HRRP Fusion Identification Method and Device Based on CPSA-Conformer

By using a CPSA-Conformer-based method, deep learning is employed to automatically extract the HRRP and spectral amplitude features of radar echo signals and perform weighted fusion. This addresses the shortcomings of traditional radar target recognition methods in complex backgrounds, achieving higher recognition accuracy and adaptability.

CN118228193BActive Publication Date: 2026-11-10XIDIAN UNIV
View PDF 1 Cites 0 Cited by

Patent Information

Application Number
CN202410377662.4
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-03-29
Publication Date
2026-11-10
Estimated Expiration
2044-03-29

AI Technical Summary

Technical Problem

In existing technologies, radar target recognition methods have limitations when facing complex backgrounds. Traditional feature fusion methods fail to fully exploit the complementary information between features from different domains, resulting in recognition quality that does not meet requirements.

Method used

The method based on CPSA-Conformer is adopted to extract HRRP and spectral amplitude features through parallel feature extraction branches, and perform weighted feature fusion. Deep learning is used to automatically learn deep features and dynamically generate weights to adaptively adjust the contribution of each feature.

Benefits of technology

It improves the accuracy and adaptability of target recognition, enhances the model's recognition performance in complex battlefields, and reduces the false positive rate.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118228193B_ABST
    Figure CN118228193B_ABST
Patent Text Reader

Abstract

This invention provides a HRRP fusion identification method and apparatus based on CPSA-Conformer, comprising: acquiring echo signals from a target feedback from a radar system; inputting the echo signals into a trained feature fusion identification model, so that the feature fusion identification model extracts HRRP features and spectral amplitude features of the echo signals through two parallel feature extraction branches; performing weighted feature fusion of the HRRP features and the spectral amplitude features to obtain fused features; and classifying the fused features to obtain the target's model number. This invention addresses the problem that traditional target identification methods cannot fully exploit the complementary information between features in different domains. By utilizing the time-domain and frequency-domain features of HRRP and combining deep learning for feature extraction and fusion, it achieves in-depth mining of features in different domains, reduces the false positive rate, and improves identification performance.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention belongs to the field of radar target recognition technology, specifically relating to an HRRP fusion recognition method and device based on CPSA-Conformer. Background Technology

[0002] Radar target identification refers to determining the type of a target using radar echo signals. However, traditional target identification methods are limited by their reliance on the target's physical characteristics, such as size, shape, and radar cross-section. This approach has limitations when dealing with complex backgrounds. A high-resolution range profile (HRRP) is the vector sum of the projections of the target's scattering sub-echoes onto the radar line of sight. The intensity of each range unit's sub-echo can reflect the distribution of the scattering cross-section of different target components on the radar line of sight, containing rich physical structural information about the target, which is very helpful for target identification tasks.

[0003] With the continuous advancement of technology, information fusion technology has gradually entered the field of radar target recognition. It analyzes and synthesizes multi-source information acquired by radar, overcoming the limitations of a single information source and improving the robustness of the system. In the field of radar information fusion processing, based on the position of the fusion module in the information flow, fusion methods can be divided into data-level fusion, feature-level fusion, and decision-level fusion. Data-level fusion refers to integrating radar signals from different stations or different bands to improve target detection capabilities. Feature-level fusion utilizes multiple features of target echoes and combines them through a series of computational rules to generate more discriminative fused features. It can comprehensively utilize the advantages of different features, thereby more comprehensively describing and identifying targets. Traditional feature fusion methods include feature fusion based on linear discriminant analysis and feature fusion based on canonical correlation analysis. Decision-level fusion analyzes and processes the output results from multiple classifiers or decision units, using specific standards and methods for calculation and evaluation to obtain more accurate and reliable recognition results. It can overcome the limitations that may exist with a single classifier, producing more reliable final recognition results through comprehensive analysis of multiple results, thus helping to improve the performance of the recognition system. Researchers often use DS evidence theory and random forest-based methods for decision fusion.

[0004] The aforementioned fusion recognition methods typically target manually extracted features, but these features often have limitations. In recent years, deep learning technology has been introduced into the field of target recognition. Deep learning automatically learns deep features through neural networks, thereby extracting target features more comprehensively. Feature fusion based on deep learning usually includes feature concatenation and feature addition, while decision fusion often employs methods such as weighted decision-making.

