Microwave photon radar strong maneuvering target identification method based on small sample incremental learning

Abstract

The invention discloses a microwave photon radar strong maneuvering target identification method based on small sample incremental learning, and belongs to the field of microwave photon radar maneuvering target identification. Time-frequency dynamic, polarization space and geometric topological features are extracted through a deep learning model, and multi-dimensional joint characterization is realized in combination with a cross-domain attention mechanism; designing a polarization-time frequency adversarial generative network, generating synthetic data conforming to an electromagnetic scattering rule under the physical constraint of a Maxwell equation, and constructing a prototype feature support set through meta learning; a dynamic knowledge distillation framework is introduced, physical consistency loss is utilized to align motion parameters and classification probability distribution of the teacher model and the student model, and lightweight increment updating is realized in combination with gradient importance masks. According to the method, the recognition precision and real-time performance of the strong maneuvering target in a complex electromagnetic environment are remarkably improved, and the method is suitable for high and overspeed maneuvering target recognition.

Description

Technical Field

[0001] The present invention relates to the field of microwave photon radar maneuvering target recognition, and in particular to a method for strongly maneuvering target recognition based on small sample incremental learning. Background Art

[0002] Traditional radar systems, limited by the bandwidth and signal processing capabilities of electronic devices, face core challenges in identifying highly maneuvering targets, including insufficient dynamic feature capture and difficulty fusing multi-domain information. The motion characteristics of highly maneuvering targets lead to dramatic fluctuations in their time-frequency energy distribution, rapid switching of polarization scattering characteristics, and nonlinear distortion of geometric profile topology. Although microwave photon radars have overcome bandwidth limitations through optical domain signal processing, their target recognition still relies on static feature libraries, making them difficult to adapt to dynamic battlefield environments. Existing methods, such as high-resolution Doppler image-based maneuver detection algorithms or recurrent neural networks (RNNs) for turning maneuver recognition, can extract single-dimensional features, but these methods suffer from weak cross-domain correlation and fail to consider the physical consistency constraints of polarization-geometry features, resulting in increased false alarm rates.

[0003] Radar target recognition faces the dual challenges of sample scarcity and model update lag. Existing small-sample learning methods such as prototypical networks and generative adversarial networks (CycleGANs) can alleviate the data shortage problem, but the generated synthetic data lacks physical interpretability, resulting in poor model generalization in real-world scenarios. Regarding incremental learning, traditional dynamically scalable networks or end-to-end incremental learning can gradually expand model capacity, but are prone to catastrophic forgetting, which can lead to a sharp decline in recognition performance for old targets. Therefore, designing a method suitable for microwave photon radar to identify highly maneuverable targets has become a hot research topic. Summary of the Invention

[0004] The purpose of the present invention is to overcome the above-mentioned defects in the background technology and propose a microwave photon strong maneuvering target recognition method based on small sample incremental learning, which improves the recognition accuracy and real-time performance of strong maneuvering targets in complex electromagnetic environments, and ultimately realizes the rapid recognition of strong maneuvering targets using microwave photon radar.

[0005] The present invention adopts the following technical solutions to solve the above technical problems:

[0006] The microwave photon radar strong maneuvering target recognition method based on small sample incremental learning includes the following steps:

[0007] Step S1: Input the target data acquired by the microwave photon radar into the multi-domain feature fusion module, extract the time-frequency energy dynamic characteristics through the long short-term memory network, mine the polarization scattering spatial pattern through the convolutional neural network, and model the geometric contour topological relationship through the graph neural network. The time-frequency energy dynamic characteristics, polarization scattering spatial pattern and geometric contour topological relationship are dynamically fused through the cross-domain attention mechanism to output multi-dimensional joint representation features;

[0008] Step S2: In the small sample data augmentation module, based on the multi-dimensional joint representation features obtained in step S1, a polarization-time-frequency adversarial generative network (PT-GAN) is designed. Synthetic data that conforms to the laws of electromagnetic scattering is generated in combination with the physical constraints of Maxwell's equations. Using this synthetic data and real data, a prototype feature support set is constructed through meta-learning.

