation process is as follows:

[0095]

[0096] Where M * R represents the order of the scattering centers after being filtered by the distance dimension. th This represents the distance threshold. After the above two-dimensional (amplitude and distance) threshold filtering, the precise electromagnetic scattering characteristics of the target are obtained. This invention will... As a feature of GTD parameters.

[0097] Step 3: Use the GTD model parameters selected in Step 2 to synthesize ultra-wideband frequency domain echoes.

[0098] Step 3-1: Construct a frequency range of [f L ,f U The full-band dictionary Ψ U The columns of the dictionary correspond to the estimated values ​​of the relative distance and frequency dependence factor of the scattering point.

[0099] Step 3-2: Using dictionary Ψ U and amplitude vector δ * Reconstructing ultrawideband frequency domain echo Y U for

[0100] Y U =Ψ U δ *

[0101] Step 4: Perform an inverse Fourier transform (IFFT) on the echo signal obtained in Step 3 to obtain its HRRP.

[0102] Step 5: Calculate the parameter feature with the largest contribution in the GTD model parameter set using the multi-feature fusion mechanism, and use this feature and the ultra-wideband HRRP as input to the multi-feature fusion identification network (MFN).

[0103] Step 5-1: Construct the MFN model. This invention proposes a parallel multi-feature fusion mechanism based on the CNN-based HRRP target recognition network. This mechanism consists of two processing branches: a CNN Block and a Transformer Block, as follows... Figure 2 As shown, the CNN Block mainly consists of three convolutional pooling blocks (CP Blocks) and a fully connected block (FC Block). The CP Block has a pooling layer after two convolutional layers, and the FC Block consists of interconnected neurons. The convolutional kernels in the three CP Blocks are all 1×3 in size with a stride of 1. The number of kernels from CP1 to CP3 are 32, 64, and 128 respectively. The feature map output by CP3 is connected to the GTD parameter features after passing through a batch normalization (BN) layer. BN is used to handle the vanishing and exploding gradient problems. The FC Block contains two fully connected layers. A dropout operation is added after the first fully connected layer to randomly discard some neurons; the activation function of the second fully connected layer is set to softmax, which yields the final object classification result of the CNN Block branch. The Transformer Block consists of two multi-encoder blocks (TE Blocks) and one fully connected block. Each TE Block consists of six stacked encoder layers, with an embedding (EMB) layer preceding the first encoder layer (Encoder1) to perform embedding operations on the input data. After feature encoding of the input data by the first TE Block, the feature map output by Encoder6 is concatenated with the GTD parameter features. This concatenated feature map is then input into the second TE Block for further feature encoding. The FC Block connects the second TE Block and the softmax classifier. The classification results from the two parallel branches, CNNBlock and Transformer Block, are weighted through a linear layer to obtain the final classification result of the MFN. A possible structural information of the proposed fusion model is shown in Table 1.

[0104] Table 1. Model Structure of the Invention

[0105]

[0106] Step 5-2: Based on the HRRP data of each frequency band obtained in Step 4, input the HRRP training set into MFN, and pre-train the model for 20 rounds based on the backpropagation algorithm, and save its network weights.

[0107] Step 5-3: Select a feature from the GTD parameter feature set obtained from the data of each frequency band, and stitch it onto the output feature map of a specific position of the two parallel processing branches as in step 4-1. Load the network weights saved in step 5-2, input the HRRP test set and the selected model feature parameter set, and continue to train the model parameters for 20 rounds.

[0108] Step 5-4: Based on the average classification results obtained in Steps 5-2 and 5-3 on the training and test sample sets, calculate the recognition contribution F of the selected parameter feature. c :

[0109]

[0110] Where loss1 and loss2 are the loss values ​​of MFN before and after GTD parameter feature concatenation (this invention uses the cross-entropy loss function), D1 and D2 represent the average feature map error matrix before and after GTD parameter feature concatenation, and ||D1-D2||2 represents the average error change of the output feature map before and after feature fusion. loss1 and D1 are calculated according to step 5-2, and loss2 and D2 are calculated according to step 5-3. D1 and D2 are obtained by the following formula:

[0111]

[0112]

[0113] Where δ l Let represent the error matrix of layer l, z represent the value of the current neuron, * represent convolution, σ′ represent the derivative with respect to the activation function, and ⊙ represent the Hadamard product.

