A distributed radar target recognition method based on bistatic angle prior and feature decoupling

Abstract

The application discloses a distributed radar target recognition method based on bistatic angle prior and feature decoupling, and belongs to the technical field of radar target recognition. The method is directed to a one-transmitting and multi-receiving distributed radar scene, and after pretreatment is performed on multi-station high-resolution range images and corresponding bistatic angle information, basic features of each station are extracted through a parameter sharing convolution network; a commonness-personality double-branch decoupling architecture is constructed to mine target commonness features irrelevant to observation angles through cross-station attention, to extract station personality features driven by angle prior, and to obtain global fusion features through fusion interaction and angle gate control weighting; and joint loss combining distributed alignment and orthogonal decoupling constraints is used to complete end-to-end training, and a target recognition result is output. The application can effectively mine multi-station complementary information, and improve the distributed radar target recognition precision and generalization performance.

Description

Technical Field

[0001] This invention belongs to the field of radar target recognition technology, specifically relating to a distributed radar target recognition method based on bistatic angle prior and feature decoupling. Background Technology

[0002] Radar target recognition is an important research direction in the field of radar information processing. Its main task is to distinguish different target types by utilizing the scattering characteristics contained in the target echo signal. High-resolution range profiles (HRRPs) can characterize the distribution characteristics of the scattering centers of a target along the radar line of sight. They have the advantages of small data volume and convenient acquisition, and have become the most commonly used information carrier in the field of radar target recognition.

[0003] In practical engineering applications, conventional monostation radars can only observe targets from a single perspective, which has several limitations in target identification tasks: First, the scattering information obtained from single-view observation is limited and cannot fully reflect the scattering characteristics of the target under different observation directions; Second, it is sensitive to changes in target attitude and local obstruction. When the target observation geometry changes, the HRRP scattering structure is prone to significant differences, resulting in the loss of effective discrimination information and a decrease in the model's generalization ability; Third, HRRP only describes the one-dimensional range scattering distribution of the target along the radar line of sight, which is difficult to fully characterize the overall scattering characteristics of the target and has certain limitations on the identification performance of targets with complex structures.

[0004] To address the aforementioned shortcomings of monostation radar, distributed multistation radar utilizes multiple spatially separated receiving stations to simultaneously observe the same target from different perspectives, enabling the acquisition of multi-dimensional complementary scattering information of the target. However, most existing multistation radar target identification methods employ shallow fusion approaches such as simple stitching, addition, or voting. While these methods can aggregate multistation observation information to some extent, they typically fail to model the complementary relationships of features and observational geometric differences between different stations, making it difficult to achieve deep interaction of multistation information. This can easily lead to insufficient utilization of effective discrimination information or the superposition of redundant information, resulting in limited fusion effectiveness.

[0005] Therefore, developing a distributed radar target identification method that can take into account the deep fusion of common information from multiple stations, the effective preservation of individual characteristics of each station, and the full utilization of observational geometric priors is a technical problem that urgently needs to be solved in this field. Summary of the Invention

[0006] The technical problem solved by this invention is to address the issues of insufficient cross-site feature depth interaction, inadequate utilization of observation geometric prior information, and decreased discriminative power caused by information aliasing between common and individual features in existing distributed radar target recognition methods. This invention proposes a distributed radar target recognition method based on bistatic angle prior and feature decoupling. First, this method extracts baseline features from each site using a parameter-sharing multi-scale convolutional module. Then, it constructs a dual-branch parallel decoupling architecture: the first branch mines inherent common features of the target that are independent of the observation angle through cross-site collaborative attention; the second branch uses bistatic angle prior to drive individual feature extraction, where the angle prior performs geometric recalibration of features through angle gating and explicitly concatenates them for individual decoding. Finally, the features from the two branches are concatenated and deeply fused with a lightweight fusion layer, then weighted and aggregated by angle scalar gating. A dual-constraint loss is designed: Central Moment Discrepancy (CMD) forces cross-site common distribution alignment, and cosine similarity forces orthogonal decoupling of common and individual features, thereby effectively improving target recognition performance.

[0007] The technical solution of this invention is: Step S1, Dataset Construction and Preprocessing For distributed radar scenarios with one transmitter and multiple receivers, each sample consists of HRRP sequences synchronously acquired by multiple spatially separated receiving stations, as well as bistatic angle information corresponding to each receiving station. The multi-station HRRP sequences constitute the station dimension input, which is preprocessed with centroid alignment and amplitude normalization. The bistatic angle information constitutes the observation geometric prior input, which is preprocessed with linear normalization.

