Space micro-motion target identification method based on complex value dynamic fusion network

Abstract

The invention discloses a spatial micro-motion target identification method based on a complex value dynamic fusion network, and mainly solves the problem that the existing spatial micro-motion target identification method cannot make full use of modal information of a target, so that the identification rate is low. According to the scheme, the method comprises the following steps: generating complex-valued RCS data and complex-valued HRRP data; preprocessing the complex value RCS data to obtain two-dimensional complex value recurrence plot RP data; dividing the RP data and the HRRP data to generate a training set and a test set; constructing a complex value dynamic fusion network comprising a multi-modal feature extraction module, a dynamic fusion module and a classifier; inputting training set data into the complex value dynamic fusion network, and training the training set data through a back propagation algorithm; and inputting the test set data into the trained complex value dynamic fusion network, and obtaining the final recognition accuracy through the forward propagation of the network. The method can make full use of the complex value characteristics of radar wide and narrow band echoes, effectively improves the target recognition accuracy and anti-noise performance, and can be used for aerospace target tracking.

Description

Technical Field

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

[0002] Micro-motion refers to the vibration or rotation of the target or target component relative to the radar line of sight in addition to the main movement, such as spin, precession, tumble, etc. When the radar uses wide and narrow band alternating modes to observe the space micro-motion target, the narrowband and broadband echoes of the target can be obtained at the same time. Among them, the narrowband radar cross-section area RCS sequence can reflect the overall scattering intensity of the target over time; the broadband high-resolution range image HRRP can characterize the structural characteristics and size information of the target. Therefore, RCS and HRRP contain complementary target separable features. If they can be fully explored and integrated, the recognition accuracy of space micro-motion targets can be effectively improved.

[0003] However, most of the existing technologies only use RCS or HRRP for single-modal feature extraction and recognition, and most of them rely on real-valued neural networks for feature extraction, abandoning the rich features contained in the phase of RCS sequence and HRRP sequence, resulting in limited performance of spatial micro-motion target recognition methods.

[0004] Yizhe Wang et al. built a deep convolutional neural network DCNN in their paper "Space Precession Target Classification Based on Radar High-Resolution Range Profiles" to extract and identify features of HRRP data. The specific implementation steps are: directly input HRRP data into the DCNN composed of convolutional layers, pooling layers and fully connected layers, automatically mine features useful for HRRP data classification through training, and finally output the classification results through the Softmax classifier. Although this method has greatly improved the recognition accuracy compared to traditional feature extraction and recognition methods, it only uses single modal features and only uses the amplitude information of the HRRP sequence, resulting in poor overall recognition performance.

[0005] Dong J et al. proposed a radar space target recognition method based on high-dimensional feature maps in their paper "HRPnet: High-Dimensional Feature Mapping for Radar Space Target Recognition". The specific implementation steps are as follows: first, a sparse autoencoder is used to extract deep generalized features from the RCS sequence and HRRP sequence respectively, and then the three features are encoded into Gram angular fields and spliced ​​into multi-channel high-dimensional features, and finally sent to the Softmax classifier to output the predicted label. Although this method improves the recognition performance compared to the single feature condition, it does not consider the mining and effective fusion of separable features of each modal data when the separability gap between the modal data is large, and only uses the amplitude information of the RCS and HRRP sequences, ignoring the separable features contained in their phases, resulting in poor target recognition performance.

[0006] The Chinese patent document with publication number CN 117289231 A discloses a radar target recognition method based on sub-band integration and multi-feature fusion mechanism, which first generates ultra-wideband data by coherent fusion of sub-band data, and then extracts time domain HRRP features and frequency domain geometric diffraction theory GTD features, and finally inputs the two features into a multi-feature fusion recognition network composed of CNN Block and TransformerBlock for training and recognition. Although this method fuses HRRP features with GTD features for training, the fusion recognition network built is a real network, which cannot be trained on complex data, and the fusion recognition network is non-dynamic. In terms of multi-modality, the amount of information of the same sample in different modes is different; in terms of the same sample, when the confidence of different modes is greatly different, the information contained in the extracted features may be contradictory. At this time, directly fusing and recognizing multi-modal features may lead to sample misclassification and damage the recognition performance of spatial micro-motion targets. Summary of the invention

[0007] The purpose of the present invention is to propose a spatial micro-motion target recognition method based on a complex-valued dynamic fusion network to solve the problems of single feature extraction, limited recognition performance, and insufficient use of echo information caused by ignoring phase information in existing methods, thereby improving the recognition performance of micro-motion targets.

