Spatial micro-motion target recognition method based on complex-valued dynamic fusion network

Abstract

This invention discloses a spatial micro-motion target recognition method based on a complex-valued dynamic fusion network, mainly addressing the problem of low recognition rates caused by the inability of existing spatial micro-motion target recognition methods to fully utilize the modal information of the target. The scheme involves: generating complex-valued RCS data and complex-valued HRRP data; preprocessing the complex-valued RCS data to obtain two-dimensional complex-valued recursive graph (RP) data; dividing the RP and HRRP data to generate training and testing sets; constructing a complex-valued dynamic fusion network containing a multi-modal feature extraction module, a dynamic fusion module, and a classifier; inputting the training set data into the complex-valued dynamic fusion network and training it using a backpropagation algorithm; inputting the testing set data into the trained complex-valued dynamic fusion network and obtaining the final recognition accuracy through forward propagation of the network. This invention can fully utilize the complex-valued characteristics of radar wide and narrow band echoes, effectively improving target recognition accuracy and noise resistance, and can be used for target tracking in aerospace.

Description

Technical Field

[0001] This invention belongs to the field of radar technology, and further relates to a method for identifying space micro-movement targets, which can be used for target tracking in aerospace. Background Technology

[0002] Micro-motion refers to the vibration or rotation of a target or target component relative to the radar line of sight, excluding the main body motion, such as spin, precession, and tumbling. When radar uses alternating wide and narrow band modes to observe micro-moving targets in space, it can simultaneously obtain both narrow-band and wide-band echoes of the target. The narrow-band radar cross-section (RCS) sequence reflects the overall scattering intensity of the target over time; the wide-band high-resolution range profile (HRRP) characterizes the target's structural characteristics and size information. Therefore, RCS and HRRP contain complementary target separable features. If these features can be fully explored and integrated, the accuracy of identifying micro-moving targets in space can be effectively improved.

[0003] However, most existing technologies only utilize RCS or HRRP for single-modal feature extraction and recognition, and most of them rely on real-valued neural networks for feature extraction, thus discarding the rich features contained in the phase of RCS and HRRP sequences, which limits the performance of spatial micro-movement target recognition methods.

[0004] In their paper "Space Precession Target Classification Based on Radar High-Resolution Range Profiles," Yizhe Wang et al. constructed a deep convolutional neural network (DCNN) for feature extraction and recognition of HRRP data. The specific steps are as follows: HRRP data is directly input into a DCNN consisting of convolutional layers, pooling layers, and fully connected layers. Through training, useful features for HRRP data classification are automatically extracted, and finally, a softmax classifier outputs the classification result. While this method significantly improves recognition accuracy compared to traditional feature extraction and recognition methods, its overall recognition performance is poor because it only utilizes single-modal features and only leverages the amplitude information of the HRRP sequence.

[0005] In their paper "HRPnet: High-Dimensional Feature Mapping for Radar Space Target Recognition," Dong J et al. proposed a radar space target recognition method based on high-dimensional feature maps. The specific implementation steps are as follows: First, a sparse autoencoder is used to extract deep generalized features from the RCS and HRRP sequences, respectively. Then, the three extracted features are encoded into Gram angle fields and concatenated into multi-channel high-dimensional features, which are finally fed into a Softmax classifier to output predicted labels. Although this method improves recognition performance compared to single-feature conditions, it fails to consider the problem of mining and effectively fusing separable features from different modalities when their separability varies significantly. Furthermore, it only utilizes the amplitude information of the RCS and HRRP sequences, ignoring the separable features inherent in their phase, thus resulting in poor target recognition performance.