[0008] Step S11, center-of-gravity alignment preprocessing: Different targets have different positions within the distance window. The HRRP (Heavy Center of Targets Rearranged) is used for alignment, shifting the target's centroid to the center of the window. First, the centroid position of the sequence is calculated. Then, the HRRP is cyclically shifted to move the target's centroid to the midpoint of the sequence. (Centroid) G It can be calculated using the following formula:

[0009] in, Indicates the echo signal to be identified; L This represents the length of the HRRP sequence.

[0010] Step S12, amplitude normalization preprocessing: To eliminate the influence of the target HRRP amplitude, the amplitude normalization process can be expressed as:

[0011] in This indicates that the maximum value is taken, and after normalization, the HRRP sequence value is mapped to the range of 0 to 1.

[0012] Step S13, Angle Information Preprocessing: Let the first s The bistatic angle vectors corresponding to each receiving station are: The four dimensions correspond to the incident pitch angle, incident azimuth angle, exit pitch angle, and exit azimuth angle in the aircraft-centered body coordinate system. The original angle values ​​range from [value missing]. Linear normalization can be expressed as:

[0013] Normalized angle vector The range of values ​​for each element is 1. .

[0014] Step S2, Multi-scale Convolutional Feature Extraction Module The preprocessed HRRP sequences from each site are input into a multi-scale convolutional feature extraction module with shared parameters across sites to extract the basic feature vectors of each site. This module shares network weights across all sites, so that the features of different sites are in a unified feature space, which facilitates subsequent fusion and significantly reduces the number of parameters.

[0015] Step S21, Multi-scale parallel convolution extraction: The single-site HRRP sequence is treated as a single-channel, one-dimensional input and fed into the first-level multi-scale convolutional layer. This layer contains three parallel convolutional branches with kernel sizes of 3, 5, and 7, respectively extracting short-range detail features, mid-range structural features, and long-range contextual features of the HRRP. Let the input HRRP be... ( L (where the sequence length is ), the outputs of the three branches are as follows:

[0016] The outputs of the three parallel branches are concatenated along the channel dimension, and then processed by batch normalization (BN) and ReLU activation function to obtain the first-level multi-scale features. :

[0017] Here, [⋅,⋅,⋅] represents the channel dimension splicing operation.

[0018] Step S22, secondary refinement of residual features: The first-level multi-scale features are fed into the first-level residual block, which consists of two 1×3 convolutional layers, batch normalization, and ReLU activation functions. Residual connections are achieved through identity mapping to alleviate the gradient vanishing problem in deep networks. (First-level residual features) for:

[0019] in This represents the composite function of two convolutional layers and activation:

[0020] Step S23, downsampling and high-level feature extraction: The first-level residual characteristics Downsampling was performed using a max-pooling layer with a kernel size of 2 and a step size of 2.

[0021] The data is then sequentially fed into the second-level multi-scale convolutional layer and the second-level residual block to extract high-level semantic features; let the output of the second-level multi-scale convolution be... Then the second-level residual block output satisfy:

[0022] Step S24, fixed-length feature output: The second-level residual characteristics After being compressed into a one-dimensional vector by an adaptive average pooling layer, a 128-dimensional basic feature vector for each site is finally obtained. :

[0023] in s = 1,2,…,S represents the station number.

[0024] Step S3, Angle Scalar Gating Module The normalized angle vectors of each station The input is fed into a scalar gating network with shared parameters across all sites, generating site-level scalar weights in the range of 0 to 1. This is used to characterize the overall contribution of the site to the identification results under the current observation geometry. The network consists of two fully connected layers and a sigmoid activation function, as shown in the following formula:

[0025] Step S4, common feature extraction The input to this module is the basic feature vector of each site. This study utilizes a multi-head self-attention mechanism to uncover deep common representations across different sites. S The basic features of each site are stacked according to the site dimension into a dimension. B × S The feature tensor of 128 F ,in BTo determine the batch size for training, the feature tensor is input into the multi-station collaborative attention module to achieve cross-station information exchange and extract common features of the target. The specific execution steps of this module are as follows: Step S41, Multi-head Self-Attention Layer: use N Head self-attention structure, for input feature tensors F Perform a linear transformation to obtain the query matrix. Q Key matrix K Value matrix V :

[0026] in, , , The weight matrix is ​​a learnable matrix; according to N The head mechanism divides features into N For each subspace, calculate the scaled dot product attention:

[0027] in, The key vector dimension of single-head attention. i =1,2,…, N Attention header number; The outputs of each head are concatenated along the channel dimension and then restored to the original feature dimension via the output mapping matrix to obtain the multi-head attention output (MHA). F ); Step S42, Residuals and Layer Normalization: Residual connections and layer normalization are used for the attention output to ensure training stability and obtain common features. :

[0028] in, The s The common characteristics of each site are: It has 128 dimensions.