[0008] The technical idea for achieving the purpose of the present invention is: through dynamic estimation of the feature information amount and the modal information amount, the multimodal information is reasonably and fully integrated to enhance the interpretability of multimodal fusion; at the same time, attention is paid to the data amplitude information and phase information, and the echo information is fully utilized to improve the micro-motion target recognition performance.

[0009] According to the above ideas, the technical solutions of the present invention include the following:

[0010] (1) Generate a dataset:

[0011] 1a) Establish three-dimensional geometric electromagnetic models of six types of spatial micro-motion 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) Preprocess the complex-valued RCS data of six types of space micro-motion targets and convert them into two-dimensional complex-valued recursive graph RP data that can reflect the dynamic characteristics of time series;

[0014] (2) The complex-valued RP data and complex-valued HRRP data of six types of spatial micro-motion targets are divided to generate training sets and test sets;

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

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

[0017] 3b) establishing a dynamic fusion module composed of a feature information dynamic estimation submodule and a modal information dynamic estimation submodule, for dynamically fusing the features of the two modalities, complex-valued RP and complex-valued HRRP;

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

[0019] 3d) cascading the multimodal feature extraction module, the dynamic fusion module and the classifier module in sequence to form a complex-valued dynamic fusion network;

[0020] (4) Input the training set data into the complex-valued dynamic fusion network and train it through the back-propagation algorithm to obtain the trained complex-valued dynamic fusion network.

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

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

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

[0024] Second, the present invention constructs a dynamic fusion module composed of a feature information dynamic estimation submodule and a modal information dynamic estimation submodule, which can dynamically fuse the features of the two modes, complex-valued RP and complex-valued HRRP. Compared with the prior art that only uses a single modal data for identification, the present invention achieves full and reasonable use of multi-modal data of micro-motion targets. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 It is a flow chart for realizing the present invention;

[0026] Figure 2 A multimodal feature extraction network structure constructed in the present invention;

[0027] Figure 3 This is the overall structure diagram of the complex-valued dynamic fusion network constructed in the present invention. DETAILED DESCRIPTION

[0028] In order to enable those skilled in the art to better understand the scheme of the present invention, the technical scheme in the embodiment of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiment of the present invention. Obviously, the described embodiment is only a part of the embodiment of the present invention, not all of the embodiments. Based on the embodiments of the present invention, other embodiments obtained by ordinary technicians in this field without creative work should all fall within the scope of protection of the present invention.

[0029] It should be noted that the step numbers in the specification and claims of the present invention are only for a clear description of the implementation scheme of the present invention to facilitate understanding, and the order of the step numbers is not limited.

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

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

[0032] 1.1) Construct three-dimensional geometric models of six types of targets, namely, flat-bottomed cone target, flat-bottomed cone-cylinder target, spherical-bottomed cone target, spherical-bottomed cone-cylinder target, frustum cone-cylinder target and grooved flat-bottomed cone target in Feko simulation software, and use the standard PO method to obtain full-angle static electromagnetic echoes of each type of target;

[0033] 1.2) According to the target micro-motion form, extract its full-angle static electromagnetic echo to generate dynamic electromagnetic echo for each type of target: set the radar carrier frequency to 10GHz, the narrowband signal bandwidth to 20MHz, the broadband signal bandwidth to 2GHz, and the operating frequency to X-band. The pulse repetition frequency of broadband and narrowband radars is 50Hz, and the observation time is 2s. The initial equivalent pitch angle of the target is set to change continuously at intervals of 1° within 21°~35°. The precession frequency and precession angle at each pitch angle are equally spaced to 11. For each type of target, 1815 broadband echo data and 1815 narrowband echo complex valued RCS data are generated respectively;

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

[0035] 1.4) In order to simulate the echo data in the actual scene as much as possible and verify the robustness of the proposed algorithm to noise, Gaussian noise is added in the range of 5dB-15dB at intervals of 5dB to obtain complex-valued HRRP data and complex-valued RCS data under different signal-to-noise ratios.