[0009] Step S3: In the physical constraint incremental learning model, the prototype feature support set constructed in step S2 is used, and a dynamic knowledge distillation framework is adopted to align the motion parameters and classification probability distribution of the teacher model and the student model through physical consistency loss. The model training update is achieved by combining the gradient importance screening mechanism based on parameter sensitivity.

[0010] Furthermore, the specific extraction process of extracting the time-frequency energy dynamic characteristics through the long short-term memory network in step S1 is as follows:

[0011] The time-frequency graph generated by the Wigner-Ville distribution is used as input, and a bidirectional long short-term memory network is used to capture the temporal correlation of the target motion. The hidden state update process of the bidirectional long short-term memory network can be expressed as:

[0012]

[0013] where x t is the input of LSTM network, f t 、i t 、o t Represent the outputs of the forget gate, input gate, and output gate respectively, is the candidate cell state, C t 、C t-1 is the cell state at the current and previous time steps, h t 、h t-1 is the hidden layer state of the current and previous time steps, σ is the Sigmoid function, W f 、W i 、W c 、W o and b f 、b i 、b C 、b o are learnable parameters.

[0014] Furthermore, the step S1 of mining polarization scattering spatial patterns through convolutional neural networks is specifically as follows:

[0015] Based on the polarization scattering matrix decomposed by Pauli, a multi-scale convolutional neural network, namely CNN, is constructed to finally mine the polarization scattering spatial pattern and calculate the polarization entropy feature H P Quantify the randomness of target scattering, H P Calculated by the following formula:

[0016]

[0017] where p i represents the proportion of scattered energy of polarization channel i, and N is the total number of polarization channels;

[0018] By introducing reciprocity theorem constraints in the convolution kernel, the network is forced to learn polarization patterns that conform to electromagnetic laws.

[0019] Furthermore, the step S1 of constructing the geometric contour topological relationship through graph neural network modeling is specifically as follows:

[0020] The contour point cloud extracted from the range-Doppler image is modeled as a graph structure. The equivariant graph neural network is used to extract the rotation and translation invariant geometric features and construct the geometric contour topological relationship. The graph convolution process of the convolution layer is defined as:

[0021]

[0022] in represents the feature of node n at layer l+1, represents the feature of node n in layer l, represents the feature of node m in layer l, is the set of adjacent nodes, φ and ψ are node highlighting function and edge aggregation function respectively, e nm The feature vector representing the edge (n,m) is used to encode the geometric and semantic relationship between nodes n and m.

[0023] Furthermore, the dynamic fusion cross-domain attention mechanism is specifically calculated as follows:

[0024] F fusion =α·F tf +β·F p +γ·F g

[0025] Among them F tf ,F p ,F gare the time-frequency, polarization, and geometric feature vectors, respectively. The weight coefficients α, β, and γ are calculated by the multi-head self-attention mechanism:

[0026]

[0027] where Q tf and K tf is the query and key vector of time-frequency features, Q p and K p is the query and key vector of the polarization feature, Q g and K g are the query and key vectors of geometric features, and d is the feature dimension.

[0028] Furthermore, the specific process of step S2 is as follows:

[0029] Step S2-1: Generate PT-GAN based on the multi-dimensional joint representation features obtained in step S1, wherein the PT-GAN includes a generator and a discriminator, wherein the generator of the PT-GAN is a noise vector. and a small number of real samples X real As input, a multi-layer deconvolution network is used to generate time-frequency-polarization joint data X = G(z, X real ), combined with the physical constraints of Maxwell's equations to generate synthetic data that conforms to the laws of electromagnetic scattering;

[0030] The loss function of the generator is composed of adversarial loss and physical constraints Joint composition:

[0031]

[0032] in To trade off the coefficient, adversarial loss D(·) is the output of the discriminator, G(z) is the output of the generator in PT-GAN, and the physical constraint term This is achieved by regularizing the differential form of Maxwell's equations, specifically:

[0033]

[0034] Physical constraints The electromagnetic field distribution E, B of the generated data is forced to satisfy the relationship between curl and divergence, Ω is the electric displacement vector, and ρ is the charge density;

[0035] The discriminator adopts the gradient penalty strategy of Wasserstein GAN to generate adversarial networks, and its loss function is defined as:

[0036]