[0114] Step 5-5: Based on steps 5-3 to 5-4, calculate the F-values ​​of these four types of parameter features sequentially. c Value, select F c The parameter with the highest value is connected to the feature map of the channel, thereby assisting in network training.

[0115] Step 6: Train the MFN model parameters using the backpropagation algorithm. The model is updated 50 times, with a batch size of 64 and a learning rate of 2e-6. AdamW is selected as the Optimizer, and StepLR is set as the Scheduler.

[0116] Step 7: Use the trained model to identify the target and obtain the classification results.

[0117] To verify the effectiveness of the identification algorithm proposed in this invention, the electromagnetic scattering fields of five similar ballistic targets were simulated using FEKO software. The target structures are as follows: Figure 3 As shown. Target 1 is a typical three-groove warhead, with the distances from the first to third grooves to the apex of the cone being 110 mm, 440 mm, and 1400 mm, respectively. The remaining targets are decoys with real warheads. The shapes of false targets 4 and 5 are very similar to target 1 to increase the difficulty of classification and to test the recognition performance of the algorithm proposed in this invention.

[0118] The experiment used two sub-bands for fusion with frequency ranges of [4.5, 5.5] and [9, 11] GHz, Δf1 = Δf2 = 2.4336 MHz, a radar incident angle of 25°, and VV polarization. For the FEKO simulation data of target 1, Figure 4 The diagram shows the corresponding HRRP obtained using data from different bands. The relative distances between the HRRP peaks generated by multi-subband fusion and those generated by the actual full-band HRRP are 30.19m, 30.45m, 31.33m, and 31.71m, respectively, corresponding to the positions of the four scattering points of the three-groove warhead. The HRRP generated by the high and low subbands has poor resolution and cannot effectively distinguish the scattering points. This demonstrates that the multi-subband fusion method used in this invention can improve radar range resolution, thus facilitating target identification. To observe the scattering information fused across the entire band, Figure 5 The results of parameter feature estimation based on the GTD model are presented; all results were obtained through 100 independent trials. Figure 4 and Figure 5 It can be seen that the ultra-wideband electromagnetic scattering parameter characteristics generated by subband synthesis in this invention have a good consistency with the target geometric structure characteristics, providing an accurate physical description of the target scattering center.

[0119] To verify the superiority of the method of this invention, comparative experiments were conducted with some classic radar target recognition methods. For each target, the radar incident angle ranged from 10° to 35°, with 1000 HRRP samples taken every 5°. Therefore, the same target sample set contained 6000 HRRP samples from subband 1, subband 2, and ultra-wideband. Three angles were randomly selected from the HRRP sample set of six angles as the training sample set, and the remaining samples were used as the test set. Considering real-world recognition scenarios, 10dB of additive white Gaussian noise was added to the target dataset. Figure 6 The visualization of HRRP for five target categories is presented.

[0120] Similarly, based on echo data of different targets in different frequency bands, the method of this invention is used to extract parameter features of sub-band 1, sub-band 2, and the synthesized ultra-wideband GTD model. Assuming the incident angle is between 10° and 35°, with an interval of 5°, 1000 independent trials are conducted for each incident angle to obtain 1000 sets of parameter estimation results. Table 2 shows the recognition contribution of the four types of parameter features normalized using the method of this invention:

[0121] Table 2. Contribution of GTD parameter features to identification F c Value (normalized)

[0122]

[0123] As can be observed from the table, the parameter characteristic R m Compared to other parameters, F in datasets across different frequency bands c The value is higher, so it is chosen as the frequency domain parameter feature used for splicing in MFN.