[0036] Step 2: preprocess the complex-valued RCS data to obtain complex-valued RP data.

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

[0038]

[0039] where t i is the element at the i-th moment in the complex-valued RCS time series, y j is the jth 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 phase space matrix columns corresponding to the time series, τ is the delay time, and l is the encoding feature dimension;

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

[0041]

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

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

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

[0045] 3.1) Select 1210 samples and corresponding labels corresponding to the initial equivalent pitch angle 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 angle 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 in parallel:

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

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

[0052] in:

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

[0054] The convolution kernel size of the first complex convolution layer is 5×5, the number of channels is 4, and the stride is 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 convolution kernel size of the second complex convolution layer is 5×5, the number of channels is 4, and the stride is 1×1;

[0057] The pooling kernel size of the second complex pooling layer is 2×2 and the step size is 2×2;

[0058] The convolution kernel size of the third complex convolution layer is 3×3, the number of channels is 8, and the stride is 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 convolution kernel size of the 4th complex convolution layer is 3×3, the number of channels is 8, and the stride is 1×1;

[0061] The pooling kernel size of the 4th complex pooling layer is 2×2 and the step size is 2×2;

[0062] A complex fully connected layer is used to map the feature dimension to 128.

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

[0064] The number of channels of the second complex convolutional layer is 8;

[0065] The number of channels of the third complex convolutional layer is 16;

[0066] The number of channels of the 4th complex convolutional layer is 32, and the stride is 2×2.

[0067] 4.2) Construct a dynamic fusion module composed of a feature information dynamic estimation submodule and a modal information dynamic estimation submodule, which is used to estimate the feature and modal information of the extracted complex-valued features, and then perform reasonable dynamic fusion. That is, the complex-valued features are first input into the feature information dynamic estimation submodule to obtain feature samples after retaining important information, and then input into the modal information dynamic estimation submodule to obtain the classification confidence of each modality. Finally, the feature samples are multiplied by the classification confidence and spliced ​​to achieve fusion.

[0068] The specific implementation is as follows:

[0069] 4.2.1) The feature information dynamic estimation submodule first introduces the feature information encoder x m →w m Evaluate the information content of two types of features, complex-valued RP and complex-valued HRRP m And apply the loss function The KL divergence of the control is then determined through the sparsity strategy to determine the key features of different samples in the spatial micro-motion target fusion recognition task. Finally, the sparse gating strategy is used to retain important information and suppress useless information, thereby converting x m Convert to The calculation formula is as follows:

[0070]

[0071] where x m is the feature sample, w m For its information volume, d m represents the complex-valued feature dimension, F CVrepresents a complex fully connected layer, abs represents the modulus value, θ(·) represents the Tanh activation function, m represents the modality number, M represents the number of modality types, Represents the average activation degree of the i-th node of the feature information, B represents the size of the current batch, ρ represents a sparse parameter close to 0, represents the feature sample after retaining important information, ⊙ represents the Hadamard product;

[0072] 4.2.2) The modal information dynamic estimation submodule will Input to the complex fully connected layer and Function activation obtains modal feature h m , and then divided into two outputs, one of which is input into the classifier f m Get h m The corresponding correct class probability TCP m And the loss function of the classifier The other input is to the belief network g m Get TCP m Estimated value of TCP m And the loss function of the belief network

[0073]

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

[0075]

[0076] Among them, the classifier f m It is composed of a complex fully connected layer and a Softmax activation function, y = [y1,...,y k ,...,y K ],y k Represents the kth element value of the category label y, and the predicted probability is K represents the number of target categories, represents the posterior probability that the feature sample under the mth mode belongs to the kth class, TCP m ∈[0,1]; belief network g m It is composed of a complex fully connected layer and a Tanh activation function. TCP m Estimated value of TCP 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 different modal features h m and the estimated value of the probability of the correct class TCP m Multiply them together, and concatenate the products of different modes to obtain the multimodal fusion feature h. Finally, the multimodal fusion feature is nonlinearly enhanced through the complex fully connected layer F to obtain the enhanced nonlinear multimodal fusion feature