[0029] Step S5, Individual Feature Extraction By constructing a cross-site, fully parameter-sharing feature extraction module, individual features related to the observation angle of each station are extracted. This supplements the missing but station-specific discriminative information in the common features, thereby explicitly modeling the scattering feature changes caused by the observation angle. The specific execution steps of this module are as follows: Step S51, Angle-driven vector-gated recalibration: The first s Angle vector of each station A 128-dimensional scaling factor with the same dimension as the basic feature is generated through a two-layer fully connected network. The formula is as follows:

[0030] The output range of tanh(⋅) is [-1, 1], therefore the scaling factor is... The range is [0.5, 1.5], which allows for bidirectional adaptive adjustment of features.

[0031] Then recalibrate the basic features:

[0032] in This indicates element-wise multiplication.

[0033] Step S52, Decoding Personality Traits: Features after angle-gated scaling With the corresponding angle vector The input vector is concatenated along the channel dimension to obtain a 132-dimensional input vector, which is then fed into a three-layer fully connected personalized feature decoder to output 128-dimensional site-specific features. :

[0034] Step S6: Weighted fusion of features from multiple stations First, the common and unique characteristics of the sites are combined along the feature dimension:

[0035] Secondly, to promote deep interaction between common and individual characteristics, a lightweight fusion layer is set after splicing, consisting of linear transformation, layer normalization, ReLU activation, and Dropout:

[0036] Finally, the angle scalar weights generated in step S3 are used. ,right S The fusion features of each site are weighted, summed, and normalized to obtain the final global fusion features. :

[0037] Step S7, Classifier Output The classifier uses a two-layer fully connected layer structure, takes 256-dimensional global fusion features as input, and outputs the predicted probability of the corresponding target category.

[0038] Step S8, Construction of joint loss function To achieve effective fusion of distributed multi-site features and ensure classification performance, consistency of common features, and decoupling of individual features, this invention constructs a joint loss function consisting of main classification loss, common constraint loss, and difference constraint loss, and performs end-to-end training and optimization of the network.

[0039] Step S81, Main classification loss : The cross-entropy loss function is used to supervise the class probabilities output by the model, in order to measure the difference between the predicted results and the true labels. The formula is as follows:

[0040] Where N represents the training batch size, and C represents the number of classes. This represents the true label of the i-th sample. This represents the probability of the corresponding class predicted by the model.

[0041] Step S82, Common Constraint Loss : To ensure the consistency of the distribution of common features among different receiving stations and to eliminate the influence of differences in observation angles, this invention employs a loss function based on the difference in central moments to align the distribution of common features among the receiving stations. The relevant formulas are as follows:

[0042] in, Representation of features x The k-th order central moment, The number of pairs of stations is denoted by K, which represents the order of the selected central moments.

[0043] Step S83, Difference Constraint Loss : To effectively decouple the inherent common features of the target from the unique features observed at each site, and to avoid information overlap between the two types of features in the representation space, this invention constructs a difference constraint loss based on cosine similarity to quantitatively constrain the representational independence of the two types of features. The relevant formulas are as follows:

[0044] in, S Indicates the number of receiving stations; This is an L2 norm operation used to calculate the magnitude of the eigenvector; This is the numerical stability coefficient, with a fixed value. .

[0045] Step S84, Joint Loss and Dynamic Weight Adjustment: The total loss function is:

[0046] in, , The dynamic weights for common constraint loss and difference constraint loss are respectively adjusted using the following strategy:

[0047]

[0048] in, This represents the target proportion coefficient of the common constraint loss, ensuring that the magnitude of the common constraint loss does not exceed the principal classification loss. This ensures the core priority of the main classification task, with a typical value of 0.2. This represents the target proportion coefficient of the difference constraint loss, ensuring that the magnitude of the difference constraint loss does not exceed that of the main classification loss. To avoid decoupling constraints interfering with the convergence of the main classification task, a typical value of 0.1 is used. The dynamic weights representing the common constraint loss are limited to the range of [0, 0.1] and are adaptively adjusted according to the training process and the loss magnitude. The dynamic weights representing the difference constraint loss are limited to the range [0, 0.05] and are adaptively adjusted with the training process and the loss magnitude.