[0005] In their paper "Fusing Feature Recognition of One-Dimensional Range Images from High-Resolution Radar," Hu Yulan et al. proposed a multi-feature fusion method for recognition, addressing the translation sensitivity of HRRP (High-Resolution Radar Target Recognition). They used PCA (Principal Component Analysis) to reduce the dimensionality of three translation-invariant features (power spectrum features, amplitude spectrum difference features, and central moment features), then concatenated and fused these features, and employed a support vector machine algorithm for recognition. This approach overcame the translation sensitivity problem while improving recognition performance to some extent. However, like traditional methods, this paper relies on feature concatenation or feature addition, failing to fully exploit the complementary information between features from different domains. Therefore, the recognition quality still does not meet the requirements for radar target recognition. Summary of the Invention

[0006] To address the aforementioned problems in the existing technology, this invention provides an HRRP fusion identification method and apparatus based on CPSA-Conformer. The technical problem to be solved by this invention is achieved through the following technical solution:

[0007] In a first aspect, the present invention provides an HRRP fusion identification method based on CPSA-Conformer, comprising:

[0008] S100, acquires the echo signal fed back by the target from the radar system;

[0009] S200, the echo signal is input into the trained feature fusion recognition model, so that the feature fusion recognition model extracts the HRRP feature and spectral amplitude feature of the echo signal through two parallel feature extraction branches; the HRRP feature and the spectral amplitude feature are weighted and fused to obtain the fused feature; the fused feature is classified to obtain the model of the target.

[0010] Secondly, the present invention provides an HRRP fusion identification device based on CPSA-Conformer, comprising:

[0011] The transmit / receive module is configured to acquire echo signals from the target feedback from the radar system;

[0012] The online identification module is configured to input the echo signal into a trained feature fusion identification model, so that the feature fusion identification model extracts the HRRP feature and spectral amplitude feature of the echo signal through two parallel feature extraction branches; performs weighted feature fusion of the HRRP feature and the spectral amplitude feature to obtain the fused feature; and classifies the target to obtain its model number based on the fused feature.

[0013] The present invention has at least the following beneficial effects:

[0014] (1) This invention uses deep learning to extract features of HRRP and spectral amplitude, automatically learns the deep features of echo, avoids the bias that may be introduced by manual feature extraction through end-to-end learning, improves the accuracy of target recognition, and its adaptability and generalization ability make the model more universal and able to cope with different types of targets.

[0015] (2) This invention improves target recognition performance by weighted fusion of multi-scale HRRP features and spectral amplitude features. By dynamically generating weights, the contribution of each feature in the fusion process is adaptively adjusted, thereby improving the performance of the recognition model. Feature weighted fusion overcomes the limitations of single features, prioritizing more informative features during recognition, thus improving the accuracy of target recognition and enhancing the model's adaptability in complex battlefield conditions.

[0016] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating an HRRP fusion identification method based on CPSA-Conformer provided by the present invention;

[0018] Figure 2 This is a schematic diagram illustrating the training and testing process of the feature fusion recognition model provided by the present invention;

[0019] Figure 3 This is a schematic diagram of HRRP feature extraction based on CPSA provided by the present invention;

[0020] Figure 4 This is a schematic diagram of spectral amplitude feature extraction based on CPSA provided by the present invention;

[0021] Figure 5 This is the overall architecture diagram of the CPSA-Conformer model provided by this invention. Detailed Implementation

[0022] The present invention will be further described in detail below with reference to specific embodiments, but the implementation of the present invention is not limited thereto.

[0023] To address the shortcomings of the existing technologies, this invention proposes an HRRP fusion recognition method based on CPSA-Conformer. This method solves the problem that most existing fusion recognition methods use feature splicing or feature addition, which cannot fully explore the complementary information between features from different domains.

[0024] Firstly, such as Figure 1 As shown, this invention provides an HRRP fusion identification method based on CPSA-Conformer, including:

[0025] S100, acquires the echo signal fed back by the target from the radar system;

[0026] S200, the echo signal is input into the trained feature fusion recognition model, so that the feature fusion recognition model extracts the HRRP feature and spectral amplitude feature of the echo signal through two parallel feature extraction branches; the HRRP feature and the spectral amplitude feature are weighted and fused to obtain the fused feature; the fused feature is classified to obtain the model of the target.