[0037] Where X is the interpolation sample, and X=∈X real +(1-)X,∈~U(0,1),λ2 controls the intensity of gradient penalty;

[0038] Step S2-2: Using the generated synthetic data and real data, a prototype feature support set is constructed through meta-learning to enhance small sample adaptability, and prototype features are extracted from the time-frequency-polarization joint data X:

[0039]

[0040] Among them, f enc is the pre-trained encoder, S c is the support set sample of category c;

[0041] The prototype network minimizes the query set sample x q With prototype P c The Euclidean distance is used to achieve classification:

[0042]

[0043] Step S2-3: Introduce the dynamic measurement function g of the relationship network φ (f enc (x q ),P c ), the relational network adopts a two-layer fully connected architecture, takes the sample features output by the encoder and the category prototype extracted in step S2-3 as input, and outputs the confidence that the sample belongs to a specific category. Its loss function is expanded to:

[0044]

[0045] g φ It is implemented by a two-layer fully connected network, and the nonlinear boundary between categories is learned through end-to-end training. q Represents the minimized query set sample x q Corresponding label, (x q ,y q ) represents the query set sample-label pair.

[0046] Furthermore, the specific process of step S3 is as follows:

[0047] Step S3-1, using the synthetic data and prototype feature support set generated in step S2, initialize the teacher model and the student model, where the teacher model is based on the initial target recognition model and the student model is the target recognition model to be updated;

[0048] Step S3-2: Through the dual feature and semantic alignment mechanism of the teacher-student architecture, the teacher model parameters are fixed, and the student model is trained using synthetic data and the prototype feature support set. Parameter-level knowledge transfer is achieved through feature space constraints. At the same time, the KL divergence loss is introduced in the output layer to constrain the classification probability distribution, so that the student model inherits the classification decision boundary of the teacher model;

[0049] Step S3-3: Embed the hypersonic vehicle motion equations as physical constraints into the knowledge distillation process, and construct a physical consistency loss function. Continue training the student model trained in step S3-2 to align the motion parameters and classification probability distributions of the teacher model and the student model.

[0050] Step S3-4: Based on the training results of steps S3-2 and S3-3, a gradient importance screening mechanism based on parameter sensitivity is designed to fine-tune only the highly sensitive parameters to achieve lightweight incremental updates of the target recognition model.

[0051] Furthermore, the teacher model f in the teacher-student architecture T The intermediate feature h T With student model f S The corresponding feature h S By constraining alignment with the L2 norm, the feature alignment loss is defined as:

[0052]

[0053] The KL divergence loss constrained classification probability distribution is:

[0054]

[0055] Where C is the total number of categories, x represents the model input, and y i Represents the prediction result of input x, p T ,p S Represents the softmax output of the teacher and student models respectively, using a dual loss function Ensure that the student model inherits the feature expression ability and classification decision boundary of the teacher model.

[0056] Furthermore, the knowledge distillation process uses the longitudinal dynamics equation in the body coordinate system:

[0057] F x =F aero +F thrust

[0058] in is the axial acceleration, q and r are the y-axis and z-axis components of the target angular velocity in the body coordinate system, v and w are the y-axis and z-axis components of the target velocity in the body coordinate system, and F aero is the aerodynamic component, F thrust is the engine thrust component, m is the mass;

[0059] The physical consistency loss function is:

[0060]

[0061] In the formula is the lift coefficient matrix of the teacher model and the student model, is the axial acceleration of the teacher model and the student model, ξ1 and ξ2 are learnable constraint weights, and the loss is minimized by the differential equation residual to force the motion parameter prediction of the student model to satisfy the aerodynamic laws.

[0062] Furthermore, the gradient importance screening mechanism based on parameter sensitivity in step S3-4 requires the definition of parameter θ j Fisher information I(θ j ):

[0063]

[0064] in is the total loss function, and a binary mask matrix M is generated by a dynamic threshold τ, whose elements M j satisfy:

[0065]

[0066] During back propagation, only highly sensitive parameters are updated, and the gradient update rule is:

[0067]

[0068] Among them, θ (t+1) ,θ (t) are the highly sensitive parameters that need to be updated at time steps t+1 and t, respectively, and η is the learning rate.

