Radar target recognition method and device under non-uniform interference based on attention network

Abstract

The application relates to a radar target recognition method and device based on an attention network under non-uniform interference, wherein the method comprises the following steps: by utilizing the characteristic that the power distribution of interference in the frequency domain is uneven in a composite interference scene, taking the frequency domain power of an interference echo as auxiliary information, constructing a frequency attention mechanism, suppressing the interference frequency band by attention weighting in the frequency domain, and enhancing the target prior; and combining a joint loss function design of a local MSE loss and a cross-entropy loss, the feature discrimination ability and the recognition accuracy of the model are enhanced under the condition of non-uniform composite interference. Thus, the problems that the robustness, generalization ability and recognition accuracy of an existing HRRP target recognition model based on deep learning are poor in a complex electromagnetic environment with strong adversarial and composite active interference are solved.

Description

Technical Field

[0001] This application relates to the field of radar target recognition technology, and in particular to a radar target recognition method and apparatus based on attention networks under non-uniform interference. Background Technology

[0002] Radar target recognition, as a core function of radar systems, has crucial application value in many fields such as traffic control and remote sensing monitoring. High-resolution range profiles (HRRPs) have become one of the key technologies for achieving automatic target recognition (ATR) due to their significant advantages, including high imaging efficiency, relatively low processing complexity, and the inclusion of fine-structure scattering information about the target.

[0003] Currently, HRRP target recognition methods can be broadly categorized into two main types: traditional methods and deep learning-based methods. The core research focus of traditional methods lies in designing robust and discriminative feature extractors (such as statistical moments, entropy features, and time-frequency analysis features). Although the features extracted by these methods typically have clear physical meaning and strong interpretability, their performance is highly dependent on the researchers' prior knowledge and design experience, and the feature extraction process often involves high computational complexity. This, to some extent, limits their widespread application in battlefield conditions facing complex environments and electromagnetic interference.

[0004] Currently, research on high-resolution radar range image target recognition based on deep learning mainly focuses on conventional electromagnetic environments without electromagnetic interference or idealized scenarios. However, the electromagnetic environment of modern battlefields is highly complex and adversarial, with a wide variety of active jamming methods. These jamming methods can be mainly categorized into two types: deceptive jamming (such as sample-and-forward jamming), which generates false targets by copying, modulating, and forwarding radar signals; and suppression jamming (including narrowband targeting jamming and broadband jamming), which aims to overwhelm the echo of the real target with high-power noise. More seriously, multiple jamming modes are often used in combination, exhibiting complex characteristics of non-uniform amplitude (i.e., uneven distribution of jamming energy in the frequency domain) in the spectrum. This phenomenon is caused by two factors: firstly, multiple suppression jamming opportunities generate different intensities of interference at different frequency bands; secondly, it stems from the fact that sample-and-forward jamming is not fully distributed in the time domain, resulting in different interference intensities received at different frequencies under agile coherent waveforms, exhibiting non-uniform amplitude. This type of complex active electromagnetic interference can have multi-dimensional and destructive effects on the target's HRRP, as described below: 1. Target echo flooding and distortion: Suppression jamming significantly reduces the signal-to-noise ratio (SNR) of the target echo, causing the key scattering center of the real target's HRRP to be flooded by noise or distorted in amplitude, or the overall structure to be severely distorted. 2. Feature space confusion: Interference-induced HRRP distortion causes the distribution of HRRP of different targets to overlap and become confused in the feature space, which greatly increases the difficulty for the classifier to effectively distinguish them.

[0005] Therefore, in complex electromagnetic environments with strong adversarial forces and complex active interference, the existing deep learning-based HRRP target recognition models have poor robustness, generalization ability, and recognition accuracy, which urgently need to be addressed. Summary of the Invention

[0006] This application provides a radar target recognition method and apparatus based on attention networks under non-uniform interference, in order to solve the problems of poor robustness, generalization ability and recognition accuracy of existing deep learning-based HRRP target recognition models in complex electromagnetic environments with strong adversarial and complex active interference.

[0007] The first aspect of this application provides a radar target recognition method under non-uniform interference based on an attention network, applied in the offline training stage, comprising the following steps: constructing a transmitted waveform signal model of the target radar, and performing signal preprocessing operations on the transmitted waveform signal model to construct a corresponding time-domain radar high-resolution range image; simulating preset multi-aircraft target echoes to construct a multi-aircraft interference-free simulation dataset, and generating corresponding colored noise based on a preset colored noise interference generation algorithm, and superimposing the colored noise and the time-domain radar high-resolution range image to obtain corresponding interference-containing data samples, and constructing a radar high-resolution range image simulation dataset based on the interference-containing data samples and the multi-aircraft interference-free simulation dataset; constructing a complex frequency attention network based on preset frequency-domain non-uniform interference characteristics, and establishing a radar target recognition model based on the complex frequency attention network, a preset radar high-resolution range image classifier, and a hybrid loss function, and training the radar target recognition model through the radar high-resolution range image simulation dataset, so as to perform radar target recognition operations under non-uniform interference environment using the trained radar target recognition model in the online recognition stage.

[0008] Optionally, in one embodiment of this application, the step of constructing a transmitted waveform signal model of the target radar and performing signal preprocessing operations on the transmitted waveform signal model to construct a corresponding time-domain high-resolution range image of the radar includes: constructing a corresponding received echo signal model based on the transmitted waveform signal model, and performing a Fourier transform on the received echo of the target radar according to the received echo signal model to generate a corresponding frequency-domain received echo signal; estimating the target velocity of the target under test through the target radar, and calculating the corresponding Doppler frequency based on the transmitted waveform signal model and the target velocity, and performing intra-pulse Doppler compensation and inter-pulse Doppler compensation on the frequency-domain received echo signal using the Doppler frequency to generate a corresponding compensation signal; performing frequency-domain matched filtering processing on each pulse in the compensation signal to obtain a corresponding matched filter signal, and performing same-frequency point summation and frequency-domain splicing operations on the matched filter signal to generate a corresponding synthetic broadband radar high-resolution range image spectrum; and performing an inverse Fourier transform on the synthetic broadband radar high-resolution range image spectrum to obtain the time-domain high-resolution range image of the radar.

[0009] Optionally, in one embodiment of this application, the step of simulating preset multi-aircraft target echoes to construct a multi-aircraft interference-free simulation dataset, generating corresponding colored noise based on a preset colored noise interference generation algorithm, and superimposing the colored noise and the time-domain radar high-resolution range image to obtain corresponding interference-containing data samples includes: calculating and simulating preset different types of aircraft based on preset three-dimensional electromagnetic field simulation software to obtain corresponding radar cross section data; calculating the time-domain radar high-resolution range image of each type of aircraft at different observation angles to construct the multi-aircraft interference-free simulation dataset based on the time-domain radar high-resolution range image; determining different interference types and the filter transfer function corresponding to each interference type based on the preset colored noise interference generation algorithm, adjusting the total interference energy at a preset signal-to-interference ratio based on the different interference types and the filter transfer function corresponding to each interference type to generate the colored noise; superimposing the colored noise and the time-domain radar high-resolution range image to generate a corresponding interference-containing signal, and performing a Fourier transform on the interference-containing signal to obtain the interference-containing data samples.

[0010] Optionally, in one embodiment of this application, the step of constructing a complex frequency attention network based on preset frequency domain non-uniform interference characteristics, and establishing a radar target recognition model based on the complex frequency attention network, a preset radar high-resolution range image classifier, and a hybrid loss function, includes: constructing a shared-weight multilayer perceptron through multiple fully connected layers, and constructing the complex frequency attention network based on a preset dual-channel pooling mechanism, activation function, and the shared-weight multilayer perceptron; constructing the radar high-resolution range image classifier based on preset convolutional layers, batch normalization layers, activation function layers, pooling layers, Flatten layers, and fully connected layers; and establishing the radar target recognition model based on the complex frequency attention network, the radar high-resolution range image classifier, and a hybrid loss function composed of a preset local MSE loss function and a cross-entropy loss function.

[0011] Optionally, in one embodiment of this application, training the radar target recognition model using the radar high-resolution range image simulation dataset includes: dividing the radar high-resolution range image simulation dataset into a training dataset, a validation dataset, and a test dataset based on a preset partitioning ratio; performing Z-score normalization on the complex components of each data in the training dataset, the validation dataset, and the test dataset to obtain corresponding normalized data; inputting the normalized data corresponding to the training dataset into the complex frequency attention network of the radar target recognition model to extract the time-domain radar high-resolution range image features corresponding to the normalized data through the complex frequency attention network, and performing Fourier transform and feature dimension transformation on the time-domain radar high-resolution range image features to generate corresponding target frequency domain features; based on the... A dual-pooling mechanism is used to compress and extract the target frequency domain features to generate corresponding feature vectors. These feature vectors are then concatenated, and a shared-weight multilayer perceptron is used to perform nonlinear transformation and linear projection on the concatenated feature vectors to obtain corresponding fused features. Based on the activation function, attention weights corresponding to the fused features are generated, and attention weighting is performed according to these attention weights to generate corresponding enhanced features. These enhanced features are then input into the radar high-resolution range image classifier to generate corresponding classification results. The radar target recognition model is trained based on the classification results and the hybrid loss function. The trained radar target recognition model is then fine-tuned using standardized data corresponding to the validation dataset and the test dataset to ensure that the radar target recognition model meets preset performance requirements.