[0049] Beneficial effects: First, addressing the issue of insufficient cross-station information interaction in existing multi-station identification methods, this invention employs a cross-station multi-head self-attention mechanism in the common feature extraction branch. This mechanism performs deep interaction and fusion modeling of the basic features of each receiving station, effectively mining complementary discriminative information among multi-station observations. On an 8-class target test set, the identification accuracy reaches 96.61%, an improvement of 6.27 percentage points compared to a single-station CNN (90.34%).

[0050] Second, addressing the problem that existing multi-station fusion methods lack observational geometric priors and struggle to allocate contributions from different stations based on differences in observational geometry, this invention introduces a bistatic angle prior-driven individual feature branch. Features are recalibrated using angle vectors and combined with angle scalar gating to achieve station-level adaptive weighted fusion, thereby improving feature representation capabilities under different observational geometric conditions. Compared to feature splicing fusion methods (91.43%), the recognition accuracy is improved by 5.18 percentage points.

[0051] Third, to address the problem of common and individual information easily overlapping during multi-station feature fusion, leading to feature redundancy and decreased discriminative ability, this invention constructs a decoupled dual-branch structure for common and individual information. Combined with central moment difference loss and cosine similarity constraints, it achieves cross-station common distribution alignment and feature independence constraints, thereby avoiding the overlap of common and individual information and improving discriminative ability. Compared to multi-view feature fusion methods (93.76%), this method improves the recognition accuracy by 2.85 percentage points. Attached Figure Description

[0052] Figure 1 A flowchart of the overall process of the distributed radar target recognition method; Figure 2 , one Schematic diagram of a three-transmitter distributed radar scenario; Figure 3 A schematic diagram of a multi-scale convolutional feature extraction structure; Figure 4 Schematic diagram of the angle scalar gating module; Figure 5 Schematic diagram of the common feature extraction module; Figure 6 A schematic diagram of the individual feature extraction module; Figure 7 A schematic diagram of the multi-station feature weighted fusion module structure; Figure 8 A schematic diagram of the confusion matrix of the simulation experiment test set; Figure 9 A schematic diagram of the t-SNE dimensionality reduction distribution of the fusion features. Detailed Implementation

[0053] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. The specific embodiments described herein are merely illustrative of the invention and are not intended to limit the scope of protection of this invention.

[0054] This embodiment provides a distributed radar target recognition method based on bistatic angle prior and feature decoupling, targeting a one-transmitter, three-receiver distributed radar scenario, and achieving target recognition for typical civil aircraft. The overall process framework of the method is as follows: Figure 1 As shown.

[0055] The detailed process of the method described in this embodiment is as follows: Step S1, Dataset Construction and Preprocessing like Figure 2As shown, this embodiment addresses a distributed radar scenario with one transmitter and three receivers, enabling target identification for typical civil aircraft. Corresponding radar echo data is collected for each target type under different flight states, including both straight-line cruise and hovering flight. Each sample consists of HRRP sequences synchronously acquired by three spatially separated receiving stations, along with bistatic observation angle information for each station. The multi-station HRRP sequences (dimension 128) constitute the station dimension input, preprocessed using centroid alignment and amplitude normalization; the bistatic angle information (dimension 4) constitutes the observation geometric prior input, preprocessed using linear normalization.

[0056] Step S11, center-of-gravity alignment preprocessing: Different targets have different positions within the distance window. The HRRP (Heavy Center of Targets Rearranged) is used for alignment, shifting the target's centroid to the center of the window. First, the centroid position of the sequence is calculated. Then, the HRRP is cyclically shifted to move the target's centroid to the midpoint of the sequence. (Centroid) G It can be calculated using the following formula:

[0057] in, This indicates the echo signal to be identified.

[0058] Step S12, amplitude normalization preprocessing: To eliminate the influence of the target HRRP amplitude, the amplitude normalization process can be expressed as:

[0059] in This indicates that the maximum value is taken, and after normalization, the HRRP sequence value is mapped to the range of 0 to 1.

[0060] Step S13, Angle Information Preprocessing: Let the first s The bistatic angle vectors corresponding to each receiving station are: The four dimensions correspond to the incident pitch angle, incident azimuth angle, exit pitch angle, and exit azimuth angle in the aircraft-centered body coordinate system. The original angle values ​​range from [value missing]. Linear normalization can be expressed as:

[0061] Where: the normalized angle vector The range of values ​​for each element is 1. .