[0069] Compared with the prior art, the present invention adopts the above technical solution and has the following technical effects:

[0070] (1) The microwave photon strong maneuvering target recognition method based on small sample incremental learning provided by the present invention, through the time-frequency-polarization-geometry multi-domain feature fusion module, based on multiple deep learning models combined with the cross-domain attention mechanism to dynamically associate physical characteristics of different dimensions, effectively improves the feature representation ability of strong maneuvering targets in complex motion states, and significantly improves the robustness and cross-scenario generalization ability of dynamic target recognition while ensuring the consistency of electromagnetic scattering mechanism.

[0071] (2) The microwave photon strong maneuvering target recognition method based on small sample incremental learning provided by the present invention ensures that the generated synthetic data strictly conforms to the electromagnetic scattering mechanism by designing a polarization-time-frequency adversarial generative network and embedding the physical constraints of the Maxwell equations. By constraining the weight update direction of the generator by physical priors, the physical interpretability of the synthetic samples is significantly improved, thereby effectively expanding the diversity of the training set that conforms to the physical laws in small sample training, and avoiding the degradation of the generalization of the model due to data bias.

[0072] (3) The microwave photon strong maneuvering target recognition method based on small sample incremental learning provided by the present invention realizes lightweight model update through a dynamic knowledge distillation framework and a dual-path constraint mechanism, and uses a physical consistency loss function to force the motion parameter estimation values ​​and classification probability distribution of the student model and the teacher model to be aligned, thereby ensuring the stability of the old category knowledge in the incremental learning process. The sensitive areas for parameter update are dynamically identified based on the gradient importance mask, and only key weights are fine-tuned instead of all parameters being updated, which significantly reduces the demand for computing power of embedded radar hardware. BRIEF DESCRIPTION OF THE DRAWINGS

[0073] Figure 1 It is a principle structure diagram of the present invention. DETAILED DESCRIPTION

[0074] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the embodiments described are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary persons in this field without making creative efforts shall fall within the scope of protection of the present invention.

[0075] Microwave photon radar strong maneuvering target recognition method based on small sample incremental learning, such as Figure 1 As shown, the following steps are included:

[0076] Step S1: Input the target data acquired by the microwave photon radar into the multi-domain feature fusion module, extract the time-frequency energy dynamic characteristics through the long short-term memory network, mine the polarization scattering spatial pattern through the convolutional neural network, and model the geometric contour topological relationship through the graph neural network. The time-frequency energy dynamic characteristics, polarization scattering spatial pattern and geometric contour topological relationship are dynamically fused through the cross-domain attention mechanism to output multi-dimensional joint representation features;

[0077] Step S2: In the small sample data augmentation module, based on the multi-dimensional joint representation features obtained in step S1, a polarization-time-frequency adversarial generative network (PT-GAN) is designed. Synthetic data that conforms to the laws of electromagnetic scattering is generated in combination with the physical constraints of Maxwell's equations. Using this synthetic data and real data, a prototype feature support set is constructed through meta-learning.

[0078] Step S3: In the physical constraint incremental learning model, the prototype feature support set constructed in step S2 is used, and a dynamic knowledge distillation framework is adopted to align the motion parameters and classification probability distribution of the teacher model and the student model through physical consistency loss. The model training update is achieved by combining the gradient importance screening mechanism based on parameter sensitivity.

[0079] Furthermore, the specific extraction process of extracting the time-frequency energy dynamic characteristics through the long short-term memory network in step S1 is as follows:

[0080] The time-frequency graph generated by the Wigner-Ville distribution is used as input, and a bidirectional long short-term memory network is used to capture the temporal correlation of the target motion. The hidden state update process of the bidirectional long short-term memory network can be expressed as:

[0081]

[0082] where x t is the input of LSTM network, f t 、i t 、o t Represent the outputs of the forget gate, input gate, and output gate respectively, is the candidate cell state, C t 、C t-1 is the cell state at the current and previous time steps, h t 、h t-1 is the hidden layer state of the current and previous time steps, σ is the Sigmoid function, W f 、W i 、W c 、W o and b f 、b i 、b C 、b o are learnable parameters.