[0124] Figure 7 The recognition rate curves of three test methods on subband and ultra-wideband samples are shown. Among all the target recognition methods tested, the ultra-wideband HRRP based on multi-subband fusion has the highest recognition accuracy compared to the subband HRRP samples, indicating that the ultra-wideband HRRP samples used in the method of this invention can effectively improve the accuracy of target recognition. Figure 7 By comparing different testing methods, compared with traditional CNN-based object recognition methods, the method of this invention not only utilizes convolution for feature extraction, but also considers the temporal information of HRRP data. That is, the MFN model uses parallel network branches to fuse target characteristics at different scales, and at the same time utilizes the spatial and temporal correlation of the input sequence, thus possessing stronger feature extraction capabilities. Therefore, MFN demonstrates its advantage in recognition performance on HRRP datasets of different frequency bands. In addition, by using R... m By parametrically concatenating the output feature maps of the two parallel branches, the MFN can utilize these frequency domain parameter features for auxiliary training to improve target recognition accuracy. Compared to deep neural networks that autonomously learn only data-driven features within a single domain from the HRRP sample set, the method of this invention simultaneously utilizes the non-parametric features of the HRRP time domain and the parametric features of the GTD frequency domain, enabling the neural network to simultaneously utilize the recognition information corresponding to the cross-domain features provided by the ballistic target sample set, thereby improving the network's recognition rate and generalization ability.