[0012] A second aspect of this application provides a radar target recognition method under non-uniform interference based on an attention network, applied in the online recognition stage, comprising the following steps: acquiring a high-resolution range image of the frequency domain radar corresponding to each target under test using a target radar; inputting the high-resolution range image of the frequency domain radar corresponding to the interference into a trained radar target recognition model, generating complex attention weights corresponding to the high-resolution range image of the frequency domain radar corresponding to the interference through the complex frequency attention network of the radar target recognition model, and performing a point-by-point multiplication operation on the high-resolution range image of the frequency domain radar corresponding to the interference and the complex attention weights to obtain corresponding filtered spectral features; inputting the filtered spectral features into a radar high-resolution range image classifier, performing multi-layer nonlinear transformation on the filtered spectral features, outputting a probability distribution corresponding to each target under test, and obtaining the recognition result corresponding to each target under test based on the probability distribution.

[0013] A third aspect of this application provides a radar target recognition device under non-uniform interference based on an attention network, applied in the offline training stage, comprising: a preprocessing module for constructing a transmitted waveform signal model of a target radar and performing signal preprocessing operations on the transmitted waveform signal model to construct a corresponding time-domain high-resolution range image of the radar; a dataset construction module for simulating preset multi-type target echoes to construct a multi-type interference-free simulation dataset, and generating corresponding colored noise based on a preset colored noise interference generation algorithm, and superimposing the colored noise and the time-domain high-resolution range image of the radar to obtain corresponding interference-containing data samples, so as to construct a radar high-resolution range image simulation dataset based on the interference-containing data samples and the multi-type interference-free simulation dataset; and a training module for constructing a complex frequency attention network based on preset frequency-domain non-uniform interference characteristics, and establishing a radar target recognition model according to the complex frequency attention network, a preset radar high-resolution range image classifier, and a hybrid loss function, and training the radar target recognition model through the radar high-resolution range image simulation dataset, so as to perform radar target recognition operations under non-uniform interference environment using the trained radar target recognition model in the online recognition stage.

[0014] Optionally, in one embodiment of this application, the preprocessing module includes: a generation unit, configured to construct a corresponding received echo signal model based on the transmitted waveform signal model, and perform a Fourier transform on the received echo of the target radar according to the received echo signal model to generate a corresponding frequency domain received echo signal; a compensation unit, configured to estimate the target velocity of the target under test through the target radar, and calculate the corresponding Doppler frequency based on the transmitted waveform signal model and the target velocity, and perform intra-pulse Doppler compensation and inter-pulse Doppler compensation on the frequency domain received echo signal through the Doppler frequency to generate a corresponding compensation signal; a filtering unit, configured to perform frequency domain matched filtering processing on each pulse in the compensation signal to obtain a corresponding matched filter signal, and perform same-frequency point summation and frequency domain splicing operations on the matched filter signal to generate a corresponding synthetic broadband radar high-resolution range image spectrum; and a transformation unit, configured to perform an inverse Fourier transform on the synthetic broadband radar high-resolution range image spectrum to obtain the time domain radar high-resolution range image.

[0015] Optionally, in one embodiment of this application, the dataset construction module includes: a simulation unit, used to perform simulation calculations on preset three-dimensional electromagnetic field simulation software for different types of aircraft to obtain corresponding radar cross section data; a first construction unit, used to calculate the time-domain radar high-resolution range image of each type of aircraft at different observation angles to construct the multi-aircraft interference-free simulation dataset based on the time-domain radar high-resolution range image; an adjustment unit, used to determine different interference types and the filter transfer function corresponding to each interference type based on a preset colored noise interference generation algorithm, and adjust the total interference energy at a preset signal-to-interference ratio based on the different interference types and the filter transfer function corresponding to each interference type to generate the colored noise; and a superposition unit, used to superimpose the colored noise and the time-domain radar high-resolution range image to generate a corresponding interference-containing signal, and perform a Fourier transform on the interference-containing signal to obtain the interference-containing data sample.

[0016] Optionally, in one embodiment of this application, the training module includes: a second construction unit, configured to construct a shared-weight multilayer perceptron through multiple fully connected layers, and to construct the complex frequency attention network based on a preset dual-channel pooling mechanism, activation function, and the shared-weight multilayer perceptron; a third construction unit, configured to construct the radar high-resolution range image classifier based on preset convolutional layers, batch normalization layers, activation function layers, pooling layers, Flatten layers, and fully connected layers; and an establishment unit, configured to establish the radar target recognition model based on the complex frequency attention network, the radar high-resolution range image classifier, and a hybrid loss function composed of a preset local MSE loss function and a cross-entropy loss function.

[0017] Optionally, in one embodiment of this application, the training module further includes: a partitioning unit, configured to partition the radar high-resolution range image simulation dataset into a training dataset, a validation dataset, and a test dataset based on a preset partitioning ratio; a processing unit, configured to perform Z-score normalization on the complex components of each data in the training dataset, the validation dataset, and the test dataset to obtain corresponding normalized data; an extraction unit, configured to input the normalized data corresponding to the training dataset into the complex frequency attention network of the radar target recognition model, so as to extract the time-domain radar high-resolution range image features corresponding to the normalized data through the complex frequency attention network, and perform Fourier transform and feature dimension transformation on the time-domain radar high-resolution range image features to generate corresponding target frequency domain features; and a stitching unit, configured to perform stitching based on the dual-channel pooling machine. The system compresses and extracts the target frequency domain features to generate corresponding feature vectors, concatenates the feature vectors, and performs nonlinear transformation and linear projection on the concatenated feature vectors through the shared weight multilayer perceptron to obtain corresponding fused features. A weighting unit generates attention weights corresponding to the fused features based on the activation function, performs attention weighting operations based on the attention weights to generate corresponding enhanced features, and inputs the enhanced features into the radar high-resolution range image classifier to generate corresponding classification results. A fine-tuning unit trains the radar target recognition model based on the classification results and the hybrid loss function, and fine-tunes the trained radar target recognition model using standardized data corresponding to the validation dataset and the test dataset to ensure that the radar target recognition model meets preset performance requirements.

[0018] A fourth aspect of this application provides a radar target recognition device under non-uniform interference based on an attention network, applied in the online recognition stage, comprising: an acquisition module for acquiring a high-resolution range image of the frequency domain radar corresponding to each target under test through a target radar; an inference module for inputting the high-resolution range image of the frequency domain radar corresponding to the interference into a trained radar target recognition model, generating complex attention weights corresponding to the high-resolution range image of the frequency domain radar corresponding to the interference through the complex frequency attention network of the radar target recognition model, and performing a point-by-point multiplication operation on the high-resolution range image of the frequency domain radar corresponding to the interference and the complex attention weights to obtain corresponding filtered spectral features; and a recognition module for inputting the filtered spectral features into a radar high-resolution range image classifier, performing multi-layer nonlinear transformation on the filtered spectral features, outputting a probability distribution corresponding to each target under test, and obtaining a recognition result corresponding to each target under test based on the probability distribution.

[0019] A fifth aspect of this application provides an electronic device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the radar target recognition method under non-uniform interference based on attention networks as described in the above embodiments.

[0020] A sixth aspect of this application provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described radar target recognition method under non-uniform interference based on an attention network.

[0021] Therefore, the embodiments of this application have the following beneficial effects: The embodiments of this application construct a transmitted waveform signal model of the target radar and perform signal preprocessing operations on the transmitted waveform signal model to construct a corresponding time-domain radar high-resolution range image. Simulations are performed on preset multi-aircraft target echoes to construct a multi-aircraft interference-free simulation dataset. Based on a preset colored noise interference generation algorithm, corresponding colored noise is generated, and the colored noise and the time-domain radar high-resolution range image are superimposed to obtain corresponding interference-containing data samples. Based on the interference-containing data samples and the multi-aircraft interference-free simulation dataset, a radar high-resolution range image simulation dataset is constructed. Based on preset frequency-domain non-uniform interference characteristics, a complex frequency attention network is constructed. A radar target recognition model is established based on the complex frequency attention network, a preset radar high-resolution range image classifier, and a hybrid loss function. The radar target recognition model is trained using the radar high-resolution range image simulation dataset, and the trained radar target recognition model is used for radar target recognition operations under non-uniform interference environments during the online recognition stage. This application, through in-depth research on the distortion mechanism of HRRP under interference environments, obtains a robust deep learning target recognition strategy with strong anti-interference capabilities. This solves the problems of poor robustness, generalization ability, and recognition accuracy of existing deep learning-based HRRP target recognition models in complex electromagnetic environments with strong adversarial forces and complex active interference.