[0079]

[0080] Where [] represents the concatenation operation;

[0081] 4.3) Select the existing Softmax classifier module to classify the above enhanced nonlinear multimodal fusion features Classification is performed to obtain the final classification result of multimodal features. The specific implementation is as follows:

[0082] Fusion of multimodal features Mapped to a K-dimensional score value vector [s1,...,s i ,...,s K ], and transform the score value vector into a probability vector [p1,...,p i ,...,p K ], the index of the maximum value element in the probability vector is the final multimodal classification result C:

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

[0084] in represents the probability that the input sample is identified as the i-th target category, s i It represents the score value of the network model on the i-th target category, i∈[1,K], K represents the number of target categories, and the role of the e index is to change the score value of (-∞,+∞) to (0,+∞) without affecting the relative size relationship. Used for normalization;

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

[0086] Step five: train the complex-valued dynamic fusion network.

[0087] 5.1) Initialize network parameters;

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

[0089] 5.2) Set the network loss function

[0090]

[0091] in, is the loss function of the Softmax classifier, y k Represents the kth element value of the category label, p k represents the posterior probability that the prediction of the multimodal fusion feature belongs to the kth class, represents the loss function of the feature information dynamic estimation submodule, represents the loss function of the modal information dynamic estimation submodule, λ1 and λ2 are respectively the balance Hyperparameters of weights;

[0092] 5.3) Using 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 training iteration number is reached, and then stop training to obtain a trained complex-valued dynamic fusion network.

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

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

[0096] The effect of the present invention can be further illustrated by the following simulation experiment:

[0097] 1. Simulation Experiment Conditions

[0098] The simulation experiment of the present invention uses the standard PO method to generate wide-band and narrow-band radar echo data of six types of space micro-motion targets, and the 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 10 GHz, the narrowband signal bandwidth is set to 20 MHz, the broadband signal bandwidth is set to 2 GHz, and the operating frequency is set to X-band; the pulse repetition frequency of both broadband and narrowband radars is 50 Hz, and the observation time is 2 s.

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

[0101] Gaussian white noise is added to the echo, and the complex-valued RCS data is preprocessed to obtain complex-valued RP data, generating echo samples with signal-to-noise ratios of 5dB, 10dB, and 15dB, respectively. The samples corresponding to the initial pitch angle of 21° to 30° are taken as the training set, and 7260 training samples can be obtained; the samples corresponding to the initial pitch angle of 31° to 35° are taken as the test set, and 3630 test samples can be obtained.

[0102] The simulation experiment hardware platform is Intel(R)Core(TM)i7-10700@2.90GHzCPU, 16GB RAM; the simulation experiment software platform is FEKO 2017, MATLAB 2021b, and Python3.8.

[0103] 2. Simulation content and results analysis

[0104] Under the above simulation conditions, the present invention, prior art 1 and prior art 2 are used to process and classify the RCS data and HRRP data constructed above, respectively. The results are shown in Table 1:

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

[0106] Signal-to-Noise Ratio 5dB 10dB 15dB Prior art 1 83.80％ 84.52 86.14％ Prior Art 2 81.46％ 85.45％ 88.04％ The present invention 95.48％ 97.93％ 99.72％