[0083] Furthermore, the step S1 of mining polarization scattering spatial patterns through convolutional neural networks is specifically as follows:

[0084] Based on the polarization scattering matrix decomposed by Pauli, a multi-scale convolutional neural network, namely CNN, is constructed to finally mine the polarization scattering spatial pattern and calculate the polarization entropy feature H P Quantify the randomness of target scattering, HP Calculated by the following formula:

[0085]

[0086] where p i represents the proportion of scattered energy of polarization channel i, and N is the total number of polarization channels;

[0087] By introducing reciprocity theorem constraints in the convolution kernel, the network is forced to learn polarization patterns that conform to electromagnetic laws.

[0088] Furthermore, the step S1 of constructing the geometric contour topological relationship through graph neural network modeling is specifically as follows:

[0089] The contour point cloud extracted from the range-Doppler image is modeled as a graph structure. The equivariant graph neural network is used to extract the rotation and translation invariant geometric features and construct the geometric contour topological relationship. The graph convolution process of the convolution layer is defined as:

[0090]

[0091] in represents the feature of node n at layer l+1, represents the feature of node n in layer l, represents the feature of node m in layer l, is the set of adjacent nodes, φ and ψ are node highlighting function and edge aggregation function respectively, e nm The feature vector representing the edge (n,m) is used to encode the geometric and semantic relationship between nodes n and m.

[0092] Furthermore, the dynamic fusion cross-domain attention mechanism is specifically calculated as follows:

[0093] F fusion =α·F tf +β·F p +γ·F g

[0094] Among them F tf ,F p ,F g are the time-frequency, polarization, and geometric feature vectors, respectively. The weight coefficients α, β, and γ are calculated by the multi-head self-attention mechanism:

[0095]

[0096] where Q tf and K tf is the query and key vector of time-frequency features, Q p and K p is the query and key vector of the polarization feature, Q g and Kg are the query and key vectors of geometric features, and d is the feature dimension.

[0097] Furthermore, the specific process of step S2 is as follows:

[0098] Step S2-1: Generate PT-GAN based on the multi-dimensional joint representation features obtained in step S1, wherein the PT-GAN includes a generator and a discriminator, wherein the generator of the PT-GAN is a noise vector. and a small number of real samples X real As input, a multi-layer deconvolution network is used to generate time-frequency-polarization joint data X = G(z, X real ), combined with the physical constraints of Maxwell's equations to generate synthetic data that conforms to the laws of electromagnetic scattering;

[0099] The loss function of the generator is composed of adversarial loss and physical constraints Joint composition:

[0100]

[0101] in To trade off the coefficient, adversarial loss D(·) is the output of the discriminator, G(z) is the output of the generator in PT-GAN, and the physical constraint term This is achieved by regularizing the differential form of Maxwell's equations, specifically:

[0102]

[0103] Physical constraints The electromagnetic field distribution E, B of the generated data is forced to satisfy the relationship between curl and divergence, Ω is the electric displacement vector, and ρ is the charge density;

[0104] The discriminator adopts the gradient penalty strategy of Wasserstein GAN to generate adversarial networks, and its loss function is defined as:

[0105]

[0106] Where X is the interpolation sample, and X=∈X real +(1-∈)X,∈~U(0,1),λ2 controls the intensity of gradient penalty;

[0107] Step S2-2: Using the generated synthetic data and real data, a prototype feature support set is constructed through meta-learning to enhance small sample adaptability, and prototype features are extracted from the time-frequency-polarization joint data X:

[0108]

[0109] Among them, f enc is the pre-trained encoder, S c is the support set sample of category c;

[0110] The prototype network minimizes the query set sample x q With prototype P c The Euclidean distance is used to achieve classification:

[0111]

[0112] Step S2-3: Introduce the dynamic measurement function g of the relationship network φ (f enc (x q ),P c ), the relational network adopts a two-layer fully connected architecture, takes the sample features output by the encoder and the category prototype extracted in step S2-3 as input, and outputs the confidence that the sample belongs to a specific category. Its loss function is expanded to:

[0113]

[0114] g φ It is implemented by a two-layer fully connected network, and the nonlinear boundary between categories is learned through end-to-end training. q Represents the minimized query set sample x q Corresponding label, (x q ,y q ) represents the query set sample-label pair.