[0022] Additional aspects and advantages of this application will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of this application. Attached Figure Description

[0023] The above and / or additional aspects and advantages of this application will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein: Figure 1 This is a flowchart of a radar target recognition method based on attention networks under non-uniform interference, provided according to an embodiment of this application, applied in the offline training phase; Figure 2A schematic diagram of a signal preprocessing process is provided for one embodiment of this application; Figure 3 A schematic diagram of a simulation dataset model construction is provided as an embodiment of this application; Figure 4 A schematic diagram illustrating the construction of a non-uniform disturbance in a simulation dataset, as provided in one embodiment of this application; Figure 5 A schematic diagram of a target HRRP provided for one embodiment of this application; Figure 5 (a) is a schematic diagram of a non-interference target HRRP provided in an embodiment of this application; Figure 5 (b) is a schematic diagram of an HRRP containing an interfering target provided in an embodiment of this application; Figure 6 A schematic diagram of a target echo spectrum provided for one embodiment of this application; Figure 6 Image (a) is a schematic diagram of an interference-free target echo spectrum provided in an embodiment of this application; Figure 6 Image (b) is a schematic diagram of an echo spectrum containing interference from an embodiment of this application; Figure 7 A schematic diagram of an HRRP classifier architecture is provided for one embodiment of this application; Figure 8 A schematic diagram of an HRRP recognition algorithm framework based on interference frequency domain information-assisted whitening is provided for one embodiment of this application; Figure 9 This is a flowchart of a radar target recognition method based on attention networks under non-uniform interference, provided according to an embodiment of this application, applied in the online recognition stage; Figure 10 A schematic diagram of recognition rate and confusion matrix under a simulated dataset is provided as an embodiment of this application; Figure 11 A schematic diagram of the recognition rate and confusion matrix under a measured dataset is provided as an embodiment of this application; Figure 12 This is an example diagram of a radar target recognition device based on an attention network under non-uniform interference applied in the offline training phase, according to an embodiment of this application. Figure 13 This is an example diagram of a radar target identification device based on an attention network under non-uniform interference, applied in the online identification stage according to an embodiment of this application; Figure 14 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application.

[0024] Among them, 10- Radar target recognition device based on attention network under non-uniform interference applied to the offline training stage, 20- Radar target recognition device based on attention network under non-uniform interference applied to the online recognition stage; 101- Preprocessing module, 102- Dataset construction module, 103- Training module; 201- Acquisition module, 202- Inference module, 203- Recognition module; 1401- Memory, 1402- Processor, 1403- Communication interface. Detailed Implementation

[0025] The embodiments of this application are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain this application, and should not be construed as limiting this application.

[0026] The following describes a radar target recognition method and apparatus based on attention networks under non-uniform interference, according to embodiments of this application, with reference to the accompanying drawings. Addressing the problems mentioned in the background section, this application provides a radar target recognition method based on attention networks under non-uniform interference. In this method, a model of the transmitted waveform signal of the target radar is constructed, and signal preprocessing is performed on the transmitted waveform signal model to construct a corresponding time-domain high-resolution radar range profile. Simulations are performed on preset multi-aircraft target echoes to construct a multi-aircraft interference-free simulation dataset. Based on a preset colored noise interference generation algorithm, corresponding colored noise is generated, and the colored noise and the time-domain high-resolution radar range profile are superimposed to obtain corresponding interference-containing data samples. Based on the interference-containing data samples and the multi-aircraft interference-free simulation dataset, a radar high-resolution range profile simulation dataset is constructed. Based on preset frequency-domain non-uniform interference characteristics, a complex-frequency attention network is constructed. A radar target recognition model is established based on the complex-frequency attention network, a preset radar high-resolution range profile classifier, and a hybrid loss function. The radar target recognition model is trained using the radar high-resolution range profile simulation dataset, so that the trained radar target recognition model can be used for radar target recognition operations under non-uniform interference environments during the online recognition phase. This application delves into the distortion mechanism of HRRP under interference environments to derive a robust deep learning-based target recognition strategy with strong anti-interference capabilities. This addresses the shortcomings of existing deep learning-based HRRP target recognition models in complex electromagnetic environments with strong adversarial forces and complex active interference, such as poor robustness, generalization ability, and recognition accuracy.

[0027] Specifically, Figure 1 This is a flowchart illustrating a radar target recognition method based on attention networks under non-uniform interference, provided as an embodiment of this application, applied during the offline training phase.

[0028] like Figure 1 As shown, the radar target recognition method based on attention networks under non-uniform interference includes the following steps: In step S101, a transmission waveform signal model of the target radar is constructed, and signal preprocessing operations are performed on the transmission waveform signal model to construct the corresponding time-domain radar high-resolution range image.

[0029] The embodiments of this application first construct a radar transmission waveform signal model and obtain the corresponding raw echo signal. The raw echo signal is then input into a preset target detection algorithm to obtain the estimated velocity of the target, which is then used for velocity compensation. Secondly, the embodiments of this application can perform frequency domain matched filtering, same-frequency point accumulation, and inter-frequency point loan synthesis operations on each pulse to obtain the corresponding time-domain radar high-resolution range image, thereby providing reliable data support for downstream tasks such as identification.

[0030] Optionally, in one embodiment of this application, a transmitted waveform signal model of the target radar is constructed, and signal preprocessing operations are performed on the transmitted waveform signal model to construct a corresponding time-domain high-resolution range image of the radar. This includes: constructing a corresponding received echo signal model based on the transmitted waveform signal model, and performing a Fourier transform on the received echo of the target radar according to the received echo signal model to generate a corresponding frequency-domain received echo signal; estimating the target velocity of the target under test through the target radar, and calculating the corresponding Doppler frequency based on the transmitted waveform signal model and the target velocity, and performing intra-pulse Doppler compensation and inter-pulse Doppler compensation on the frequency-domain received echo signal using the Doppler frequency to generate a corresponding compensation signal; performing frequency-domain matched filtering processing on each pulse in the compensation signal to obtain a corresponding matched filter signal, and performing same-frequency point summation and frequency-domain splicing operations on the matched filter signal to generate a corresponding synthetic broadband radar high-resolution range image spectrum; and performing an inverse Fourier transform on the synthetic broadband radar high-resolution range image spectrum to obtain a time-domain high-resolution range image of the radar.

[0031] Specifically, in actual implementation, embodiments of this application can first construct a transmission waveform. The signal model is shown below:

[0032] in, This is the baseband waveform of the nth pulse; Let n be the carrier frequency of the nth pulse; For the fast time dimension.

[0033] Secondly, this application uses a single-scattering point echo as an example to introduce and explain the signal preprocessing mechanism. This application's embodiments can construct the received echo after down-conversion demodulation. Signal model:

[0034] in, R is the echo delay of the nth pulse; R is the target distance; V is the target velocity under the "stop-go-stop" model.

[0035] Will receive the echo Transform to frequency domain form using Fourier transform:

[0036] in, The symbol represents the convolution operation; This is the Fourier transform.

[0037] Furthermore, the embodiments of this application can be applied to... and Simplify:

[0038] in, This is the frequency domain form of the baseband waveform.

[0039]

[0040] Based on the above formula, the embodiments of this application can be obtained The simplified expression:

[0041] To simplify the symbols, let ,get The simplified expression:

[0042] Intrapulse Doppler compensation was performed on the above equations respectively. Interpulse Doppler compensation ( After that, the expression for the compensated signal is obtained:

[0043] Since the target velocity is negligible compared to the speed of light, the frequency domain scaling factor will be used in subsequent calculations. Consider it as 1,

[0044] At this point, embodiments of this application can perform frequency domain matched filtering on each pulse:

[0045] in, Conj is the conjugate operator.

[0046] The result after matched filtering The final synthesized broadband HRRP spectrum is obtained by summing the frequencies at the same frequency and concatenating the frequencies at different frequencies. The target's time-domain radar high-resolution range image is .

[0047] Figure 2 This is a schematic diagram of the signal preprocessing process. (Example) Figure 2 As shown, the signal preprocessing process in this embodiment is as follows: 1. Target detection and velocity estimation: Using a pre-defined target detection and velocity estimation method for agile coherent waveforms, the velocity estimate of the target is obtained. This is used for subsequent speed compensation; 2. Imaging algorithm speed compensation: Based on the above transmission waveform The signal model is used to perform intra-pulse Doppler compensation and inter-pulse Doppler compensation on the echo to generate corresponding compensation signals: 1) Intrapulse Doppler compensation: Calculate the Doppler frequency based on the target velocity estimate. ,in, It is the carrier frequency of the nth pulse; Echo spectrum of all N pulses Perform spectrum shifting, i.e. .

[0048] 2) Interpulse Doppler compensation: Based on target speed estimation:

[0049] 3. Matched filtering processing: Each pulse is subjected to frequency domain matched filtering to obtain the corresponding matched filtered signal, i.e.:

[0050] in, Conj is the conjugate operator.