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

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

Claims

1. A spatial micro-motion target recognition method based on a complex-valued dynamic fusion network, characterized in that: include: (1) Generate a dataset: 1a) Establish three-dimensional geometric electromagnetic models of six types of spatial micro-motion 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) Preprocess the complex-valued RCS data of six types of space micro-motion targets and convert them into two-dimensional complex-valued recursive graph RP data that can reflect the dynamic characteristics of time series; (2) The complex-valued RP data and complex-valued HRRP data of six types of spatial micro-motion targets are divided to generate training sets and test sets; (3) Constructing a complex-valued dynamic fusion network: 3a) establishing a multimodal feature extraction module consisting of two convolutional neural networks in parallel to extract complex-valued RP features and complex-valued HRRP features; 3b) establishing a dynamic fusion module composed of a feature information dynamic estimation submodule and a modal information dynamic estimation submodule, for dynamically fusing the features of the two modalities, complex-valued RP and complex-valued HRRP; 3c) Selecting the existing Softmax classifier to obtain the final multimodal classification result; 3d) cascading the multimodal feature extraction module, the dynamic fusion module and the classifier module in sequence to form a complex-valued dynamic fusion network; (4) Input the training set data into the complex-valued dynamic fusion network and train it through the back-propagation algorithm to obtain the trained complex-valued dynamic fusion network. (5) Input the test set data into the trained complex-valued dynamic fusion network to obtain the spatial micro-motion target recognition results.

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

3. The method according to claim 1, characterized in that: In step 1b), the original echo is dynamically extracted as follows: 1b1) Setting the initial equivalent pitch angle, spin frequency, cone frequency, and precession angle of various targets, and then calculating the change curve of the equivalent pitch angle; 1b2) According to the equivalent pitch angle change curve of the corresponding target, the data of the corresponding equivalent pitch angle is found in the static electromagnetic simulation data, the dynamic electromagnetic echo data is obtained, and Gaussian white noise is added to it to obtain the required dynamic echo.

4. The method according to claim 1, characterized in that In step 1c), the complex RCS data of the six types of space micro-motion targets are preprocessed as follows: 1c1) For the real-valued RCS time series T = (t1, t2, ..., t i ,…,t N ), denote the delay time as τ, the encoding feature dimension as l, and calculate the phase space matrix Y of the time series T: where t i is the i-th moment in the real-valued time series, y j is 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 phase space matrix columns corresponding to the time series; 1c2) Calculate the recurrence graph corresponding to the time series T according to the phase space matrix Y: Where RP(p,q) represents the recurrence graph calculated by 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 RP state, and y p and q denote the p-th column and the q-th column in the phase space matrix Y respectively; 1c3) According to the situation that inappropriate ε may lead to the loss of time series detail information, ε is ignored to ensure the integrity of information, and the modified RP(p,q) is obtained: RP(p,q)=||y p -y q ||2 1c4) For complex-valued RCS time series The complex-valued recursion graph CRP is calculated based on the recursion graph RP(p,q) to retain the phase information of the complex-valued time series. The calculation formula is: in, and Represent the real and imaginary parts of a complex number respectively.

5. The method according to claim 1, characterized in that In step (2), the complex-valued RP data and complex-valued HRRP data of the six types of space micro-motion targets are divided according to the initial equivalent pitch angle, which is implemented as follows: 2a) Select 1210 samples and corresponding labels corresponding to the initial equivalent pitch angle of each type of target 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 type of target between 31° and 35° to form a test set.

6. The method according to claim 1, characterized in that Step 3a) constructs two convolutional neural networks in the multimodal feature extraction module, and their structures and parameters are as follows: The structure of each convolutional neural network is: 1st complex convolution layer → 1st complex pooling layer → 2nd complex convolution layer → 2nd complex pooling layer → 3rd complex convolution layer → 3rd complex pooling layer → 4th complex convolution layer → 4th complex pooling layer → complex fully connected layer; The parameters of the two convolutional neural networks are different, where: The parameters of the first convolutional neural network are: The convolution kernel size of the first complex convolution layer is 5×5, the number of channels is 4, and the stride is 1×1; The pooling kernel size of the first complex pooling layer is 2×2 and the step size is 2×2; The convolution kernel size of the second complex convolution layer is 5×5, the number of channels is 4, and the stride is 1×1; The pooling kernel size of the second complex pooling layer is 2×2 and the step size is 2×2; The convolution kernel size of the third complex convolution layer is 3×3, the number of channels is 8, and the stride is 1×1; The pooling kernel size of the third complex pooling layer is 2×2 and the step size is 2×2; The convolution kernel size of the 4th complex convolution layer is 3×3, the number of channels is 8, and the stride is 1×1; The pooling kernel size of the 4th complex pooling layer is 2×2 and the step size is 2×2; A complex fully connected layer is used to map the feature dimension to 128. The parameters of the second convolutional neural network differ from the first convolutional neural network in that: The number of channels of the second complex convolutional layer is 8; The number of channels of the third complex convolutional layer is 16; The number of channels of the 4th complex convolutional layer is 32, and the stride is 2×2.