[0115] Furthermore, the specific process of step S3 is as follows:

[0116] Step S3-1, using the synthetic data and prototype feature support set generated in step S2, initialize the teacher model and the student model, where the teacher model is based on the initial target recognition model and the student model is the target recognition model to be updated;

[0117] Step S3-2: Through the dual feature and semantic alignment mechanism of the teacher-student architecture, the teacher model parameters are fixed, and the student model is trained using synthetic data and the prototype feature support set. Parameter-level knowledge transfer is achieved through feature space constraints. At the same time, the KL divergence loss is introduced in the output layer to constrain the classification probability distribution, so that the student model inherits the classification decision boundary of the teacher model;

[0118] Step S3-3: Embed the hypersonic vehicle motion equations as physical constraints into the knowledge distillation process, and construct a physical consistency loss function. Continue training the student model trained in step S3-2 to align the motion parameters and classification probability distributions of the teacher model and the student model.

[0119] Step S3-4: Based on the training results of steps S3-2 and S3-3, a gradient importance screening mechanism based on parameter sensitivity is designed to fine-tune only the highly sensitive parameters to achieve lightweight incremental updates of the target recognition model.

[0120] Furthermore, the teacher model f in the teacher-student architecture T The intermediate feature h T With student model f S The corresponding feature h S By constraining alignment with the L2 norm, the feature alignment loss is defined as:

[0121]

[0122] The KL divergence loss constrained classification probability distribution is:

[0123]

[0124] Where C is the total number of categories, x represents the model input, and y i Represents the prediction result of input x, p T ,p S Represents the softmax output of the teacher and student models respectively, using a dual loss function Ensure that the student model inherits the feature expression ability and classification decision boundary of the teacher model.

[0125] Furthermore, the knowledge distillation process uses the longitudinal dynamics equation in the body coordinate system:

[0126] F x =F aero +F thrust

[0127] in is the axial acceleration, q and r are the y-axis and z-axis components of the target angular velocity in the body coordinate system, v and w are the y-axis and z-axis components of the target velocity in the body coordinate system, and F aero is the aerodynamic component, F thrust is the engine thrust component, m is the mass;

[0128] The physical consistency loss function is:

[0129]

[0130] In the formula is the lift coefficient matrix of the teacher model and the student model, is the axial acceleration of the teacher model and the student model, ξ1 and ξ2 are learnable constraint weights, and the loss is minimized by the differential equation residual to force the motion parameter prediction of the student model to satisfy the aerodynamic laws.

[0131] Furthermore, the gradient importance screening mechanism based on parameter sensitivity in step S3-4 requires the definition of parameter θ j Fisher information I(θ j ):

[0132]

[0133] in is the total loss function, and a binary mask matrix M is generated by a dynamic threshold τ, whose elements M j satisfy:

[0134]

[0135] During back propagation, only highly sensitive parameters are updated, and the gradient update rule is:

[0136]

[0137] Among them, θ (t+1) ,θ (t) are the highly sensitive parameters that need to be updated at time steps t+1 and t, respectively, and η is the learning rate.

[0138] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A microwave photon radar strong maneuvering target recognition method based on small sample incremental learning, characterized by: The steps include: Step S1: Input the target data acquired by the microwave photon radar into the multi-domain feature fusion module, extract the time-frequency energy dynamic characteristics through the long short-term memory network, mine the polarization scattering spatial pattern through the convolutional neural network, and model the geometric contour topological relationship through the graph neural network. The time-frequency energy dynamic characteristics, polarization scattering spatial pattern and geometric contour topological relationship are dynamically fused through the cross-domain attention mechanism to output multi-dimensional joint representation features; Step S2: In the small sample data augmentation module, based on the multi-dimensional joint representation features obtained in step S1, a polarization-time-frequency adversarial generative network (PT-GAN) is designed. Synthetic data that conforms to the laws of electromagnetic scattering is generated in combination with the physical constraints of Maxwell's equations. Using this synthetic data and real data, a prototype feature support set is constructed through meta-learning. Step S3: In the physical constraint incremental learning model, the prototype feature support set constructed in step S2 is used, and a dynamic knowledge distillation framework is adopted to align the motion parameters and classification probability distribution of the teacher model and the student model through physical consistency loss. The model training update is achieved by combining the gradient importance screening mechanism based on parameter sensitivity.