[0051] 4. Coherent accumulation of points with the same frequency:

[0052] 5. Frequency domain stitching of different frequency points to generate the corresponding synthetic broadband radar high-resolution range image spectrum:

[0053] 6. Perform an inverse Fourier transform on the spectrum of the synthetic broadband radar high-resolution range image to obtain the time-domain radar high-resolution range image:

[0054] Therefore, the embodiments of this application perform corresponding signal preprocessing operations to obtain a high-resolution range image of the time-domain radar, thereby providing reliable data for the identification of radar targets under non-uniform interference.

[0055] In step S102, the preset multi-aircraft target echoes are simulated to construct a multi-aircraft interference-free simulation dataset. Based on the preset colored noise interference generation algorithm, corresponding colored noise is generated, and colored noise and time-domain radar high-resolution range image are superimposed to obtain corresponding interference-containing data samples. Based on the interference-containing data samples and the multi-aircraft interference-free simulation dataset, a radar high-resolution range image simulation dataset is constructed.

[0056] Those skilled in the art should understand that, due to the non-uniform distribution of interference energy in radar echoes in actual battlefield environments (such as strong interference in some frequency bands and weak interference in others), datasets under the traditional uniform noise assumption are difficult to meet the needs of complex adversarial scenarios.

[0057] Therefore, the embodiments of this application can generate training and testing data that are more consistent with the actual electromagnetic environment by simulating the echoes of multiple target models and combining them with controllable colored noise interference.

[0058] Optionally, in one embodiment of this application, simulations are performed on the echoes of preset multi-aircraft targets to construct a multi-aircraft interference-free simulation dataset. Based on a preset colored noise interference generation algorithm, corresponding colored noise is generated, and the colored noise and a time-domain radar high-resolution range image are superimposed to obtain corresponding interference-containing data samples. This includes: calculating and simulating different types of aircraft using preset three-dimensional electromagnetic field simulation software to obtain corresponding radar cross-section data; calculating time-domain radar high-resolution range images of each type of aircraft at different observation angles to construct a multi-aircraft interference-free simulation dataset based on the time-domain radar high-resolution range images; determining different interference types and the filter transfer function corresponding to each interference type based on the preset colored noise interference generation algorithm; adjusting the total interference energy at a preset signal-to-interference ratio based on different interference types and the filter transfer function corresponding to each interference type to generate colored noise; superimposing the colored noise and the time-domain radar high-resolution range image to generate corresponding interference-containing signals; and performing a Fourier transform on the interference-containing signals to obtain interference-containing data samples.

[0059] It should be noted that the steps for constructing the radar high-resolution range image simulation dataset in this application embodiment are as follows: Step 1: Generation of multi-model interference-free simulation dataset: Figure 3 This is a schematic diagram illustrating the construction of the simulation dataset model. (Example:) Figure 3 As shown, in order to construct a representative HRRP target library, this application uses six types of aircraft of different sizes and uses as benchmark targets, and obtains their high-fidelity RCS data through CST electromagnetic calculation simulation.

[0060] Based on the above signal preprocessing method, HRRP data for each aircraft model under multi-angle observation are obtained; the final multi-aircraft model interference-free simulation dataset is denoted as... Each target category contains 800 samples, covering typical observation conditions; Step 2, Superposition of non-uniform interference in the frequency domain: To simulate the non-uniform spectral characteristics of actual interference, this application employs a colored noise interference generation algorithm. The core idea is to control the distribution of interference energy in the frequency domain under a fixed signal-to-interference ratio (SJR). The specific process is as follows: 1. Interference type definition: Narrowband strong interference: The interference energy is concentrated in 20%~30% of the frequency band, and the amplitude is significantly higher than other frequency bands; Broadband weak interference: The interference energy is distributed in more than 80% of the frequency band, and the interference intensity is inconsistent between different frequency bands.

[0061] 2. Color noise generation: like Figure 4 As shown, for white noise Frequency domain filtering is performed by designing the filter transfer function. Controlling the shape of the interference spectrum; for example: Narrowband interference: It is a bandpass filter; Broadband interference: It is a multi-peak comb filter; Adjust the total interference energy according to the target SJR (-45dB~-35dB) to ensure that the interference intensity matches the actual scenario.

[0062] 3. Interference superposition: The generated colored noise With the target HRRP time domain signal The signals are superimposed to obtain the interference signal, and then transformed to the frequency domain using FFT, such as... Figure 5 and Figure 6 As shown, where, Figure 5 (a) in the diagram is a schematic of HRRP for a non-interference target; Figure 5 (b) in the diagram is a schematic diagram of HRRP with interfering targets; Figure 6 (a) in the diagram is a schematic diagram of the echo spectrum of an interference-free target; Figure 6(b) in the diagram is a schematic diagram of the echo spectrum of the target with interference. Five interference-containing data samples are generated for each interference-free HRRP sample to form the final training samples. And form sample-label pairs with the sample labels. .

[0063]

[0064] in, The total number of samples in the simulation dataset. Sample size The label corresponding to each sample.

[0065] Therefore, the embodiments of this application generate training and testing data that are more consistent with the actual electromagnetic environment by simulating the echoes of multiple target models and combining them with controllable colored noise interference, thereby effectively ensuring the quality and efficiency of subsequent model construction and training.

[0066] In step S103, a complex frequency attention network is constructed based on the preset frequency domain non-uniform interference characteristics. A radar target recognition model is established based on the complex frequency attention network, the preset radar high-resolution range image classifier, and the hybrid loss function. The radar target recognition model is trained using a radar high-resolution range image simulation dataset so that the trained radar target recognition model can be used to perform radar target recognition operations under non-uniform interference environment during the online recognition stage.

[0067] Subsequently, embodiments of this application can construct a radar target recognition model, which mainly consists of the following core components: (1) Complex frequency attention network: designed based on the characteristics of non-uniform interference in the frequency domain, and dynamically suppresses interference frequency bands through complex frequency domain attention weights; (2) HRRP classifier: It contains multi-layer convolutional neural networks and fully connected layers, which are used to extract deep features and complete target classification; (3) Hybrid loss function: Combining cross-entropy loss and local mean square error loss to achieve joint optimization of classification accuracy and feature preservation ability.

[0068] Furthermore, embodiments of this application can train a radar target recognition model using a radar high-resolution range image simulation dataset, and then use the trained radar target recognition model to perform radar target recognition operations under non-uniform interference environments.

[0069] Therefore, this application embodiment can utilize the characteristic of uneven power distribution of interference in the frequency domain in a compound interference scenario, and use the frequency domain power containing interference echoes as auxiliary information to construct a frequency attention mechanism. By performing attention weighting in the frequency domain, interference frequency bands are suppressed and target priors are enhanced. In addition, this application embodiment enhances the feature discrimination ability and recognition accuracy of the model under non-uniform compound interference conditions by designing a joint loss function combining local MSE loss and cross-entropy loss.

[0070] Optionally, in one embodiment of this application, a complex frequency attention network is constructed based on preset frequency domain non-uniform interference characteristics, and a radar target recognition model is established based on the complex frequency attention network, a preset radar high-resolution range image classifier, and a hybrid loss function. This includes: constructing a shared-weight multilayer perceptron through multiple fully connected layers, and constructing a complex frequency attention network based on a preset dual-channel pooling mechanism, activation function, and shared-weight multilayer perceptron; constructing a radar high-resolution range image classifier based on preset convolutional layers, batch normalization layers, activation function layers, pooling layers, Flatten layers, and fully connected layers; and establishing a radar target recognition model based on the complex frequency attention network, the radar high-resolution range image classifier, and a hybrid loss function composed of a preset local MSE loss function and a cross-entropy loss function.

[0071] Understandably, the attention mechanism of the complex frequency attention network essentially implements a data-driven, differentiable adaptive frequency domain filter. Non-uniform suppression and deceptive interference physically involve injecting high-power signals into specific frequency bands, a characteristic that manifests as discrete or localized high-amplitude spectral blocks in the frequency domain power spectrum. Therefore, the frequency domain power spectrum constitutes a relatively optimal observation domain for the spectral distribution characteristics of the interference energy. The core function of the attention mechanism is dynamic weight allocation. After training, this complex frequency attention network can learn the negative correlation between the amplitude of the frequency domain power spectrum and the reliability of the target information.

[0072] Specifically, for high-power frequencies indicating strong interference on the power spectrum, the complex frequency attention network will assign a low weight close to zero; for frequencies that are not interfered with or have weak interference, it will assign a high weight close to 1.