[0062] Step S2, Multi-scale Convolutional Feature Extraction Module The structure of this module is as follows: Figure 3As shown, the input is a single-site, single-channel 128-dimensional HRRP sequence, and the output is a 128-dimensional single-site basic feature vector. The specific structure is as follows: Step S21, First-level multi-scale convolutional layer: The input channel has 1 channel, the total output channel has 64 channels, and there are 3 parallel convolutional branches with kernel sizes of 3, 5, and 7, and corresponding padding of 1, 2, and 3. The number of channels per branch is 21, 21, and 22.

[0063] The outputs of the three parallel branches are concatenated along the channel dimension, and then sequentially pass through a batch normalization layer and a ReLU activation function to output the first-level multi-scale features. :

[0064] Here, [⋅,⋅,⋅] represents the channel dimension splicing operation.

[0065] Step S22, First-level residual block: The input channel count is 64, consisting of two 1×3 convolutional layers, a batch normalization layer, and a ReLU activation function. Residual connections are achieved through identity mapping. First-level residual features. for:

[0066] in This represents the composite function of two convolutional layers and activation:

[0067] Step S23, Max Pooling Layer: The first-level residual characteristics The feature length is halved by downsampling through a max-pooling layer with a kernel size of 2 and a stride of 2.

[0068] Step S24, Second-level multi-scale convolutional layer: The input has 64 channels, and the total output has 128 channels. It contains 3 parallel convolutional branches with kernel sizes of 3, 5, and 7, and corresponding padding of 1, 2, and 3, respectively. The number of channels per branch is 43, 43, and 42. After concatenating the channel dimensions, the branch outputs are sequentially passed through a batch normalization layer and a ReLU activation function to output the second-level multi-scale features. ; Step S25, Second-level residual block: The number of input channels is 128, and the structure is consistent with the first-level residual block. Step S26, Adaptive average pooling layer: Output features of the second-level residual block Compressed into a 128-dimensional single-site basic feature vector :

[0069] Step S3, Angle Scalar Gating Module The structure of this module is as follows: Figure 4 As shown, the input is the normalized angle vector of each station. The output is a site-level scalar weight; it consists of two fully connected layers, a ReLU activation function, and a Sigmoid activation function; the first fully connected layer has an input dimension of 4 and an output dimension of 32, and the second fully connected layer has scalar weights with an input dimension of 32 and an output dimension of 1. :

[0070] Step S4, common feature extraction This module is structured as follows. Figure 5 As shown, the input is a B×3×128-dimensional feature tensor composed of the basic features of the three stations stacked together. F (B is the training batch size), the output is the common features of the sites, and the specific implementation parameters are as follows: Step S41, Multi-head Self-Attention Layer: A 4-head self-attention structure is adopted, with an embedding dimension of 128 and a single-head dimension of 32; the input features are linearly transformed to obtain the query matrix Q, the key matrix K, and the value matrix V.

[0071] in, , , The weight matrix is ​​a learnable matrix; Then calculate the scaling dot product attention for each of the four subspaces. :

[0072] in, The key vector dimension of single-head attention. i =1,2,3,4 are the attention head numbers; After concatenating the outputs of each head along the channel dimension, the outputs are restored to the original 128-dimensional feature dimension via the output mapping matrix, resulting in the multi-head attention output (MHA). F ); Step S42, Residuals and Layer Normalization: The output of the multi-head self-attention layer is added to the input features using a residual connection, and then subjected to layer normalization to obtain common features. :

[0073] in, The s The common characteristics of each site are: It has 128 dimensions.

[0074] Step S5, Individual Feature Extraction The structure of this module is as follows: Figure 6 As shown, it consists of two parts: angle-driven residual vector gating and individual feature decoder.

[0075] Step S51, Angle-driven vector-gated recalibration: For the s Angle vector of each station 128-dimensional scaling factors are generated through a two-layer fully connected network. :

[0076] The first fully connected layer has an input dimension of 4 and an output dimension of 64, while the second fully connected layer has an input dimension of 64 and an output dimension of 128. (Scaling factor) The range is [0.5, 1.5], which allows for bidirectional adaptive adjustment of the basic features.

[0077] Then recalibrate the basic features:

[0078] in This indicates element-wise multiplication.

[0079] Step S52, Decoding Personality Traits: Features after angle-gated scaling With the corresponding angle vector The input vector is concatenated along the channel dimension to obtain a 132-dimensional input vector. This vector is then fed into a three-layer fully connected network, which outputs 128-dimensional individual features. :

[0080] The decoder's specific structure is as follows: The first fully connected layer has an input dimension of 132 and an output dimension of 256. It is followed by a LeakyReLU activation function with a negative half-axis slope of 0.2 and a Dropout layer with a dropout rate of 0.2. The second fully connected layer has an input dimension of 256 and an output dimension of 128, followed by a LeakyReLU activation function with a negative half-axis slope of 0.2. The third fully connected layer has an input dimension of 128 and an output dimension of 128. Weight initialization: All fully connected layers are initialized uniformly using Xavier, and the bias term is initialized to 0.01 to avoid all outputs being zero in the initial training phase.