[0006] Chinese patent document CN 117289231 A discloses a radar target recognition method based on subband synthesis and multi-feature fusion mechanism. It first coherently fuses subband data to generate ultra-wideband data, then extracts time-domain HRRP features and frequency-domain geometric diffraction theory (GTD) features. Finally, it inputs these two types of features into a multi-feature fusion recognition network composed of CNN Block and Transformer Block for training. Although this method fuses HRRP and GTD features for training, the constructed fusion recognition network is a real network, which cannot be trained on complex data, and the fusion recognition network is non-dynamic. Regarding multimodal recognition, the information content of the same sample differs across modalities; for the same sample, when the confidence levels differ significantly across modalities, the information contained in the extracted features may contradict each other. In such cases, directly fusing multimodal features for recognition may lead to misclassification of samples, impairing the recognition performance of spatially moving targets. Summary of the Invention

[0007] The purpose of this invention is to propose a spatial micro-moving target recognition method based on a complex-valued dynamic fusion network, in order to solve the problems of existing methods such as single feature extraction, limited recognition performance, and insufficient utilization of echo information due to neglect of phase information. This improves the recognition performance of micro-moving targets.

[0008] The technical approach to achieving the objective of this invention is as follows: by dynamically estimating the amount of feature information and modal information, multimodal information is reasonably and fully integrated to enhance the interpretability of multimodal fusion; at the same time, attention is paid to the amplitude and phase information of the data, and echo information is fully utilized to improve the performance of micro-moving target recognition.

[0009] Based on the above ideas, the technical solution of the present invention includes the following:

[0010] (1) Generate dataset:

[0011] 1a) Establish three-dimensional geometric electromagnetic models of six types of space micro-moving targets, and obtain their narrowband and broadband echoes respectively;

[0012] 1b) Dynamically extract the original echo to obtain complex-valued RCS data and complex-valued HRRP data;

[0013] 1c) The complex-valued RCS data of six types of spatial micro-moving targets are preprocessed and converted into two-dimensional complex-valued recursive graph (RP) data that can reflect the dynamic characteristics of the time series.

[0014] (2) Divide the complex value RP data and complex value HRRP data of six types of spatial micro-motion targets to generate training set and test set;

[0015] (3) Constructing a complex-valued dynamic fusion network:

[0016] 3a) Establish a multimodal feature extraction module consisting of two convolutional neural networks connected in parallel, used to extract complex-valued RP features and complex-valued HRRP features;

[0017] 3b) Establish a dynamic fusion module composed of a feature information dynamic estimation submodule and a modality information dynamic estimation submodule, which is used to dynamically fuse the features of the two modalities, complex value RP and complex value HRRP.

[0018] 3c) Select the existing Softmax classifier to obtain the final multimodal classification result;

[0019] 3d) The multimodal feature extraction module, dynamic fusion module, and classifier module are cascaded in sequence to form a complex value dynamic fusion network;

[0020] (4) Input the training set data into the complex value dynamic fusion network and train it using the backpropagation algorithm to obtain the trained complex value dynamic fusion network.

[0021] (5) Input the test set data into the trained complex value dynamic fusion network to obtain the spatial micro-movement target recognition result.

[0022] Compared with the prior art, the present invention has the following advantages:

[0023] First, by constructing a multimodal feature extraction module composed of two convolutional neural networks connected in parallel, this invention can simultaneously extract the amplitude and phase information of RP features and HRRP features. Compared with the prior art, which only uses the amplitude information of RCS data and HRRP data for identification, this invention makes full use of echo information.

[0024] Secondly, by constructing a dynamic fusion module composed of a feature information dynamic estimation submodule and a modal information dynamic estimation submodule, the present invention can dynamically fuse the features of two modalities, complex value RP and complex value HRRP. Compared with the existing technology that only uses single modal data for identification, the present invention achieves full and reasonable utilization of multimodal data of micro-moving targets. Attached Figure Description

[0025] Figure 1 This is a flowchart illustrating the implementation of the present invention;

[0026] Figure 2 This refers to the multimodal feature extraction network structure constructed in this invention;

[0027] Figure 3 This is a diagram of the overall structure of the complex-valued dynamic fusion network constructed in this invention. Detailed Implementation

[0028] To enable those skilled in the art to better understand the present invention, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of the present invention.