[0073] Model-based filters are often designed according to a specific optimization objective, such as the Wiener filter, which aims to minimize the mean square error (MSE) between the estimated and true signals, and the whitening filter, which aims to maximize the signal-to-noise ratio (SNR). However, these model-driven methods face two inherent limitations when dealing with complex non-uniform interference: 1. Strong dependence on precise statistical priors: Such methods usually require known precise statistical models of signals and interferences; for example, the optimal design of Wiener filters must use the power spectral density functions of the target and interference as prior knowledge. However, in modern complex and variable electromagnetic environments, such prior information is often dynamic and difficult to obtain precisely.

[0074] 2. Inconsistency between signal-level optimization and task-level optimization: Filters based on fixed mathematical criteria optimize for signal fidelity, but this does not guarantee optimal separability in the classification feature space. In other words, a filter operation that is optimal under MSE or SNR metrics may inadvertently suppress or distort weak but highly discriminative feature components that are crucial for downstream classification tasks. This results in excellent performance at the signal level but suboptimal or even worse performance in the final task performance metrics (such as classification accuracy).

[0075] The frequency attention mechanism in this application fundamentally avoids the aforementioned problems. First, it is a data-driven adaptive method that requires no explicit prior statistical models of signals or interference. Through end-to-end learning from large amounts of data, the complex frequency attention network implicitly captures the energy distribution patterns of interference in the frequency domain from the data itself and generates matching suppression strategies, thus exhibiting strong adaptability and generalization performance against dynamic and unknown interference patterns. Secondly, and more importantly, the filtering behavior of this attention mechanism is not isolated and optimized solely for signal fidelity. Instead, it is directly supervised and driven by the final task performance metric (i.e., the hybrid loss function) within the end-to-end training framework of the entire network. That is, the ultimate goal of learning and optimizing attention weights is to maximize classification accuracy. Therefore, the network autonomously learns a selective strategy: it not only suppresses high-power interference bands but also retains or even enhances the frequency components that contribute most to improving the classification confidence of a specific target category, regardless of their power. This data-driven frequency domain attention mechanism ensures a high degree of consistency between the signal processing process and the final classification objective. It enables the network to learn a feature representation that is optimal at the semantic level, rather than just at the signal level, thereby fundamentally improving the robustness and accuracy of target recognition in complex interference environments.

[0076] In the embodiments of this application, a complex frequency attention network can be constructed by using a shared weight multilayer perceptron composed of two fully connected layers, a preset dual-path pooling mechanism, and activation functions. like Figure 7As shown, the HRRP classifier in this embodiment mainly consists of a convolutional network, a Flatten network, a fully connected network, and an activation function. The HRRP classifier extracts features from a one-dimensional HRRP sequence at multiple scales in the time domain through convolutional networks of different sizes, and then completes the classification task through the Flatten network, the fully connected network, and the Softmax function in sequence.

[0077] Furthermore, the hybrid loss function in this application embodiment consists of cross-entropy loss ( ) and local mean square error loss ( It consists of two parts, using adaptive weighting coefficients. A balance is struck to simultaneously optimize the model's global classification ability and local feature matching accuracy, thereby improving the robustness and generalization performance of HRRP target recognition, as detailed below: 1. Cross-entropy loss ( ): Cross-entropy loss is a standard loss function in classification tasks, used to measure the difference between the probability distribution predicted by the model and the true label. It represents the predicted class probability given an input sample. (in k (number of categories), and their actual labels. (In one-hot encoding form), the expression for calculating cross-entropy loss is:

[0078] 2. Local mean square error loss ( ): Because HRRP data has local sensitivity characteristics (such as key regions rich in structural information, such as peaks and energy concentration areas), relying solely on cross-entropy loss may cause the model to ignore local details and affect matching accuracy. Therefore, this application introduces local mean squared error loss to calculate the error between the predicted waveform and the true waveform in specific key regions, thereby enhancing the model's ability to fit local features.

[0079] Predicted waveforms given HRRP samples and the actual waveform First, select a central key region (such as an energy concentration area) on the waveform, and then calculate the mean square error of these regions:

[0080] Among them, subscript It is the time-domain point index of HRRP. The number of sampling points in the key region is used as a model hyperparameter, and is generally taken as... This loss function can constrain the model to maintain high-precision matching in local structures, effectively prevent the model from overfitting to HRRP edge noise points, and avoid losing subtle but important target features due to global optimization.

[0081] 3. Adaptive weighted total loss ( ): To balance the contributions of the two types of losses, this embodiment of the application calculates the total loss using a weighted summation method:

[0082] in, is an adjustable weighting coefficient used to control the degree of influence of local matching loss. Experiments show that when At the same time, the model can effectively improve its sensitivity to local features while maintaining high classification accuracy.

[0083] Optionally, in one embodiment of this application, training a radar target recognition model using a radar high-resolution range image simulation dataset includes: dividing the radar high-resolution range image simulation dataset into a training dataset, a validation dataset, and a test dataset based on a preset partitioning ratio; performing Z-score normalization on the complex components of each data point in the training dataset, validation dataset, and test dataset to obtain corresponding normalized data; inputting the normalized data corresponding to the training dataset into the complex frequency attention network of the radar target recognition model to extract the time-domain radar high-resolution range image features corresponding to the normalized data through the complex frequency attention network, and performing Fourier transform and feature dimension transformation on the time-domain radar high-resolution range image features to generate corresponding target frequency domain features; Using a dual-pooling mechanism, target frequency domain features are compressed and extracted to generate corresponding feature vectors. These feature vectors are then concatenated, and a shared-weight multilayer perceptron is used to perform nonlinear transformation and linear projection on the concatenated feature vectors to obtain the corresponding fused features. Based on the activation function, attention weights corresponding to the fused features are generated, and attention weighting is performed according to these attention weights to generate corresponding enhanced features. These enhanced features are then input into a radar high-resolution range image classifier to generate the corresponding classification results. The radar target recognition model is trained based on the classification results and a hybrid loss function. The trained radar target recognition model is then fine-tuned using standardized data corresponding to the validation and test datasets to ensure that the radar target recognition model meets the preset performance requirements.

[0084] In actual implementation, the process of training the radar target recognition model using a radar high-resolution range image simulation dataset in this embodiment is as follows: 1. Dataset construction and training / test set partitioning: Based on the simulation dataset containing six types of military aircraft targets constructed above, a total of 4,800 samples were randomly divided into training and test sets in an 8:2 ratio. To ensure the generalization ability of the model, an additional 10% was allocated from the training set as a validation set to monitor the training process. 2. Data preprocessing: In the data preprocessing stage, the complex components of the frequency domain data are Z-score standardized according to the modulus to obtain the corresponding standardized data and eliminate the impact of dimensional differences on model training. 3. Hyperparameter and optimizer settings: The training hyperparameters were optimized through multiple experiments. The AdamW optimizer was used with an initial learning rate of 5e-4 and a batch size of 64 to ensure both training efficiency and gradient estimation accuracy. The entire training process lasted 70 epochs, with an early stopping tolerance of 10 epochs to prevent overfitting.

[0085] 4. Reshape operation: like Figure 8 As shown, in this embodiment of the application, a complex frequency attention network can be used to perform frequency domain transformation and dimension reconstruction operations on the input features (i.e., high-resolution range image features of the radar) to generate corresponding target frequency domain features; let the input HRRP features be... Where H represents the spectral length, and 2 represents the real and imaginary parts of the complex spectrum. This complex frequency attention network first converts the time-domain features into a frequency-domain representation using a Fast Fourier Transform, and then performs a Reshape operation to transform the feature dimensions into the form H / C×2×C, where C is a preset hyperparameter for the number of channels, which is taken here as... C =2 *Freq_ num This operation enables the reorganization of frequency domain features along the channel dimension, laying the foundation for subsequent multi-scale feature extraction.

[0086] 5. Pooling operation: A dual-pooling mechanism in a complex frequency attention network is employed to compress and extract target frequency domain features. Specifically, embodiments of this application can perform max pooling and average pooling operations in parallel, generating feature vectors of dimension C respectively. The max pooling branch highlights significant frequency domain features by extracting local spectral maximum values, while the average pooling branch obtains the overall interference power frequency domain distribution characteristics by calculating the global average value of the spectrum. These two complementary feature representation methods together constitute a multi-angle description of the frequency domain features.

[0087] 6. SharedMLP (Shared Weighted Multilayer Perceptron): Feature fusion and nonlinear transformation are achieved through a shared-weight multilayer perceptron to obtain the corresponding fused features. This complex-frequency attention network concatenates the feature vectors output by dual-pooling and feeds them into a shared MLP network consisting of two fully connected layers. The first layer uses the ReLU activation function for nonlinear transformation, and the second layer performs linear projection. This design achieves deep fusion of features from different pooling methods, while effectively controlling the model complexity through a parameter-sharing mechanism.

[0088] 7. Sigmoid activation: The features processed by SharedMLP are used to generate attention weights in the range [0,1] through the Sigmoid activation function. This weight reflects the importance of each frequency domain component; finally, attention weighting is achieved through Hadamard product to generate the corresponding enhanced features: , where ⊙ denotes element-wise multiplication. This operation achieves adaptive enhancement of the target feature frequency band and suppression of interference frequency bands, significantly improving the distinguishability of the features.