2. The microwave photon radar strong maneuvering target recognition method based on small sample incremental learning according to claim 1 is characterized in that: In step S1, the specific extraction process of extracting the time-frequency energy dynamic characteristics through the long short-term memory network is as follows: The time-frequency graph generated by the Wigner-Ville distribution is used as input, and a bidirectional long short-term memory network is used to capture the temporal correlation of the target motion. The hidden state update process of the bidirectional long short-term memory network can be expressed as: where x t is the input of LSTM network, f t 、i t 、o t Represent the outputs of the forget gate, input gate, and output gate respectively, is the candidate cell state, C t 、C t-1 is the cell state at the current and previous time steps, h t 、h t-1 is the hidden layer state of the current and previous time steps, σ is the Sigmoid function, W f 、W i 、W c 、W o and b f 、b i 、b C 、b o are learnable parameters.

3. The microwave photon radar strong maneuvering target recognition method based on small sample incremental learning according to claim 1 is characterized in that: The specific steps of mining polarization scattering spatial patterns through convolutional neural networks in step S1 are as follows: Based on the polarization scattering matrix decomposed by Pauli, a multi-scale convolutional neural network, namely CNN, is constructed to finally mine the polarization scattering spatial pattern and calculate the polarization entropy feature H P Quantify the randomness of target scattering, H P Calculated by the following formula: where p i represents the proportion of scattered energy of polarization channel i, and N is the total number of polarization channels; By introducing reciprocity theorem constraints in the convolution kernel, the network is forced to learn polarization patterns that conform to electromagnetic laws.

4. The microwave photon radar strong maneuvering target recognition method based on small sample incremental learning according to claim 1 is characterized in that: The specific steps in step S1 regarding constructing geometric contour topological relationships through graph neural network modeling are as follows: The contour point cloud extracted from the range-Doppler image is modeled as a graph structure. The equivariant graph neural network is used to extract the rotation and translation invariant geometric features and construct the geometric contour topological relationship. The graph convolution process of the convolution layer is defined as: in represents the feature of node n at layer l+1, represents the feature of node n in layer l, represents the feature of node m in layer l, is the set of adjacent nodes, φ and ψ are node highlighting function and edge aggregation function respectively, e nm The feature vector representing the edge (n,m) is used to encode the geometric and semantic relationship between nodes n and m.

5. The microwave photon radar strong maneuvering target recognition method based on small sample incremental learning according to claim 1 is characterized in that: The dynamic fusion cross-domain attention mechanism is specifically calculated as follows: F fusion =α·F tf +β·F p +γ·F g Among them F tf ,F p ,F g are the time-frequency, polarization, and geometric feature vectors, respectively. The weight coefficients α, β, and γ are calculated by the multi-head self-attention mechanism: where Q tf and K tf is the query and key vector of time-frequency features, Q p and K p is the query and key vector of the polarization feature, Q g and K g are the query and key vectors of geometric features, and d is the feature dimension.

6. The microwave photon radar strong maneuvering target recognition method based on small sample incremental learning according to claim 1 is characterized in that: The specific process of step S2 is as follows: Step S2-1: Generate PT-GAN based on the multi-dimensional joint representation features obtained in step S1, wherein the PT-GAN includes a generator and a discriminator, wherein the generator of the PT-GAN is a noise vector. and a small number of real samples X real As input, a multi-layer deconvolution network is used to generate time-frequency-polarization joint data X = G(z, X real ), combined with the physical constraints of Maxwell's equations to generate synthetic data that conforms to the laws of electromagnetic scattering; The loss function of the generator is composed of adversarial loss and physical constraints Joint composition: in To trade off the coefficient, adversarial loss D(·) is the output of the discriminator, G(z) is the output of the generator in PT-GAN, and the physical constraint term This is achieved by regularizing the differential form of Maxwell's equations, specifically: Physical constraints The electromagnetic field distribution E, B of the generated data is forced to satisfy the relationship between curl and divergence, Ω is the electric displacement vector, and ρ is the charge density; The discriminator adopts the gradient penalty strategy of Wasserstein GAN to generate adversarial networks, and its loss function is defined as: Where X is the interpolation sample, and X=∈X real +(1-∈)X,∈~U(0,1),λ2 controls the intensity of gradient penalty; Step S2-2: Using the generated synthetic data and real data, a prototype feature support set is constructed through meta-learning to enhance small sample adaptability, and prototype features are extracted from the time-frequency-polarization joint data X: Among them, f enc is the pre-trained encoder, S c is the support set sample of category c; The prototype network minimizes the query set sample x q With prototype P c The Euclidean distance is used to achieve classification: Step S2-3: Introduce the dynamic measurement function g of the relationship network φ (f enc (x q ),P c ), the relational network adopts a two-layer fully connected architecture, takes the sample features output by the encoder and the category prototype extracted in step S2-3 as input, and outputs the confidence that the sample belongs to a specific category. Its loss function is expanded to: g φ It is implemented by a two-layer fully connected network, and the nonlinear boundary between categories is learned through end-to-end training. q Represents the minimized query set sample x q Corresponding label, (x q ,y q ) represents the query set sample-label pair.