[0027] The feature fusion recognition model includes a high-resolution HRRP feature extraction branch, a spectral amplitude feature extraction branch, and a weighted feature fusion network. The high-resolution HRRP feature extraction branch and the spectral amplitude feature extraction branch run in parallel, and their outputs are both input to the weighted feature fusion network. The high-resolution HRRP feature extraction branch is used to extract the HRRP features of the input signal to the feature fusion recognition model. The spectral amplitude feature extraction branch is used to extract the spectral amplitude features of the input signal. The weighted feature fusion network is used to fuse the HRRP features and the spectral amplitude features to obtain fused features and output the classification result of the fused features.

[0028] refer to Figure 2 The training process of the feature fusion recognition model of the present invention includes:

[0029] Step a: Obtain M real HRRP data sets and divide them into training datasets and test datasets. Each training dataset and test dataset contains C categories, where each category represents the type or specific model of the target.

[0030] Step b: Train the feature fusion recognition model using the training set to obtain the trained feature fusion recognition model, and test the trained feature fusion recognition model using the test dataset to select the well-trained feature fusion recognition model.

[0031] This invention trains a feature fusion recognition model using a training set. The test echo signal to be recognized is input into the trained model to obtain the test results. If the test results are unsatisfactory, the recognition model parameters can be adjusted until a suitable feature fusion model is selected as the trained feature fusion recognition model.

[0032] Combination Figures 3-5The high-resolution HRRP feature extraction branch includes a first input layer, a first convolutional layer, a first pooling layer, a first activation layer, a first PSA module, a second convolutional layer, a second pooling layer, a second activation layer, a second PSA module, a third convolutional layer, a third pooling layer, a third activation layer, a third PSA module, a first fully connected layer, and a second fully connected layer, connected in sequence.

[0033] The first convolutional layer has 1 input channel, and the first to third convolutional layers have 32, 64, and 128 convolutional kernels, respectively. The kernel size is 1×3, and the kernel stride is 1. The first to third PSA modules have 32, 64, and 128 convolutional kernels, respectively, and each PSA module contains four convolutional kernels of sizes 1×3, 1×5, 1×7, and 1×9. The pooling strategy of the first to third pooling layers is max pooling, the pooling kernel size is 1×2, and the pooling kernel stride is 2. The first to third activation layers all use the ReLU activation function.

[0034] Each PSA module divides the number of channels into G groups, with a kernel size of 1×K and a kernel stride of 1; where K∈Set={1,3,5,7,9…}. G is the number of groups, and K is the kernel size.

[0035] The feature map of the i-th convolutional kernel in each PSA module is represented as:

[0036] F i =Conv(1×k) i G i (X) i = 0, 1, 2…S-1;

[0037] Where, k i =2×(i+1)+1 represents the size of the i-th convolutional kernel. Let S represent the size of the i-th group, and F represent the number of groups. i Feature maps representing different scales;

[0038] The feature map output by each PSA module is represented as Out = Cat([Y0, Y1, ..., Y...). S-1 Each PSA module is concatenated to obtain a multi-scale feature map with C channels, represented as:

[0039] F = Cat([F0,F1,…,F) S-1 ]);

[0040] Where F∈R C×1×W For a multi-scale feature map with channel C, based on the obtained multi-scale feature map F... iWeight information is extracted through the channel attention mechanism. The extracted attention weights can be represented as follows:

[0041] Z i =SEWeight(F i ), i = 0, 1, 2…S-1;

[0042] Z i ∈R C ' ×1×1 To handle attention weights, the SE module is used to obtain weights from feature maps at different scales. By fusing the cross-dimensional vectors, the attention vector for the entire multi-scale channel is obtained:

[0043]

[0044] After obtaining the overall attention vector, weights are generated using Softmax:

[0045]

[0046] By concatenating the weights, we obtain the attention vector for the entire channel:

[0047]

[0048] By multiplying the multi-scale features by their corresponding weights, the fused feature map is obtained:

[0049] Y i =F i ⊙att i i = 1, 2, 3, ..., S-1;

[0050] Finally, by concatenating the data, we obtain the complete feature map output by each PSA module:

[0051] Out = Cat([Y0,Y1,…,Y) S-1 ]).