[0089] 8. The HRRP classifier extracts features from one-dimensional HRRP sequences at multiple scales in the temporal domain using convolutional networks of different sizes: 1) Time-frequency domain conversion The frequency domain feature data, i.e. the enhanced features, processed by the complex frequency attention network is subjected to an inverse fast Fourier transform operation to convert the data from the frequency domain back to the time domain. The output is also divided into real and imaginary parts, and the dimension is changed back to H so that subsequent classification operations can be performed in the time domain. 2) Convolutional layers (Conv Block 1, Conv Block 2, Conv Block 3) Conv (convolutional layer): Convolutional kernels are used to perform convolution operations on one-dimensional HRRP sequences to extract local patterns and features in the sequence; convolutional kernels of different sizes can capture local features of target scattering point distribution at different scales.

[0090] Batch Normalization: Normalizes the one-dimensional data after convolution, standardizing the data distribution, which helps to speed up model training convergence, avoid gradient problems, and improve the model's generalization ability.

[0091] ReLU (Activation Function Layer): The ReLU activation function is used to introduce non-linearity into the model; by setting values ​​less than 0 to 0 and keeping values ​​greater than 0 unchanged, the model can learn the complex non-linear relationships in the HRRP sequence.

[0092] Pooling (pooling layer): The average pooling method is used to downsample one-dimensional data, reducing the amount of data and computation, while retaining key features, preventing model overfitting, and enhancing the stability of HRRP sequence feature extraction. The three convolutional modules gradually filter and refine the features of the HRRP sequence.

[0093] 3) Flatten layer Flattening the multidimensional feature data processed by the convolutional network into a one-dimensional vector, and converting the spatial features extracted from the HRRP sequence into a form suitable for processing by the fully connected layer, is a transitional link connecting the convolutional layer and the fully connected layer. 4) Fully connected layers (FC Block 1, FC Block 2) Linear layer: As a fully connected layer, its neurons are connected to all neurons in the previous layer. It performs a linear transformation on the flattened one-dimensional feature vector, mapping the HRRP sequence features to a new feature space.

[0094] Batch Normalization Layer: Normalizes the feature vectors output by the fully connected layer, stabilizing the training process and helping the model to better learn HRRP sequence-related features.

[0095] ReLU (Activation Function Layer): Introduces non-linearity again to further explore the complex relationships between HRRP sequence features and improve the model's ability to express HRRP data.

[0096] Dropout: During training, some neurons are randomly set to output 0, preventing the model from overfitting to the HRRP training data and improving the model's generalization performance to different HRRP sequences in real-world applications. Two fully connected modules further integrate and transform HRRP sequence features.

[0097] Softmax: Finally, the values ​​output by the fully connected module are converted into a probability distribution through the Softmax function to obtain the probability that the input HRRP sequence belongs to each category. This probability is used to calculate the cross-entropy loss in the subsequent mixture loss function to generate the corresponding classification results.

[0098] Finally, embodiments of this application use classification results and a hybrid loss function to train a radar target recognition model, and fine-tune and optimize the trained radar target recognition model using standardized data corresponding to the validation dataset and the test dataset, so that the radar target recognition model meets the corresponding performance requirements.

[0099] Therefore, the embodiments of this application, through in-depth research on the distortion mechanism of HRRP under interference environment, can provide a reliable and robust target recognition scheme for agile phase coherent radar under complex electromagnetic interference environment, and solve the problems of poor robustness and decreased recognition accuracy of current target recognition under complex interference conditions.

[0100] According to the radar target recognition method based on attention network under non-uniform interference proposed in the offline training stage of this application, the method constructs a model of the transmitted waveform signal of the target radar and performs signal preprocessing operations on the transmitted waveform signal model to construct a corresponding time-domain radar high-resolution range image. It simulates the echoes of multiple target models to construct a multi-model interference-free simulation dataset, and generates corresponding colored noise based on a preset colored noise interference generation algorithm. The colored noise and the time-domain radar high-resolution range image are then superimposed to obtain corresponding interference-containing data samples. Based on the interference-containing data samples and the multi-model interference-free simulation dataset, a radar high-resolution range image simulation dataset is constructed. Based on the preset frequency-domain non-uniform interference characteristics, a complex frequency attention network is constructed. A radar target recognition model is established based on the complex frequency attention network, a preset radar high-resolution range image classifier, and a hybrid loss function. The radar target recognition model is trained using the radar high-resolution range image simulation dataset, so that the trained radar target recognition model can be used for radar target recognition operations under non-uniform interference environments during the online recognition stage. This application conducts an in-depth study of the distortion mechanism of HRRP under interference environment in order to obtain a robust deep learning target recognition strategy with strong anti-interference ability.

[0101] Figure 9 This is a flowchart illustrating a radar target recognition method based on attention networks under non-uniform interference, provided as an embodiment of this application, applied in the online recognition stage.

[0102] like Figure 9 As shown, the radar target recognition method based on attention networks under non-uniform interference includes the following steps: In step S901, the high-resolution range image of the frequency domain radar corresponding to each target under test, which includes interference, is acquired by the target radar.

[0103] In step S902, the high-resolution range image of the frequency domain radar containing interference is input into the trained radar target recognition model so that the complex attention weights corresponding to the high-resolution range image of the frequency domain radar containing interference are generated through the complex frequency attention network of the radar target recognition model. The high-resolution range image of the frequency domain radar containing interference and the complex attention weights are multiplied point by point to obtain the corresponding filtered spectrum features.

[0104] In step S903, the filtered spectrum features are input into the radar high-resolution range image classifier to perform multi-layer nonlinear transformation on the filtered spectrum features and output the probability distribution corresponding to each target to be tested, so as to obtain the recognition result corresponding to each target to be tested based on the probability distribution.

[0105] In the online identification stage, the forward propagation process of this application embodiment can first input the HRRP frequency domain data containing interference (i.e., the high-resolution range image of the frequency domain radar containing interference) into a complex frequency attention network. The complex frequency attention network automatically learns the attention weight of each frequency point through a deep neural network, and multiplies the original spectrum with the attention weight in complex form point by point to obtain the corresponding filtered spectrum features, thereby realizing the dynamic suppression of the interference frequency band.

[0106] Subsequently, the embodiments of this application feed the filtered spectral features (i.e., filtered spectral features) into the HRRP classifier, and finally output the probability distribution of each type of target through multi-layer nonlinear transformation, so as to obtain the recognition result corresponding to each target to be tested based on the probability distribution.

[0107] It is understood that in the embodiments of this application, the complex frequency attention mechanism is essentially a weighted concatenation of the target echo frequency domain. Therefore, in order to specifically compare the recognition performance of the recognition method of this application embodiment with that of the prior art in the face of non-uniform frequency domain interference, this application embodiment uses the Wiener filtering method, which is the optimal weighting method with minimizing MSE as the optimization objective, as the baseline model under the same dataset. Under the condition that the classifier performance is the same, the recognition effect is compared, and the recognition accuracy and confusion matrix of the proposed method and the Wiener filtering method are output, such as... Figure 10 and Figure 11 As shown, where, Figure 10 This is a schematic diagram of the recognition rate and confusion matrix under the simulation dataset; Figure 11 This is a schematic diagram illustrating the recognition rate and confusion matrix under the actual test dataset. Figure 10 and Figure 11 As shown, compared with the Wiener filtering method, the method proposed in this application embodiment has a better recognition accuracy.

[0108] According to the embodiments of this application, a radar target recognition method based on attention networks under non-uniform interference, applied to the online recognition stage, is proposed. This method acquires a high-resolution range image of the radar in the frequency domain with interference for each target. The high-resolution range image is then input into a trained radar target recognition model. A complex attention network of the radar target recognition model generates complex attention weights corresponding to the high-resolution range image with interference. The high-resolution range image with interference and the complex attention weights are multiplied point-by-point to obtain the corresponding filtered spectral features. These filtered spectral features are then input into a radar high-resolution range image classifier to perform multi-layer nonlinear transformations, outputting a probability distribution for each target. The recognition result for each target is then obtained based on the probability distribution. This application, through in-depth research into the distortion mechanism of HRRP under interference environments, obtains a robust deep learning target recognition strategy with strong anti-interference capabilities.

[0109] Secondly, with reference to the accompanying drawings, a radar target recognition device based on attention networks under non-uniform interference, according to an embodiment of this application, is described.

[0110] Figure 12 This is a block diagram of a radar target recognition device based on an attention network under non-uniform interference, applied in the offline training phase according to an embodiment of this application.

[0111] like Figure 12 As shown, the radar target recognition device 10 based on attention network under non-uniform interference includes: a preprocessing module 101, a dataset construction module 102, and a training module 103.

[0112] The preprocessing module 101 is used to construct the transmitted waveform signal model of the target radar and perform signal preprocessing operations on the transmitted waveform signal model to construct the corresponding time-domain radar high-resolution range image.