[0029] It should be noted that the step numbers in the specification and claims of this invention are only for the purpose of clearly describing the embodiments of this invention and facilitating understanding, and their order is not limited.

[0030] Reference Figure 1 The implementation steps of this embodiment include the following:

[0031] Step 1: Establish the target model and generate dynamic electromagnetic echo data.

[0032] 1.1) Construct three-dimensional geometric models of six types of targets in Feko simulation software: flat-bottomed cone targets, flat-bottomed cone-shaped targets, spherical-bottomed cone targets, spherical-bottomed cone-shaped targets, frustum-shaped cone-shaped targets, and grooved flat-bottomed cone targets. Use the standard PO method to obtain the full-angle domain static electromagnetic echo of each type of target.

[0033] 1.2) Extract the full-angle static electromagnetic echo based on the target's micro-motion pattern to generate dynamic electromagnetic echoes for each target type: Set the radar carrier frequency to 10GHz, the narrowband signal bandwidth to 20MHz, the wideband signal bandwidth to 2GHz, and the operating frequency to the X-band. The pulse repetition frequency of both the wideband and narrowband radars is 50Hz, and the observation time is 2s. The initial equivalent elevation angle of the target is set to vary continuously in 1° intervals within the range of 21° to 35°. The precession frequency and precession angle at each elevation angle are taken at equal intervals of 11. 1815 wideband echo data and 1815 narrowband echo complex-valued RCS data are generated for each target type.

[0034] 1.3) Perform range pulse compression on each broadband echo data to obtain the complex HRRP data of each target;

[0035] 1.4) To simulate echo data in real-world scenarios as closely as possible and to verify the robustness of the proposed algorithm to noise, Gaussian noise was added in 5dB intervals within the range of 5dB-15dB to obtain complex HRRP data and complex RCS data under different signal-to-noise ratios.

[0036] Step 2: Preprocess the complex value RCS data to obtain complex value RP data.

[0037] 2.1) Represent the complex-valued RCS time series as follows: And calculate the phase space matrix Y of the time series T:

[0038]

[0039] where t i Let y be the element at time i in the complex-valued RCS time series. j Let be the j-th column of the phase space matrix, i∈[1,N], j∈[1,M], N is the length of the sequence, M=N-(l-1)τ is the number of columns in the phase space matrix corresponding to the time series, τ is the delay time, and l is the dimension of the encoded features;

[0040] 2.2) Calculate the recurrence graph corresponding to the time series T based on the phase space matrix Y:

[0041]

[0042] Where CRP(p,q) represents the complex recursive graph calculated using the p-th and q-th columns of the phase space matrix Y, || ||2 represents L2 normalization, and p,q∈[1,M]. and They represent the real and imaginary parts of a complex number, respectively.

[0043] In this embodiment, the delay time and coding feature dimension in the time series T are initialized to τ=1 and l=1.

[0044] Step 3: Divide the complex-valued HRRP data and complex-valued RP data into training and test sets based on the initial equivalent pitch angle.

[0045] 3.1) Select 1210 samples and their corresponding labels corresponding to the initial equivalent pitch angles of 21° to 30° for each type of target to form a training set;

[0046] 3.2) Select 605 samples and corresponding labels corresponding to the initial equivalent pitch angles of 31° to 35° for each type of target to form a test set.

[0047] Step 4: Construct a complex-valued dynamic fusion network.

[0048] refer to Figure 3 The implementation of this step includes the following:

[0049] 4.1) Construct a multimodal feature extraction module consisting of two convolutional neural networks connected in parallel:

[0050] The two convolutional neural networks mentioned are respectively as follows: Figure 2 (a) and Figure 2 As shown in (b), Figure 2 (a) shows the first convolutional neural network used to extract complex value RP features; Figure 2 (b) shows a second convolutional neural network used to extract complex HRRP features. They have the same structure but slightly different parameters.