[0052] refer to Figure 4The spectral amplitude feature extraction branch of this invention includes a preprocessing module and a Conformer encoder connected in sequence. The preprocessing module is used to segment the input signal to obtain multiple spectral amplitude subsequences, and then to encode each spectral amplitude subsequence to obtain an encoded sequence. The input encoding includes linear encoding and positional encoding. The Conformer encoder is used to encode the encoded sequence to obtain spectral amplitude features. The input signal is a spectral amplitude sequence. The structure of the Conformer encoder includes: a first normalization layer, a multi-head attention mechanism module based on relative positional encoding, a fourth convolutional layer, a second normalization layer, and a feedforward layer. The number of distance units for spectral amplitude sequence segmentation in the preprocessing module is set to 32, resulting in 8 spectral amplitude subsequences, a linear encoding dimension of 128, a Conformer encoder with 4 heads, a multi-head attention mechanism module with 4 heads, and a Conformer feedforward layer output dimension of 512.

[0053] The multi-head attention mechanism module divides the position-encoded output vector into h groups of heads, and multiplies each feature dimension by a randomly initialized matrix W. q W k W v We obtain three matrices, Q, K, and V, representing the query, key, and value, respectively, as follows:

[0054] Q i =XW i Q ,K i =XW i K V = XW i V ;

[0055] Each set of query Q, key K, and value V is obtained by scaling the dot product attention:

[0056]

[0057] Where, d k The dimension of the input data is represented as follows: The attention score calculated for each head is expressed as:

[0058] head i =Attention(Q) i ,K i V i );

[0059] The outputs of multiple attention heads are concatenated together and coupled with the weight matrix W. o Multiplying these together yields the output of the multi-head attention mechanism module, represented as:

[0060] MHSA(Q,K,V)=Concat(head0,...,head3)W o .

[0061] Combination Figures 3-5 The weighted feature fusion network of the present invention comprises three parts: an HRRP branch feature processing module, a Conformer branch feature processing module, and a fusion module.

[0062] The HRRP branch feature processing module is used to process the HRRP features of the input signal of the feature fusion recognition model, and the Conformer branch feature processing module is used to process the spectral amplitude features of the input signal. The fusion module is used to calculate the weights when fusing the HRRP features and the spectral amplitude features, and to fuse the HRRP features and the spectral amplitude features according to the weights corresponding to the two branches to obtain fused features, and to classify the fused features to obtain the recognition result of the input signal.

[0063] The HRRP branch feature processing module includes a fifth convolutional layer, a fourth pooling layer, a sixth convolutional layer, a fifth pooling layer, and a third fully connected layer connected in sequence. The fifth pooling layer and the third fully connected layer are connected by a residual connection. The Conformer branch feature processing module includes a global average pooling layer, a third normalization layer, and a fully connected layer connected in sequence. The third normalization layer is followed by a GELU activation function. The fusion module includes a feature concatenation layer, a linear weight generation layer, a fourth fully connected layer, and a fifth fully connected layer connected in sequence. The fifth fully connected layer is followed by an activation function.

[0064] As an optional embodiment of the present invention, S200 includes:

[0065] S210, the echo signal is input into the trained feature fusion recognition model so that the feature fusion recognition model extracts the HRRP feature f1 and the spectral amplitude feature f2 of the echo signal through two parallel feature extraction branches;

[0066] S220, using the respective weights of the HRRP feature f1 and the spectral amplitude feature f2, the HRRP feature and the spectral amplitude feature are fused to obtain a fused feature; the weights of the HRRP feature f1 and the spectral amplitude feature f2 are respectively expressed as:

[0067]

[0068]

[0069] The fusion feature is represented as: coal_feature = w1f1 + w2f2; where coal_feature is the fusion feature;

[0070] S230, the model number of the target is obtained by classifying the fused features.

[0071] Secondly, the present invention provides an HRRP fusion identification device based on CPSA-Conformer, comprising:

[0072] The transmit / receive module is configured to acquire echo signals from the target feedback from the radar system;

[0073] The online identification module is configured to input the echo signal into a trained feature fusion identification model, so that the feature fusion identification model extracts the HRRP feature and spectral amplitude feature of the echo signal through two parallel feature extraction branches; performs weighted feature fusion of the HRRP feature and the spectral amplitude feature to obtain the fused feature; and classifies the target to obtain its model number based on the fused feature.