[0081] Step S6: Weighted fusion of features from multiple stations The structure of this module is as follows: Figure 7 As shown. First, the common and unique characteristics of the sites are concatenated along the feature dimension:

[0082] Secondly, to promote deep interaction between common and individual characteristics, a lightweight fusion layer is set after splicing, consisting of linear transformation (256→256), layer normalization, ReLU activation, and Dropout (0.2):

[0083] Finally, the angle scalar weights generated in step S3 are used. The weighted summation and normalization of the fusion features from the three sites are used to obtain the final global fusion features. :

[0084] Step S7, Classifier Output The classifier takes 256-dimensional global fusion features as input and outputs the predicted probability of the corresponding target class. The specific structure is as follows: The first fully connected layer has an input dimension of 256 and an output dimension of 128, followed by a batch normalization layer and a ReLU activation function. The second fully connected layer has an input dimension of 128 and an output dimension of 8, which is the number of target categories.

[0085] Step S8, Joint Loss Function and Dynamic Weight Adjustment During the training phase, a joint loss function consisting of classification loss, common constraint loss, and difference constraint loss is used to optimize the network parameters end-to-end and dynamically adjust the auxiliary loss weights.

[0086] Step S81, Main classification loss : The cross-entropy loss function is used to supervise the class probabilities output by the model, in order to measure the difference between the predicted results and the true labels. The formula is as follows:

[0087] Where N represents the training batch size, and C represents the number of classes. This represents the true label of the i-th sample. This represents the probability of the corresponding class predicted by the model.

[0088] Step S82, Common Constraint Loss : To ensure the consistency of the distribution of common features among different receiving stations and to eliminate the influence of differences in observation angles, this invention employs a loss function based on the difference in central moments to align the distribution of common features among the receiving stations. The relevant formulas are as follows:

[0089] in, Representation of features x The k-th order central moment, The number of pairs of stations is denoted by K, which represents the order of the selected central moments. In this embodiment, K = 3.

[0090] Step S83, Difference Constraint Loss : To effectively decouple the inherent common features of the target from the unique features observed at each site, and to avoid information overlap between the two types of features in the representation space, this invention constructs a difference constraint loss based on cosine similarity to quantitatively constrain the representational independence of the two types of features. The relevant formulas are as follows:

[0091] in, S This represents the number of receiving stations; in this embodiment, it is taken as... S = 3; This is an L2 norm operation used to calculate the magnitude of the eigenvector; This is the numerical stability coefficient, with a fixed value. .

[0092] Step S84, Joint Loss and Dynamic Weight Adjustment The total loss function is:

[0093] in, , The dynamic weights for common constraint loss and difference constraint loss are respectively adjusted using the following strategy:

[0094]

[0095] in, This represents the target proportion coefficient of the common constraint loss, ensuring that the magnitude of the common constraint loss does not exceed the principal classification loss. This ensures the core priority of the main classification task; This represents the target proportion coefficient of the difference constraint loss, ensuring that the magnitude of the difference constraint loss does not exceed that of the main classification loss. To avoid decoupling constraints interfering with the convergence of the main classification task; The dynamic weights representing the common constraint loss are limited to the range of [0, 0.1] and are adaptively adjusted according to the training process and the loss magnitude. The dynamic weights representing the difference constraint loss are limited to the range [0, 0.05] and are adaptively adjusted with the training process and the loss magnitude.

[0096] Optionally, the batch size during training is 64, and the initial learning rate is [missing information]. The optimizer is AdamW, and the weight decay factor is... The learning rate is decayed using a cosine annealing strategy, and the total number of iterations is set to 100.

[0097] In an optional embodiment of the present invention, the effectiveness of the distributed radar target recognition method based on bistatic angle prior and feature decoupling provided in the above embodiment is verified by simulation experiments, specifically as follows: I. Simulation Conditions The hardware platform for the simulation experiment of this invention is: Windows operating system, Intel Core i7-14650HX processor @ 2.20 GHz, 16 GB of memory, and NVIDIA GeForce RTX 5060 graphics card.

[0098] The software platform for the simulation experiment of this invention is: Python 3.10, PyTorch 2.0, CUDA 11.7.