[0051] The structure of each convolutional neural network is as follows: 1st complex convolutional layer → 1st complex pooling layer → 2nd complex convolutional layer → 2nd complex pooling layer → 3rd complex convolutional layer → 3rd complex pooling layer → 4th complex convolutional layer → 4th complex pooling layer → complex fully connected layer;

[0052] in:

[0053] The parameters of the first convolutional neural network are:

[0054] The first complex convolutional layer has a kernel size of 5×5, 4 channels, and a stride of 1×1.

[0055] The pooling kernel size of the first complex pooling layer is 2×2, and the step size is 2×2.

[0056] The second complex convolutional layer has a kernel size of 5×5, 4 channels, and a stride of 1×1.

[0057] The pooling core size of the second complex pooling layer is 2×2, and the step size is 2×2.

[0058] The third complex convolutional layer has a kernel size of 3×3, 8 channels, and a stride of 1×1.

[0059] The pooling kernel size of the third complex pooling layer is 2×2, and the step size is 2×2.

[0060] The fourth complex convolutional layer has a kernel size of 3×3, 8 channels, and a stride of 1×1.

[0061] The pooling kernel size of the fourth complex pooling layer is 2×2, and the step size is 2×2.

[0062] Fully connected layers are used to map the feature dimensions to 128.

[0063] The parameters of the second convolutional neural network differ from those of the first convolutional neural network in that:

[0064] The second complex convolutional layer has 8 channels;

[0065] The third complex convolutional layer has 16 channels;

[0066] The fourth complex convolutional layer has 32 channels and a stride of 2×2.

[0067] 4.2) Construct a dynamic fusion module consisting of a feature information dynamic estimation submodule and a modal information dynamic estimation submodule. This module is used to estimate the information content of the extracted complex-valued features and modalities, and then perform reasonable dynamic fusion. Specifically, the complex-valued features are first input into the feature information dynamic estimation submodule to obtain feature samples that retain important information. Then, they are input into the modal information dynamic estimation submodule to obtain the classification confidence scores for each modality. Finally, the feature samples are multiplied by the classification confidence scores and concatenated to achieve fusion.

[0068] Its specific implementation is as follows:

[0069] The feature information quantity dynamic estimation submodule described in 4.2.1) first introduces a feature information quantity encoder x m →w m Assess the information content of the two types of features, complex value RP and complex value HRRP. m And its application is determined by the loss function. The KL divergence is controlled, and then a sparsity strategy is used to determine the key features of different samples in the spatial micro-movement target fusion and recognition task. Finally, a sparse gating strategy is used to retain important information and suppress useless information, thereby enabling x to be... m Convert to The calculation formula is as follows:

[0070]

[0071] Where x m For the feature sample, w m For its information content, d m F represents the dimension of the complex-valued feature. CVThis represents a fully connected layer, where abs represents the modulus value, θ(·) represents the Tanh activation function, m represents the mode number, and M represents the number of mode types. This represents the average activation level of the i-th node in terms of feature information content. B represents the size of the current batch, and ρ represents a sparse parameter close to 0. This represents the feature sample after retaining important information, and ⊙ represents the Hadamard product;

[0072] 4.2.2) The modal information dynamic estimation submodule will Input to a fully connected layer and use Function activation to obtain modal features h m Then it is split into two outputs, one of which is input into the classifier f. m Get h m Corresponding correct class probability TCP m and the loss function of the classifier Another input is sent to the confidence network g. m Get TCP m TCP estimated value m Loss function of confidence network

[0073]

[0074] TCP m =g m (h m )

[0075]

[0076] Wherein, classifier f m It is composed of fully connected layers and a softmax activation function, y = [y1,...,y k ,...,y K ], y k Let the k-th element of the category label y be the predicted probability. K represents the number of target categories. Let represent the posterior probability that a feature sample in the m-th modality is predicted to belong to the k-th class. TCP m ∈[0,1]; belief network g m TCP is composed of a fully connected layer and a Tanh activation function. m TCP estimated value m ∈[0,1];

[0077] In this embodiment, the sparsity parameter ρ is initialized to 0.05, the number of modal types M = 2, the batch size B = 128, and the number of target categories K = 6;

[0078] 4.2.3) The above-mentioned different modal features h m The estimated value of the correct class probability TCP m The multiplication of different modalities is performed, and the products are concatenated to obtain the multimodal fusion feature h. Finally, the nonlinear representation capability of the multimodal fusion feature is enhanced by a complex fully connected layer F, resulting in the enhanced nonlinear multimodal fusion feature.