[0074] This invention provides a HRRP fusion identification method and apparatus based on CPSA-Conformer, comprising: transmitting a signal to a target using a radar system and receiving the echo signal; inputting the echo signal into a trained feature fusion identification model, so that the feature fusion identification model extracts the HRRP features and spectral amplitude features of the echo signal through two parallel feature extraction branches; performing weighted feature fusion of the HRRP features and the spectral amplitude features to obtain fused features; and classifying the fused features to obtain the target's model number. This invention addresses the problem that traditional target identification methods cannot fully exploit the complementary information between features in different domains. By utilizing the time-domain and frequency-domain features of HRRP and combining deep learning for feature extraction and fusion, it achieves in-depth mining of features in different domains, reduces the false positive rate, and improves identification performance.

[0075] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0076] Although this application has been described herein in conjunction with various embodiments, those skilled in the art will understand and implement other variations of the disclosed embodiments by reviewing the accompanying drawings, the disclosure, and the appended claims in carrying out the claimed application. In the claims, the word "comprising" does not exclude other components or steps, and "a" or "an" does not exclude a plurality.

[0077] The above description, in conjunction with specific preferred embodiments, provides a further detailed explanation of the present invention. It should not be construed that the specific implementation of the present invention is limited to these descriptions. For those skilled in the art, various simple deductions or substitutions can be made without departing from the concept of the present invention, and all such modifications and substitutions should be considered within the scope of protection of the present invention.

Claims

1. A HRRP fusion identification method based on CPSA-Conformer, characterized in that, include: S100, acquires the echo signal fed back by the target from the radar system; S200, the echo signal is input into the trained feature fusion recognition model, so that the feature fusion recognition model extracts the HRRP feature and spectral amplitude feature of the echo signal through two parallel feature extraction branches; the HRRP feature and the spectral amplitude feature are weighted and fused to obtain the fused feature; The target's model number is obtained by classifying the fused features; The feature fusion recognition model includes: a high-resolution HRRP feature extraction branch, a spectral amplitude feature extraction branch, and a weighted feature fusion network; the high-resolution HRRP feature extraction branch and the spectral amplitude feature extraction branch run in parallel, and their output results are both input into the weighted feature fusion network; The high-resolution HRRP feature extraction branch is used to extract the HRRP features of the input signal into the feature fusion recognition model; the spectral amplitude feature extraction branch is used to extract the spectral amplitude features of the input signal; and the weighted feature fusion network is used to fuse the HRRP features and the spectral amplitude features to obtain fused features and output the classification result of the fused features. The high-resolution HRRP feature extraction branch includes a first input layer, a first convolutional layer, a first pooling layer, a first activation layer, a first PSA module, a second convolutional layer, a second pooling layer, a second activation layer, a second PSA module, a third convolutional layer, a third pooling layer, a third activation layer, a third PSA module, a first fully connected layer, and a second fully connected layer, connected in sequence. In this configuration, the first convolutional layer has 1 input channel; the first to third convolutional layers have 32, 64, and 128 convolutional kernels, respectively, with each kernel size set to 1×3 and each kernel stride set to 1; the first to third PSA modules have 32, 64, and 128 convolutional kernels, respectively, with each PSA module containing four convolutional kernels of sizes 1×3, 1×5, 1×7, and 1×9; the first to third pooling layers all use max pooling, with each pooling kernel size set to 1×2 and each pooling kernel stride set to 2; and the first to third activation layers all use the ReLU activation function. The spectral amplitude feature extraction branch includes a preprocessing module and a Conformer encoder connected in sequence. The preprocessing module is used to segment the input signal to obtain multiple spectral amplitude subsequences, and then to perform input encoding on each spectral amplitude subsequence to obtain an encoded sequence. The input encoding includes linear encoding and positional encoding. The Conformer encoder is used to perform encoding processing on the encoded sequence to obtain spectral amplitude features. The structure of the Conformer encoder includes: a first normalization layer, a multi-head attention mechanism module based on relative position encoding, a fourth convolutional layer, a second normalization layer, and a feedforward layer; In this configuration, the number of distance units for spectral amplitude sequence segmentation in the preprocessing module is set to 32, resulting in 8 spectral amplitude subsequences, a linear encoding dimension of 128, a Conformer encoder of 4, a multi-head attention mechanism module of 4 heads, and a Conformer feedforward layer output dimension of 512.