[0113] The dataset construction module 102 is used to simulate the echoes of multiple target models to construct a multi-model interference-free simulation dataset. Based on a preset colored noise interference generation algorithm, it generates corresponding colored noise and superimposes the colored noise and the time-domain radar high-resolution range image to obtain the corresponding interference-containing data sample. Based on the interference-containing data sample and the multi-model interference-free simulation dataset, a radar high-resolution range image simulation dataset is constructed.

[0114] Training module 103 is used to construct a complex frequency attention network based on the preset frequency domain non-uniform interference characteristics, and to establish a radar target recognition model based on the complex frequency attention network, the preset radar high-resolution range image classifier and the hybrid loss function. The radar target recognition model is trained through a radar high-resolution range image simulation dataset so that the trained radar target recognition model can be used to perform radar target recognition operations under non-uniform interference environment in the online recognition stage.

[0115] Optionally, in one embodiment of this application, the preprocessing module 101 includes: a generation unit, a compensation unit, a filtering unit, and a transformation unit.

[0116] The generation unit is used to construct a corresponding received echo signal model based on the transmitted waveform signal model, and to perform a Fourier transform on the received echo of the target radar according to the received echo signal model to generate the corresponding frequency domain received echo signal.

[0117] The compensation unit is used to estimate the target velocity of the target under test through the target radar, and calculate the corresponding Doppler frequency based on the transmitted waveform signal model and the target velocity. It also performs intra-pulse Doppler compensation and inter-pulse Doppler compensation on the frequency domain received echo signal through the Doppler frequency to generate the corresponding compensation signal.

[0118] The filtering unit is used to perform frequency domain matched filtering on each pulse in the compensation signal to obtain the corresponding matched filtered signal, and to perform same-frequency point summation and frequency domain splicing operations on the matched filtered signal to generate the corresponding synthetic broadband radar high-resolution range image spectrum.

[0119] The transformation unit is used to perform inverse Fourier transform on the spectrum of the synthetic broadband radar high-resolution range image to obtain the time-domain radar high-resolution range image.

[0120] Optionally, in one embodiment of this application, the dataset construction module 102 includes: a simulation unit, a first construction unit, an adjustment unit, and an overlay unit.

[0121] The simulation unit is used to perform simulation calculations on different types of aircraft based on preset three-dimensional electromagnetic field simulation software to obtain corresponding radar cross section data.

[0122] The first building unit is used to calculate the time-domain radar high-resolution range image of each type of aircraft at different observation angles, so as to build a multi-aircraft interference-free simulation dataset based on the time-domain radar high-resolution range image.

[0123] The adjustment unit is used to determine different interference types and the filter transfer function corresponding to each interference type based on a preset colored noise interference generation algorithm, and to adjust the total interference energy at a preset signal-to-interference ratio based on the different interference types and the filter transfer function corresponding to each interference type to generate colored noise.

[0124] The superposition unit is used to superimpose colored noise and time-domain radar high-resolution range image to generate corresponding interference-containing signals, and to perform Fourier transform on the interference-containing signals to obtain interference-containing data samples.

[0125] Optionally, in one embodiment of this application, the training module 103 includes: a second construction unit, a third construction unit, and an establishment unit.

[0126] The second building unit is used to construct a shared-weight multilayer perceptron through multiple fully connected layers, and to construct a complex-frequency attention network based on a preset dual-pooling mechanism, activation function, and shared-weight multilayer perceptron.

[0127] The third building unit is used to construct a radar high-resolution range image classifier based on a preset convolutional layer, batch normalization layer, activation function layer, pooling layer, Flatten layer and fully connected layer.

[0128] The establishment unit is used to build a radar target recognition model based on a complex frequency attention network, a radar high-resolution range image classifier, and a hybrid loss function composed of a preset local MSE loss function and a cross-entropy loss function.

[0129] Optionally, in one embodiment of this application, the training module 103 further includes: a partitioning unit, a processing unit, an extraction unit, a splicing unit, a weighting unit, and a fine-tuning unit.

[0130] The partitioning unit is used to divide the radar high-resolution range image simulation dataset into training dataset, validation dataset, and test dataset based on a preset partitioning ratio.

[0131] The processing unit is used to perform Z-score normalization on the complex components of each data point in the training dataset, validation dataset, and test dataset to obtain the corresponding normalized data.

[0132] The extraction unit is used to input the standardized data corresponding to the training dataset into the complex frequency attention network of the radar target recognition model, so as to extract the time-domain radar high-resolution range image features corresponding to the standardized data through the complex frequency attention network, and perform Fourier transform and feature dimension transformation on the time-domain radar high-resolution range image features to generate the corresponding target frequency domain features.

[0133] The concatenation unit is used to compress and extract the target frequency domain features based on the dual-channel pooling mechanism to generate the corresponding feature vectors, and then concatenate the feature vectors. The concatenated feature vectors are then subjected to nonlinear transformation and linear projection through a shared-weight multilayer perceptron to obtain the corresponding fused features.

[0134] The weighting unit is used to generate attention weights corresponding to the fused features based on the activation function, and to perform attention weighting operation based on the attention weights to generate the corresponding enhanced features. The enhanced features are then input into the radar high-resolution range image classifier to generate the corresponding classification results.

[0135] The fine-tuning unit is used to train the radar target recognition model based on the classification results and the hybrid loss function, and to fine-tune the trained radar target recognition model with standardized data corresponding to the validation dataset and the test dataset so that the radar target recognition model meets the preset performance requirements.

[0136] It should be noted that the foregoing explanation of the radar target recognition method based on attention network under non-uniform interference applied to the offline training stage also applies to the radar target recognition device based on attention network under non-uniform interference applied to the offline training stage, and will not be repeated here.

[0137] According to the embodiments of this application, a radar target recognition device based on an attention network under non-uniform interference, applied in the offline training stage, includes a preprocessing module 101 for constructing a transmitted waveform signal model of the target radar and performing signal preprocessing operations on the transmitted waveform signal model to construct a corresponding time-domain high-resolution radar range image; a dataset construction module 102 for simulating preset multi-type target echoes to construct a multi-type interference-free simulation dataset, and generating corresponding colored noise based on a preset colored noise interference generation algorithm, and superimposing the colored noise and the time-domain high-resolution radar range image to obtain corresponding interference-containing data samples, so as to construct a radar high-resolution range image simulation dataset based on the interference-containing data samples and the multi-type interference-free simulation dataset; and a training module 103 for constructing a complex frequency attention network based on preset frequency domain non-uniform interference characteristics, and establishing a radar target recognition model based on the complex frequency attention network, a preset radar high-resolution range image classifier, and a hybrid loss function, and training the radar target recognition model through the radar high-resolution range image simulation dataset, so as to use the trained radar target recognition model to perform radar target recognition operations under non-uniform interference environment in the online recognition stage. This application conducts an in-depth study of the distortion mechanism of HRRP under interference environment in order to obtain a robust deep learning target recognition strategy with strong anti-interference ability.

[0138] Figure 13This is a block diagram of a radar target recognition device based on an attention network under non-uniform interference, applied in the online recognition stage according to an embodiment of this application.

[0139] like Figure 13 As shown, the radar target identification device 20 based on attention network under non-uniform interference applied in the online identification stage includes: an acquisition module 201, an inference module 202, and an identification module 203.

[0140] The acquisition module 201 is used to acquire the high-resolution range image of the frequency domain radar corresponding to each target under test, which contains interference, through the target radar. The inference module 202 is used to input the interference-containing frequency domain radar high-resolution range image into the trained radar target recognition model, so as to generate the complex attention weights corresponding to the interference-containing frequency domain radar high-resolution range image through the complex frequency attention network of the radar target recognition model, and to perform a point-by-point multiplication operation on the interference-containing frequency domain radar high-resolution range image and the complex attention weights to obtain the corresponding filtered spectrum features. The recognition module 203 is used to input the filtered spectrum features into the radar high-resolution range image classifier to perform multi-layer nonlinear transformation on the filtered spectrum features and output the probability distribution corresponding to each target to be tested, so as to obtain the recognition result corresponding to each target to be tested based on the probability distribution.

[0141] It should be noted that the foregoing explanation of the radar target recognition method embodiment based on attention network under non-uniform interference in the online recognition stage also applies to the radar target recognition device based on attention network under non-uniform interference in the online recognition stage embodiment, and will not be repeated here.

[0142] According to the embodiments of this application, a radar target recognition device based on an attention network under non-uniform interference, applied in the online recognition stage, includes an acquisition module 201 for acquiring the interference-laden frequency domain high-resolution range image of each target to be tested via the target radar; an inference module 202 for inputting the interference-laden frequency domain high-resolution range image into a trained radar target recognition model, generating complex attention weights corresponding to the interference-laden frequency domain high-resolution range image through the complex frequency attention network of the radar target recognition model, and performing point-by-point multiplication of the interference-laden frequency domain high-resolution range image and the complex attention weights to obtain the corresponding filtered spectral features; and a recognition module 203 for inputting the filtered spectral features into a radar high-resolution range image classifier, performing multi-layer nonlinear transformations on the filtered spectral features, outputting the probability distribution corresponding to each target to be tested, and obtaining the recognition result corresponding to each target to be tested based on the probability distribution. This application, through in-depth research on the distortion mechanism of HRRP under interference environment, obtains a robust deep learning target recognition strategy with strong anti-interference capability.