[0099] The dataset used in the simulation experiments of this invention consists of bistatic HRRP echo data of eight typical civil aircraft targets. The data was generated using the electromagnetic simulation software FEKO, with a simulation center frequency of 3 GHz, a signal bandwidth of 100 MHz, and a total of 7922 samples. To verify the generalization ability of the proposed method under different conditions, the dataset was divided according to flight trajectory type: samples corresponding to training trajectories were assigned to the training set, and samples corresponding to test trajectories were assigned to the test set, with a sample ratio of approximately 1:1 between the training and test sets.

[0100] II. Simulation Content and Results To verify the performance advantages of the method of the present invention, three baseline methods were set up for comparative experiments. All baseline methods used the same training set, test set and hardware and software environment as the present invention. Baseline 1: CNN single-station recognition method, consisting of 3 convolutional layers and 3 fully connected layers. The number of channels in the 3 convolutional layers are 16, 32 and 64 respectively, the kernel size is 1×5 and the stride is 1. The number of neurons in the fully connected layers are 512, 144 and 8 respectively. Baseline 2: A CNN-based feature concatenation and fusion method is used to achieve multi-station feature fusion by concatenating the features extracted from each station and inputting them into a classifier based on a single-station CNN model. Baseline 3: The multi-view feature fusion method refers to the distributed radar multi-view feature fusion method proposed by Xi'an Jiaotong University in its patent application "Target Recognition Method and System Based on Distributed Radar Dual-View Attention Feature Fusion" (Patent Application No.: CN202410563379.0, Publication No.: CN118334319A), which is referred to here as the multi-view feature fusion method. This method consists of a weight-sharing convolutional network module, a view attention module, a feature fusion module, a feature attention module, a temporal network module, and a classification and recognition module. The convolutional network module consists of three convolutional layers and one pooling layer. The kernel sizes of the three convolutional layers are 7, 5, and 3, respectively, with a stride of 2 for each layer. The number of channels is 32, 64, and 128, respectively. The pooling layer has a kernel size of 3 and a stride of 2. The viewpoint attention module performs spatial attention weighting on the feature maps of each viewpoint. The feature fusion module concatenates the weighted features of the three viewpoints along the channel dimension. The feature attention module generates channel weights and weighted fused features by using global average pooling and global max pooling followed by a shared multilayer perceptron. The temporal network module uses GRU to capture temporal correlations. The classification and recognition module consists of an input layer, a hidden layer, and an output layer.

[0101] Table 1 shows a comparison of the test set recognition accuracy of the method of this invention with that of three baseline methods: Table 1 Comparison of target recognition accuracy of different methods

[0102] The method of this invention identifies the confusion matrix on the test set, such as... Figure 8 As shown, the t-SNE feature distribution is visualized as follows: Figure 9 As shown.

[0103] The comparison results show that the method of the present invention can effectively improve the target recognition accuracy in distributed radar scenarios and has good inter-class separability.

[0104] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications are also considered to be within the scope of protection of the present invention.

Claims

1. A distributed radar target recognition method based on bistatic angle prior and feature decoupling, applied to target recognition in a single-transmitter, multiple-receiver distributed radar system, characterized in that... Includes the following steps: Step S1: Construct a multi-station HRRP dataset and corresponding bistatic angle information, and perform preprocessing on each. Step S2: Extract the basic feature vectors of each site through the multi-scale convolutional feature extraction module with cross-site full parameter sharing; Step S3: Generate the observation geometric contribution weights for each station through the angle scalar gating module; Step S4: Extract common target features independent of the observation angle using the common feature extraction module; Step S5: Extract the site-specific features that are geometrically related to the observations using the individual feature extraction module; Step S6: Fuse common features and individual features, and combine angle scalar weights to complete multi-station adaptive weighted aggregation to obtain global fused features; Step S7: Input the global fusion features into the classifier and output the target category prediction result; Step S8: Construct a joint loss function consisting of main classification loss, commonality constraint loss, and difference constraint loss, and combine it with a dynamic weighting mechanism to perform end-to-end training and optimization of the network.

2. The distributed radar target recognition method based on bistatic angle prior and feature decoupling as described in claim 1, characterized in that, The preprocessing in step S1 specifically includes: Perform centroid alignment and amplitude normalization on HRRP sequentially: calculate the centroid position of the sequence and move the target centroid to the center of the distance window through cyclic shifting, and then linearly map the sequence amplitude to the 0~1 interval; Linear normalization is performed on the bistatic angle information: the four-dimensional angle vector containing the incident elevation angle, incident azimuth angle, exit elevation angle, and exit azimuth angle is linearly mapped to the interval [-1,1].