7. The method according to claim 1, characterized in that In step 3b), the dynamic fusion module dynamically fuses the features of the two modes, complex-valued RP and complex-valued HRRP, as follows: The feature information dynamic estimation submodule first introduces a feature information encoder x m →w m Evaluate the information content of two types of features, complex-valued RP and complex-valued HRRP m And apply the loss function The KL divergence of the control is then determined through the sparsity strategy to determine the key features of different samples in the spatial micro-motion target fusion recognition task. Finally, the sparse gating strategy is used to retain important information and suppress useless information, thereby converting x m Convert to The calculation formula is as follows: where x m is the feature sample, w m For its information volume, d m represents the complex-valued feature dimension, F CV represents a complex fully connected layer, abs represents the modulus value, θ(·) represents the Tanh activation function, m represents the modality number, M represents the number of modality types, Represents the average activation degree of the i-th node of the feature information, B represents the size of the current batch, ρ represents a sparse parameter close to 0, represents the feature sample after retaining important information, ⊙ represents the Hadamard product; The modal information dynamic estimation submodule is used to Input to the complex fully connected layer and Function activation obtains modal feature h m , and then divided into two outputs, one of which is input into the classifier f m Get h m The corresponding correct class probability TCP m And the loss function of the classifier The other input is to the belief network g m Get TCP m Estimated value of TCP m And the loss function of the belief network TCP m =g m (h m ) Among them, the classifier f m It is composed of a complex fully connected layer and a Softmax activation function, y = [y1,...,y k ,...,y K ],y k Represents the kth element value of the category label y, and the predicted probability is K represents the number of target categories, represents the posterior probability that the feature sample under the mth mode belongs to the kth class, TCP m ∈[0,1]; belief network g m It is composed of a complex fully connected layer and a Tanh activation function. TCP m Estimated value of TCP m ∈[0, 1]; The above different modal features h m and the estimated value of the probability of the correct class TCP m Multiply them together, and concatenate the products of different modes to obtain the multimodal fusion feature h. Finally, the multimodal fusion feature is nonlinearly enhanced through the complex fully connected layer F to obtain the enhanced nonlinear multimodal fusion feature h=[TCP 1 h 1 ,...,TCP M h M ] Where [] represents a concatenation operation.

8. The method according to claim 1, characterized in that: In step 3c), the Softmax classifier is used to obtain the final multimodal classification result, which is implemented as follows: Fusion of multimodal features Mapped to a K-dimensional score vector [s1,...,s i ,...,s K ], and transform the score value vector into a probability vector [p1,...,p i ,...,p K ], the index of the maximum value element in the probability vector is the final multimodal classification result C: C=argmax(p1,...,p i ,...,p K ) in represents the probability that the input sample is identified as the i-th target category, s i It represents the score value of the network model on the i-th target category, i∈[1,K], K represents the number of target categories, and the role of the e index is to change the score value of (-∞,+∞) to (0,+∞) without affecting the relative size relationship. Used for normalization.

9. The method according to claim 1, characterized in that: In step (4), the complex-valued dynamic fusion network is trained by the back-propagation algorithm, which is implemented as follows: 4a) Initialize network parameters; 4b) Set the network’s loss function in, is the loss function of the Softmax classifier, y k Represents the kth element value of the category label, p k represents the posterior probability that the prediction of the multimodal fusion feature belongs to the kth class, represents the loss function of the feature information dynamic estimation submodule, represents the loss function of the modal information dynamic estimation submodule, λ1 and λ2 are respectively the balance Hyperparameters of weights; 4c) Using 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, and then stop training to obtain a trained complex-valued dynamic fusion network.

Images