7. The microwave photon radar strong maneuvering target recognition method based on small sample incremental learning according to claim 1 is characterized in that: The specific process of step S3 is as follows: Step S3-1, using the synthetic data and prototype feature support set generated in step S2, initialize the teacher model and the student model, where the teacher model is based on the initial target recognition model and the student model is the target recognition model to be updated; Step S3-2: Through the dual feature and semantic alignment mechanism of the teacher-student architecture, the teacher model parameters are fixed, and the student model is trained using synthetic data and the prototype feature support set. Parameter-level knowledge transfer is achieved through feature space constraints. At the same time, the KL divergence loss is introduced in the output layer to constrain the classification probability distribution, so that the student model inherits the classification decision boundary of the teacher model; Step S3-3: Embed the hypersonic vehicle motion equations as physical constraints into the knowledge distillation process, and construct a physical consistency loss function. Continue training the student model trained in step S3-2 to align the motion parameters and classification probability distributions of the teacher model and the student model. Step S3-4: Based on the training results of steps S3-2 and S3-3, a gradient importance screening mechanism based on parameter sensitivity is designed to fine-tune only the highly sensitive parameters to achieve lightweight incremental updates of the target recognition model.

8. The microwave photon radar strong maneuvering target recognition method based on small sample incremental learning according to claim 7 is characterized in that: The teacher model f in the teacher-student architecture T The intermediate feature h T With student model f S The corresponding feature h S By constraining alignment with the L2 norm, the feature alignment loss is defined as: The KL divergence loss constrained classification probability distribution is: Where C is the total number of categories, x represents the model input, and y i Represents the prediction result of input x, p T ,p S Represents the softmax output of the teacher and student models respectively, using a dual loss function Ensure that the student model inherits the feature expression ability and classification decision boundary of the teacher model.

9. The microwave photon radar strong maneuvering target recognition method based on small sample incremental learning according to claim 7 is characterized in that: The knowledge distillation process uses the longitudinal dynamics equation in the body coordinate system: in is the axial acceleration, q and r are the y-axis and z-axis components of the target angular velocity in the body coordinate system, v and w are the y-axis and z-axis components of the target velocity in the body coordinate system, and F aero is the aerodynamic component, F thrust is the engine thrust component, m is the mass; The physical consistency loss function is: In the formula is the lift coefficient matrix of the teacher model and the student model, is the axial acceleration of the teacher model and the student model, ξ1 and ξ2 are learnable constraint weights, and the loss is minimized by the differential equation residual to force the motion parameter prediction of the student model to satisfy the aerodynamic laws.

10. The microwave photon radar strong maneuvering target recognition method based on small sample incremental learning according to claim 7 is characterized in that: The gradient importance screening mechanism based on parameter sensitivity in step S3-4 requires the definition of parameter θ j Fisher information I(θ j ): in is the total loss function, and a binary mask matrix M is generated by a dynamic threshold τ, whose elements M j satisfy: During back propagation, only highly sensitive parameters are updated, and the gradient update rule is: Among them, θ (t+1) ,θ (t) are the highly sensitive parameters that need to be updated at time steps t+1 and t, respectively, and η is the learning rate.

Images