3. The distributed radar target recognition method based on bistatic angle prior and feature decoupling as described in claim 1, characterized in that, The multi-scale convolutional feature extraction module in step S2 adopts a two-stage multi-scale convolution and residual cascade structure: The first level contains three parallel one-dimensional convolutional branches with kernel sizes of 3, 5, and 7, which extract short-range details, mid-range structure, and long-range contextual features, respectively. After the branch outputs are concatenated, they are batch normalized and ReLU activation function to output the first-level multi-scale features, and then downsampling is completed by the first-level residual block and the max pooling layer with a stride of 2. The downsampled features are sequentially input into the second-level multi-scale convolutional layer and the second-level residual block to extract high-level semantic features, and finally compressed into a fixed-length basic feature vector through an adaptive average pooling layer.

4. The distributed radar target recognition method based on bistatic angle prior and feature decoupling as described in claim 1, characterized in that, The angle scalar gating module in step S3 shares all parameters across stations and consists of two fully connected layers and a Sigmoid activation function. It takes a normalized angle vector of a single station as input and outputs a station-level scalar weight in the range of 0 to 1, which is used to characterize the overall contribution of the corresponding station to the recognition result under the current observation geometry.

5. The distributed radar target recognition method based on bistatic angle prior and feature decoupling as described in claim 1, characterized in that, Step S4 specifically involves: The basic features of each station are stacked into a feature tensor according to the station dimension. A multi-head self-attention mechanism is used to realize deep cross-station information interaction. After the outputs of each attention head are concatenated, the original feature dimension is restored by output mapping. Then, residual connection is performed with the input features and layer normalization is performed to obtain the inherent common features of the target that are independent of the observation angle.

6. The distributed radar target recognition method based on bistatic angle prior and feature decoupling as described in claim 1, characterized in that, Step S5 specifically involves: The normalized angle vector is passed through a fully connected network and a hyperbolic tangent activation function to generate a dimension-level scaling factor that is consistent with the dimension of the basic feature. The scaling factor is then multiplied element-wise with the basic feature to complete the angle-driven feature geometric recalibration. The recalibrated features and angle vectors are concatenated along the channel dimension and input into a fully connected personalized feature decoder, which outputs site-specific features consistent with the basic feature dimension.

7. The distributed radar target recognition method based on bistatic angle prior and feature decoupling as described in claim 1, characterized in that, Step S6 specifically involves: The common and unique features of a single site are concatenated along the channel dimension, and a lightweight fusion layer consisting of linear transformation, layer normalization, ReLU activation and Dropout is used to complete the deep interaction of features. Using the site-level scalar weights generated in step S3, the multi-site fusion features are weighted, summed, and normalized to obtain the global fusion features, expressed as: ； Where S is the total number of receiving stations. For the first s scalar weights of each receiving station For the first s The single-station fusion characteristics of each receiving station.

8. The distributed radar target recognition method based on bistatic angle prior and feature decoupling as described in claim 1, characterized in that, The expression for the joint loss function in step S8 is: ； in, The primary classification loss is adopted, using the cross-entropy loss function. The common constraint loss is used to align the distribution of common features across different sites; This is a difference-constrained loss, used to constrain the independence of representations of common and individual characteristics; , These are dynamic weighting coefficients.

9. The method as described in claim 8, characterized in that, The common constraint loss is constructed based on the difference in central moments, specifically as follows: Pairwise comparisons of the central moments of common characteristics at each site are performed, and the distribution differences are measured using the second-order norm, expressed as: ； Where S is the total number of receiving stations, and K is the order of the central moment. Representation of features x of k The central moment of the first order, This represents the number of pairs of sites.

10. The method as described in claim 8, characterized in that, The difference constraint loss is constructed based on cosine similarity, specifically as follows: Calculate the absolute value of the cosine similarity between common and unique features within a single site, using the mean of all sites as a constraint term. The expression is: ； in, S Indicates the total number of receiving stations; It is an L2 norm; This is the numerical stability coefficient, with a value of [value missing]. .

11. The method as described in claim 8, characterized in that, The dynamic weight coefficients are adaptively adjusted according to the training process and the loss magnitude. The adjustment strategy is as follows: ； ； Among them, the common constraint loss target proportion coefficient The typical value is 0.2, representing the dynamic weight of the common constraint loss. The value range is limited to [0, 0.1]; the difference constraint loss target proportion coefficient. The typical value is 0.1, representing the dynamic weight of the difference constraint loss. The value range is limited to [0, 0.05].

Images