2. The HRRP fusion identification method based on CPSA-Conformer according to claim 1, characterized in that, The multi-head attention mechanism module divides the position-encoded output vector into h groups of heads, and multiplies each feature dimension by a randomly initialized matrix. We obtain three matrices, Q, K, and V, representing the query, key, and value, respectively, as follows: ; Each set of query Q, key K, and value V is obtained by scaling the dot product attention: ; in, The dimension of the input data is represented as follows: The attention score calculated for each head is expressed as: ; The outputs of multiple attention heads are concatenated together and coupled with the weight matrix. Multiplying these together yields the output of the multi-head attention mechanism module, represented as: 。 3. The HRRP fusion identification method based on CPSA-Conformer according to claim 1, characterized in that, The weighted feature fusion network consists of three parts: the HRRP branch feature processing module, the Conformer branch feature processing module, and the fusion module. The HRRP branch feature processing module is used to process the HRRP features of the input signal of the feature fusion recognition model, and the Conformer branch feature processing module is used to process the spectral amplitude features of the input signal. The fusion module is used to calculate the weights when fusing the HRRP features and the spectral amplitude features, and to fuse the HRRP features and the spectral amplitude features according to the weights corresponding to the two branches to obtain fused features, and to classify the fused features to obtain the recognition result of the input signal.

4. The HRRP fusion identification method based on CPSA-Conformer according to claim 3, characterized in that, The HRRP branch feature processing module includes a fifth convolutional layer, a fourth pooling layer, a sixth convolutional layer, a fifth pooling layer, and a third fully connected layer connected in sequence. The fifth pooling layer and the third fully connected layer are connected by a residual connection. The Conformer branch feature processing module includes a global average pooling layer, a third normalization layer, and a fully connected layer connected in sequence. The third normalization layer is followed by a GELU activation function. The fusion module includes a feature concatenation layer, a linear weight generation layer, a fourth fully connected layer, and a fifth fully connected layer connected in sequence. The fifth fully connected layer is followed by an activation function.

5. The HRRP fusion identification method based on CPSA-Conformer according to claim 1, characterized in that, The training process of the feature fusion recognition model includes: Step a: Obtain M real HRRP data sets and divide them into training datasets and test datasets. Each training dataset and test dataset contains C categories, where each category represents the type or specific model of the target. Step b: Train the feature fusion recognition model using the training dataset to obtain the trained feature fusion recognition model, and test the trained feature fusion recognition model using the test dataset to select the well-trained feature fusion recognition model.

6. The HRRP fusion identification method based on CPSA-Conformer according to claim 1, characterized in that, S200 includes: S210, the echo signal is input into the trained feature fusion recognition model, so that the feature fusion recognition model extracts the HRRP features of the echo signal through two parallel feature extraction branches. and spectral amplitude characteristics ; S220, utilizing the HRRP feature With the aforementioned spectral amplitude characteristics The HRRP features are fused with the spectral amplitude features based on their respective weights to obtain fused features; the HRRP features and the spectral amplitude characteristics The corresponding weights are expressed as follows: ; ; The fusion feature is represented as follows: ;in, Features of fusion; S230, the model number of the target is obtained by classifying the fused features.

7. An HRRP fusion identification device based on CPSA-Conformer, characterized in that, The HRRP fusion identification method based on CPSA-Conformer as described in any one of claims 1 to 6 is used to implement the HRRP fusion identification method based on CPSA-Conformer, wherein the HRRP fusion identification based on CPSA-Conformer includes: The transmit / receive module is configured to acquire echo signals from the target feedback from the radar system; The online identification module is configured to input the echo signal into a trained feature fusion identification model, so that the feature fusion identification model extracts the HRRP feature and spectral amplitude feature of the echo signal through two parallel feature extraction branches; performs weighted feature fusion of the HRRP feature and the spectral amplitude feature to obtain the fused feature; and classifies the target to obtain its model number based on the fused feature.

Citation Information

Patent Citations

  • Two-stage weak supervision segmentation method for concrete structure defects

    CN120356217A