[0079]

[0080] [] indicates a splicing operation;

[0081] 4.3) The existing Softmax classifier module is selected for processing the enhanced nonlinear multimodal fusion features described above. The classification process is performed to obtain the final classification results of the multimodal features. The specific implementation is as follows:

[0082] Multimodal fusion features Mapped to a K-dimensional rating vector [s1,...,s i ,...,s K The Softmax function is used to transform the rating vector into a probability vector [p1,...,p]. i ,...,p K The index of the element with the maximum value in the probability vector is the final multimodal classification result C.

[0083] C = argmax(p1,...,p i ,...,p K )

[0084] in s represents the probability that the input sample is identified as the i-th target category. i Let represent the score of the network model on the i-th target category, i∈[1,K], where K represents the number of target categories. The e-index is used to transform the score values ​​in (-∞,+∞) to (0,+∞) without affecting the relative magnitudes. Used for normalization;

[0085] 4.4) The above-mentioned multimodal feature extraction module, dynamic fusion module and Softmax classifier module are cascaded in sequence to form a complex value dynamic fusion network.

[0086] Step 5: Train the complex-valued dynamic fusion network.

[0087] 5.1) Initialize network parameters;

[0088] In this example, the network is set to 60 training iterations, a learning rate of 0.1, a learning decay coefficient of 0.96, a momentum of 0.96, and a weight decay coefficient of 0.0001.

[0089] 5.2) Set the network loss function

[0090]

[0091] in, Let y be the loss function of the Softmax classifier. k p represents the value of the k-th element of the category label. k This represents the posterior probability that the predicted multimodal fusion feature belongs to the k-th class. The loss function represents the dynamic estimation submodule for feature information. Let λ1 and λ2 represent the loss function of the modal information dynamic estimation submodule, respectively. The hyperparameters of the weights;

[0092] 5.3) Using the loss function Calculate the error between the network output value and the target value, and use the SGD optimizer to update the network parameters between each layer;

[0093] 5.4) Repeat steps 5.1) to 5.3) until the error between the network output value and the target value is basically stable or the number of training iterations is reached, then stop training to obtain the trained complex value dynamic fusion network.

[0094] Step 6: Calculate the recognition results of the test set.

[0095] The complex value RP data and complex value HRRP data of the test set are input into the trained complex value dynamic fusion network, and the final recognition accuracy is obtained through the forward propagation of the network.

[0096] The effects of this invention can be further illustrated by the following simulation experiments:

[0097] I. Simulation Experiment Conditions

[0098] The simulation experiment of this invention uses the standard PO method to generate wide and narrow band radar echo data of six types of spatial micro-moving targets. These six types of targets are: flat-bottomed cone, flat-bottomed cone cylinder, spherical-bottomed cone, spherical-bottomed cone cylinder, frustum cone cylinder, and grooved flat-bottomed cone.

[0099] In the FEKO software, the radar carrier frequency is set to 10GHz, the narrowband signal bandwidth to 20MHz, the wideband signal bandwidth to 2GHz, and the operating frequency to the X-band; the pulse repetition frequency of both the wideband and narrowband radars is 50Hz, and the observation time is 2s.

[0100] The initial equivalent pitch angle of the target is set to vary continuously in 1° intervals within the range of 21°-35°. The precession frequency and precession angle at each pitch angle are taken at equal intervals of 11. 1815 narrowband and broadband echo data are generated for each type of target. Therefore, a total of 10890 narrowband and broadband echo data can be generated for the six types of targets.