Claims

1. A radar target recognition method based on sub-band integration and multi-feature fusion mechanism, characterized in that, Includes the following steps: S1. Based on the GTD model, the radar echoes of the two sub-bands are modeled to obtain the discretized frequency response. Using relevant compensation to enable Coherence, specifically: Definition of the first The sub-band radar will transmit linear frequency modulated signals. Send to target, number The frequency response of the radar target echo is: ， in, yes Fourier transform, The number of target scattering centers, and They represent the first The amplitude and frequency dependence factors of each scattering center Indicates the first The initial frequency of each subband, Indicates the first The scattering point and the th The relative distance between the radars The speed of light; Build an overcomplete dictionary , No. The first of the sub-bands The element of a bit is represented as: ， in, Indicates the first Discretized frequencies for individual radar units It is the frequency dependence factor of the scattering point. , Indicates the scattering point relative to the first The relative distance between radar reference points , Based on the radar's range resolution settings, ; No. Discrete frequency response of a sub-band radar The matrix form is as follows: ， in, , It is the first Number of frequency points in each subband Indicates the scattering center relative to the first The unknown complex amplitude of the sub-band For noise vectors, dictionary Contains and The value information; Using sub-band 1 as a reference, compensation and Phase difference: Assuming the range of frequency band 1 is The range of frequency band 2 is Then the full frequency band range is The frequency domain expressions for sub-band 1 and sub-band 2 are respectively , The phase difference between the two subbands is expressed as The cost function is then constructed as follows: ， in, , These represent the target echoes of subband 1 and subband 2 extended to the full frequency band, respectively; minimizing the cost function yields... At this point, the simulated full-band echo is obtained by arranging the sub-band vectors. for ， Find Then, a compressed sensing method is used to solve the problem: ， In the formula, and They represent frequency ranges respectively. The dictionary and vectors at that time; The non-zero elements represent the number of scattering points, which in The corresponding columns are the amplitude, position, and frequency dependence factors of the target scattering point; parameter set It describes the characteristics of the radially distributed scattering centers of the target, reflecting the target's physical properties; S2. The scattering center features are screened, and the robust electromagnetic scattering features of the target are obtained through a scattering center feature extraction method based on dual threshold discrimination. Specifically: Define the target relative to the entire frequency band One scattering point, The complex amplitude of the scattering center is normalized: ， Set a distance window near the target center based on the target size. Scattering centers located outside this distance window are removed, and the window function is defined as follows: ， in, The relative position of the scattering center and These represent the lower and upper boundaries of the distance window interval, respectively. and The corresponding position value is determined by the size of the target to be identified; Set an amplitude threshold to filter the amplitude of scattering centers within the distance window: ， in, This indicates the order of the scattering center after amplitude filtering. Represents the unit step function. To determine the threshold, after amplitude-dimensional filtering, the feature parameter set of the scattering center is: ; Based on the obtained first-order scattering centers and their relative positions, perform a first-order forward difference operation on them: ， Set a distance threshold within the distance window. To eliminate false scattering centers appearing near the target's strong scattering center: ， in, Represents the order of the scattering centers after being filtered by the distance dimension. The distance threshold is represented; after two-dimensional threshold filtering, the precise electromagnetic scattering characteristics of the target are obtained as GTD parameter features; S3. Use the filtered GTD parameters to synthesize the ultra-wideband frequency domain echo, specifically: The constructed frequency range is Full-band dictionary The columns of the dictionary correspond to the estimated values ​​of the relative distance and frequency dependence factor of the scattering points; Using a dictionary and amplitude vector Reconstructing ultra-wideband frequency domain echo for: ， S4. Perform inverse Fourier transform on the acquired echo signal to obtain its HRRP as the training set and test set; S5. Construct a multi-feature fusion recognition network consisting of two processing branches: a CNN Block and a Transformer Block. The CNN Block comprises three convolutional pooling blocks and a fully connected block. Each convolutional pooling block consists of a pooling layer after two convolutional layers. The fully connected block consists of interconnected neurons. The convolutional kernels in the three convolutional pooling blocks are all 1×3 in size with a stride of 1, and the number of kernels is 32, 64, and 128 respectively. The feature map output from the final convolutional pooling block is batch normalized, concatenated with the GTD parameter features, and then input into the fully connected block. The fully connected block contains two fully connected layers. The first fully connected layer is followed by a Dropout operation to randomly discard some neurons. The activation function of the second fully connected layer is set to softmax, yielding the final object classification result of the CNN Block branch. The Transformer Block... The Block consists of two multi-encoder blocks and one fully connected block. Each multi-encoder block consists of six encoder layers stacked together, with an embedding layer placed before the first encoder layer to perform an embedding operation on the input data. After the input data is feature-encoded by the first multi-encoder block, the output feature map is concatenated with the GTD parameter features. The concatenated feature map is then input into the second multi-encoder block for feature encoding again before being input into the fully connected block. The fully connected block is used to connect the second multi-encoder block and the softmax classifier. The classification results of the two parallel branches, CNN Block and Transformer Block, are weighted through a linear layer to obtain the final classification result of the multi-feature fusion recognition network. The obtained HRRP training data of each frequency band is input into the multi-feature fusion recognition network, and the model is pre-trained for 20 rounds based on the backpropagation algorithm, and its network weights are saved. Select a feature from the GTD parameter feature set obtained from data of each frequency band, stitch it onto the output feature map of a specific position of the two parallel processing branches in the multi-feature fusion recognition network, load the saved network weights, input the HRRP test set and the selected model feature parameter set, and continue to train the model parameters for 20 rounds. Based on the average classification results obtained from the training and test sample sets, the recognition contribution of the selected parameter feature is calculated. : ， in These are the loss values ​​of the multi-feature fusion recognition network before and after GTD parameter feature concatenation. and These represent the average feature map error matrices before and after GTD parameter feature concatenation. This represents the average error change of the output feature map before and after feature fusion; and It is calculated by the following formula: ， ， in, express The error matrix of the layer, This represents the current value of the neuron. Represents convolution. This represents the differentiation with respect to the activation function. It represents the Hadamardi (or Hadama) stack; Calculate all class parameter features sequentially Value, selection The parameter with the highest value is connected to the feature map of the channel to assist in network training; S6. Use the backpropagation algorithm to train the model parameters of the multi-feature fusion recognition network. The model is updated 50 times to obtain the trained multi-feature fusion recognition network. S7. Use the trained multi-feature fusion recognition network to perform target recognition and obtain classification results.

Images