[0101] Gaussian white noise was added to the echoes, and complex-valued RCS data was preprocessed to obtain complex-valued RP data, generating echo samples with signal-to-noise ratios of 5dB, 10dB, and 15dB, respectively. Samples corresponding to initial pitch angles of 21°–30° were used as the training set, yielding 7260 training samples; samples corresponding to initial pitch angles of 31°–35° were used as the test set, yielding 3630 test samples.

[0102] The simulation experiment hardware platform is an Intel(R) Core(TM) i7-10700@2.90GHz CPU with 16GB RAM; the simulation experiment software platform is FEKO 2017, MATLAB 2021b, and Python 3.8.

[0103] II. Simulation Content and Result Analysis

[0104] Under the above simulation conditions, the constructed RCS data and HRRP data were processed and classified using the present invention, prior art 1, and prior art 2, respectively. The results are shown in Table 1.

[0105] Table 1 Comparison of identification results between the present invention and prior art.

[0106] Signal-to-noise ratio 5dB 10dB 15dB Prior Art 1 83.80％ 84.52 86.14％ Existing technology 2 81.46％ 85.45％ 88.04％ This invention 95.48％ 97.93％ 99.72％

[0107] In Table 1, Existing technology 1 is to build a deep convolutional neural network to identify real-valued HRRP data; Existing technology 2 is to use a sparse autoencoder to perform feature fusion recognition on real-valued RCS data and real-valued HRRP data.

[0108] As shown in Table 1, compared with the single-feature recognition of prior art 1, the target classification and recognition rate of the present invention can be improved by at least 11.68%, 13.41%, and 13.58% under different signal-to-noise ratios, respectively; compared with the multi-feature real-value data fusion recognition of prior art 2, the target classification and recognition rate of the present invention can be improved by at least 14.02%, 12.48%, and 11.68% under different signal-to-noise ratios, respectively. Simulation experiments show that the complex-valued dynamic fusion network-based recognition method proposed in this invention can fully utilize the complex-valued characteristics of radar wide and narrow band echoes, effectively improve the target recognition accuracy, and has high noise resistance and application value.

Claims

1. A spatial micro-movement target recognition method based on complex-valued dynamic fusion networks, characterized in that, include: (1) Generate dataset: 1a) Establish three-dimensional geometric electromagnetic models of six types of space micro-moving targets, and obtain their narrowband and broadband echoes respectively; 1b) Dynamically extract the original echo to obtain complex-valued RCS data and complex-valued HRRP data; 1c) The complex-valued RCS data of six types of spatial micro-moving targets are preprocessed and converted into two-dimensional complex-valued recursive graph (RP) data that can reflect the dynamic characteristics of the time series. (2) Divide the complex value RP data and complex value HRRP data of six types of spatial micro-motion targets to generate training set and test set; (3) Constructing a complex-valued dynamic fusion network: 3a) Establish a multimodal feature extraction module consisting of two convolutional neural networks connected in parallel, used to extract complex-valued RP features and complex-valued HRRP features; 3b) Establish a dynamic fusion module composed of a feature information dynamic estimation submodule and a modality information dynamic estimation submodule, which is used to dynamically fuse the features of the two modalities, complex-valued RP and complex-valued HRRP. Its implementation is as follows: The feature information dynamic estimation submodule first introduces a feature information encoder x m →w m Assess the information content of the two types of features, complex value RP and complex value HRRP. m And its application is determined by the loss function. The KL divergence is controlled, and then a sparsity strategy is used to determine the key features of different samples in the spatial micro-movement target fusion and recognition task. Finally, a sparse gating strategy is used to retain important information and suppress useless information, thereby enabling x to be... m Convert to The calculation formula is as follows: where x m For the feature sample, w m For its information content, d m F represents the dimension of the complex-valued feature. CV This represents a fully connected layer, where abs represents the modulus value, θ(·) represents the Tanh activation function, m represents the mode number, and M represents the number of mode types. This represents the average activation level of the i-th node in terms of feature information content. B represents the size of the current batch, and ρ represents a sparse parameter close to 0. This represents the feature sample after retaining important information, and ⊙ represents the Hadamard product; The modal information dynamic estimation submodule is used to... The input is fed into a fully connected layer and activated using the ReLU function to obtain the modal features h. m Then it is split into two outputs, one of which is input into the classifier f. m Get h m Corresponding correct class probability TCP m and the loss function of the classifier Another input is sent to the confidence network g. m Get TCP m The estimated value Loss function of confidence network Wherein, classifier f m It is composed of fully connected layers and a softmax activation function, y = [y1',...,y k ',…,y K '],y k ' represents the k-th element value of the category label y, with a predicted probability of . K represents the number of target categories. Let represent the posterior probability that a feature sample in the m-th modality is predicted to belong to the k-th class. TCP m ∈[0,1]; belief network g m TCP is composed of a fully connected layer and a Tanh activation function. m The estimated value The above different modal features h m The estimated value of the correct class probability The multiplication of different modalities is performed, and the products are concatenated to obtain the multimodal fusion feature h. Finally, the nonlinear representation capability of the multimodal fusion feature is enhanced by a complex fully connected layer F, resulting in the enhanced nonlinear multimodal fusion feature. [] indicates a splicing operation; 3c) Select the existing Softmax classifier to obtain the final multimodal classification result; 3d) The multimodal feature extraction module, dynamic fusion module, and classifier module are cascaded in sequence to form a complex value dynamic fusion network; (4) Input the training set data into the complex value dynamic fusion network and train it using the backpropagation algorithm to obtain the trained complex value dynamic fusion network. (5) Input the test set data into the trained complex value dynamic fusion network to obtain the spatial micro-movement target recognition result.

2. The method according to claim 1, characterized in that, Step 1a) Establish three-dimensional geometric and electromagnetic models of six types of spatially micro-moving targets, as follows: 1a1) Construct three-dimensional geometric models of six types of targets in Feko simulation software: flat-bottomed cone target (1), flat-bottomed cone cylinder target (2), spherical-bottomed cone target (3), spherical-bottomed cone cylinder target (4), frustum cone cylinder target (5), and grooved flat-bottomed cone target (6); 1a2) Initial parameter settings are performed in the Feko simulation software. After all settings are completed, the 3D model is subdivided and the static electromagnetic echo of each target in the full-angle domain is calculated.

3. The method according to claim 1, characterized in that, In step 1b), the original echo is dynamically extracted as follows: 1b1) Set the initial equivalent pitch angle, spin frequency, conic frequency, and precession angle for various targets, and then calculate the change curve of the equivalent pitch angle; 1b2) Based on the equivalent pitch angle change curve of the corresponding target, find the corresponding equivalent pitch angle data in the static electromagnetic simulation data, obtain the dynamic electromagnetic echo data, and add Gaussian white noise to it to obtain the required dynamic echo.

4. The method according to claim 1, characterized in that, Step 1c) involves preprocessing the complex RCS data of six types of spatially moving targets as follows: 1c1) For a real-valued RCS time series T=(t1,t2,…,t…) i ,…,t N Let the delay time be τ and the encoded feature dimension be l. Calculate the phase space matrix Y of the time series T: Where t i Let y be the i-th time in the real-valued time series. j Let be the j-th column of the phase space matrix, i∈[1,N], j∈[1,M], N is the length of the real-valued sequence, and M=N-(l-1)τ is the number of columns in the phase space matrix corresponding to the time series; 1c2) Calculate the recurrence graph corresponding to the time series T based on the phase space matrix Y: Where RP(p,q) represents the recursive graph calculated using the p-th and q-th columns of the phase space matrix Y, |||² represents L2 normalization, p,q∈[1,M], ε is the threshold for determining the state of RP, and y p and y q These represent the p-th and q-th columns in the phase space matrix Y, respectively. 1c3) Considering that an inappropriate ε might lead to the loss of detailed information in the time series, ε is ignored to ensure the integrity of the information, resulting in the modified RP(p,q): RP(p,q)=||y p -y q ||2 1c4) For complex-valued RCS time series The complex-valued recursive graph CRP is calculated based on the recursive graph RP(p,q) to preserve the phase information of the complex-valued time series. The calculation formula is as follows: in, and They represent the real and imaginary parts of a complex number, respectively.

5. The method according to claim 1, wherein In step (2), the complex RP data and complex HRRP data of the six types of space micro-motion targets are divided according to the initial equivalent pitch angle, as follows: 2a) Select 1210 samples and their corresponding labels corresponding to the initial equivalent pitch angle of each target class between 21° and 30° to form a training set; 2b) Select 605 samples and corresponding labels corresponding to the initial equivalent pitch angle of each target type between 31° and 35° to form a test set.

6. The method according to claim 1, characterized in that, Step 3a) Construct two convolutional neural networks in the multimodal feature extraction module, with the following structure and parameters: The structure of each convolutional neural network is as follows: 1st complex convolutional layer → 1st complex pooling layer → 2nd complex convolutional layer → 2nd complex pooling layer → 3rd complex convolutional layer → 3rd complex pooling layer → 4th complex convolutional layer → 4th complex pooling layer → complex fully connected layer; The two convolutional neural networks have different parameters, where: The parameters of the first convolutional neural network are: The first complex convolutional layer has a kernel size of 5×5, 4 channels, and a stride of 1×1. The pooling kernel size of the first complex pooling layer is 2×2, and the step size is 2×2. The second complex convolutional layer has a kernel size of 5×5, 4 channels, and a stride of 1×1. The pooling core size of the second complex pooling layer is 2×2, and the step size is 2×2. The third complex convolutional layer has a kernel size of 3×3, 8 channels, and a stride of 1×1. The pooling kernel size of the third complex pooling layer is 2×2, and the step size is 2×2. The fourth complex convolutional layer has a kernel size of 3×3, 8 channels, and a stride of 1×1. The pooling kernel size of the fourth complex pooling layer is 2×2, and the step size is 2×2. Fully connected layers are used to map the feature dimensions to 128; The parameters of the second convolutional neural network differ from those of the first convolutional neural network in that: The second complex convolutional layer has 8 channels; The third complex convolutional layer has 16 channels; The fourth complex convolutional layer has 32 channels and a stride of 2×2.

7. The method according to claim 1, characterized in that, In step 3c), the Softmax classifier is selected to obtain the final multimodal classification result, as follows: Multimodal fusion features Mapped to a K-dimensional rating vector [s1,...,s i ,...,s K The Softmax function is used to transform the rating vector into a probability vector [p1,...,p]. i ,...,p K The index of the element with the maximum value in the probability vector is the final multimodal classification result C. C=argmax(p1,…,p i ,…,p K ) in s represents the probability that the input sample is identified as the i-th target category. i Let represent the score of the network model on the i-th target category, i∈[1,K], where K represents the number of target categories. The e-index is used to transform the score values ​​in (-∞,+∞) to (0,+∞) without affecting the relative magnitudes. Used for normalization.

8. The method according to claim 1, characterized in that, In step (4), the complex-valued dynamic fusion network is trained using the backpropagation algorithm, as follows: 4a) Initialize network parameters; 4b) Set the network loss function in, Let y be the loss function of the Softmax classifier. k p represents the value of the k-th element of the category label. k This represents the posterior probability that the predicted multimodal fusion feature belongs to the k-th class. The loss function represents the dynamic estimation submodule for feature information. Let λ1 and λ2 represent the loss function of the modal information dynamic estimation submodule, respectively. The hyperparameters of the weights; 4c) Using the loss function Calculate the error between the network output value and the target value, and use the SGD optimizer to update the network parameters between each layer; 4d) Repeat steps 4a) to 4c) until the error between the network output value and the target value is basically stable, then stop training and obtain the trained complex value dynamic fusion network.

Images