[0143] Figure 14 A schematic diagram of the structure of an electronic device provided in an embodiment of this application. The electronic device may include: The memory 1401, the processor 1402, and the computer program stored on the memory 1401 and executable on the processor 1402.

[0144] When the processor 1402 executes the program, it implements the radar target recognition method based on attention network under non-uniform interference provided in the above embodiments.

[0145] Furthermore, electronic devices also include: Communication interface 1403 is used for communication between memory 1401 and processor 1402.

[0146] The memory 1401 is used to store computer programs that can run on the processor 1402.

[0147] The memory 1401 may include high-speed RAM memory, and may also include non-volatile memory, such as at least one disk storage.

[0148] If the memory 1401, processor 1402, and communication interface 1403 are implemented independently, then the communication interface 1403, memory 1401, and processor 1402 can be interconnected via a bus to complete communication between them. The bus can be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus, or an Extended Industry Standard Architecture (EISA) bus, etc. Buses can be categorized as address buses, data buses, control buses, etc. For ease of representation, Figure 14 The bus is represented by a single thick line, but this does not mean that there is only one bus or one type of bus.

[0149] Optionally, in a specific implementation, if the memory 1401, processor 1402, and communication interface 1403 are integrated on a single chip, then the memory 1401, processor 1402, and communication interface 1403 can communicate with each other through an internal interface.

[0150] The processor 1402 may be a central processing unit (CPU), an application specific integrated circuit (ASIC), or one or more integrated circuits configured to implement the embodiments of this application.

[0151] This application also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the above-described radar target recognition method under non-uniform interference based on attention networks.

[0152] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0153] 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 at least one of that feature. In the description of this application, "N" means at least two, such as two, three, etc., unless otherwise explicitly specified.

[0154] Any process or method described in the flowchart or otherwise herein can be understood as representing a module, segment, or portion of code comprising one or N executable instructions for implementing custom logic functions or processes, and the scope of the preferred embodiments of this application includes additional implementations in which functions may be performed not in the order shown or discussed, including substantially simultaneously or in reverse order depending on the functions involved, as should be understood by those skilled in the art to which embodiments of this application pertain.

[0155] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-included system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device. More specific examples (a non-exhaustive list) of computer-readable media include: an electrical connection having one or more wires (electronic device), a portable computer disk drive (magnetic device), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Alternatively, the computer-readable medium may be paper or other suitable media on which the program can be printed, since the program can be obtained electronically by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in a computer memory.

[0156] It should be understood that the various parts of this application can be implemented using hardware, software, firmware, or a combination thereof. In the above embodiments, the N steps or methods can be implemented using software or firmware stored in memory and executed by a suitable instruction execution system. If implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0157] Those skilled in the art will understand that all or part of the steps of the methods in the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, the program includes one or a combination of the steps of the method embodiments.

[0158] Furthermore, the functional units in the various embodiments of this application can be integrated into a processing module, or each unit can exist physically separately, or two or more units can be integrated into a module. The integrated module can be implemented in hardware or as a software functional module. If the integrated module is implemented as a software functional module and sold or used as an independent product, it can also be stored in a computer-readable storage medium.

[0159] The storage medium mentioned above can be a read-only memory, a disk, or an optical disk, etc. Although embodiments of this application have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting this application. Those skilled in the art can make changes, modifications, substitutions, and variations to the above embodiments within the scope of this application.

Claims

1. A radar target recognition method based on attention networks under non-uniform interference, applied in the offline training stage, characterized in that, Includes the following steps: A model of the transmitted waveform signal of the target radar is constructed, and signal preprocessing is performed on the transmitted waveform signal model to construct the corresponding time-domain radar high-resolution range image; Simulations are performed on the echoes of multiple target models to construct a multi-model interference-free simulation dataset. Based on a preset colored noise interference generation algorithm, corresponding colored noise is generated and superimposed with the colored noise and the time-domain radar high-resolution range image to obtain corresponding interference-containing data samples. Based on the interference-containing data samples and the multi-model interference-free simulation dataset, a radar high-resolution range image simulation dataset is constructed. Based on the preset frequency domain non-uniform interference characteristics, a complex frequency attention network is constructed, and a radar target recognition model is established according to the complex frequency attention network, the preset radar high-resolution range image classifier and the hybrid loss function. The radar target recognition model is trained through the radar high-resolution range image simulation dataset so that the trained radar target recognition model can be used to perform radar target recognition operations under non-uniform interference environment in the online recognition stage. The process includes simulating the echoes of multiple target aircraft models to construct a multi-aircraft model interference-free simulation dataset, generating corresponding colored noise based on a preset colored noise interference generation algorithm, and superimposing the colored noise and the time-domain radar high-resolution range image to obtain corresponding interference-containing data samples, including: Based on the preset three-dimensional electromagnetic field simulation software, simulation calculations are performed on different types of aircraft to obtain the corresponding radar cross section data. Calculate the time-domain radar high-resolution range image of each type of aircraft at different observation angles, and construct the multi-aircraft model interference-free simulation dataset based on the time-domain radar high-resolution range image; Based on a preset colored noise interference generation algorithm, different interference types and the filter transfer function corresponding to each interference type are determined. Based on the different interference types and the filter transfer function corresponding to each interference type, the total interference energy is adjusted at a preset signal-to-interference ratio to generate the colored noise. The colored noise and the time-domain radar high-resolution range image are superimposed to generate a corresponding interference-containing signal, and the interference-containing signal is subjected to Fourier transform to obtain the interference-containing data sample; The method involves constructing a complex frequency attention network based on preset frequency domain non-uniform interference characteristics, and establishing a radar target recognition model based on the complex frequency attention network, a preset radar high-resolution range image classifier, and a hybrid loss function, including: A shared-weight multilayer perceptron is constructed by using multiple fully connected layers, and the complex frequency attention network is built based on a preset dual-path pooling mechanism, activation function, and the shared-weight multilayer perceptron. The radar high-resolution range image classifier is constructed based on a pre-defined convolutional layer, batch normalization layer, activation function layer, pooling layer, Flatten layer, and fully connected layer. The radar target recognition model is established based on the complex frequency attention network, the radar high-resolution range image classifier, and a hybrid loss function composed of a preset local MSE loss function and a cross-entropy loss function.

2. The method according to claim 1, characterized in that, The process of constructing a model of the target radar's transmitted waveform signal and performing signal preprocessing on the transmitted waveform signal model to construct a corresponding time-domain high-resolution range image of the radar includes: Based on the transmitted waveform signal model, a corresponding received echo signal model is constructed, and the received echo of the target radar is subjected to Fourier transform according to the received echo signal model to generate the corresponding frequency domain received echo signal. The target velocity of the target to be measured is estimated by the target radar, and the corresponding Doppler frequency is calculated based on the transmitted waveform signal model and the target velocity. Intra-pulse Doppler compensation and inter-pulse Doppler compensation are performed on the frequency domain received echo signal using the Doppler frequency to generate the corresponding compensation signal. Each pulse in the compensation signal is subjected to frequency domain matched filtering to obtain the corresponding matched filter signal. The matched filter signal is then summed at the same frequency points and spliced ​​in the frequency domain to generate the corresponding synthetic broadband radar high-resolution range image spectrum. The spectrum of the synthetic broadband radar high-resolution range image is subjected to inverse Fourier transform to obtain the time-domain radar high-resolution range image.

3. The method according to claim 1, characterized in that, The step of training the radar target recognition model using the radar high-resolution range image simulation dataset includes: Based on a preset division ratio, the radar high-resolution range image simulation dataset is divided into a training dataset, a validation dataset, and a test dataset. The complex components of each data point in the training dataset, the validation dataset, and the test dataset are Z-score standardized to obtain the corresponding standardized data. The standardized data corresponding to the training dataset is input into the complex frequency attention network of the radar target recognition model, so as to extract the time-domain radar high-resolution range image features corresponding to the standardized data through the complex frequency attention network, and perform Fourier transform and feature dimension transformation on the time-domain radar high-resolution range image features to generate the corresponding target frequency domain features. Based on the dual-pooling mechanism, the target frequency domain features are compressed and extracted to generate corresponding feature vectors. The feature vectors are then concatenated, and the concatenated feature vectors are subjected to nonlinear transformation and linear projection through the shared weight multilayer perceptron to obtain the corresponding fused features. Based on the activation function, attention weights corresponding to the fused features are generated, and attention weighting operations are performed according to the attention weights to generate corresponding enhanced features. The enhanced features are then input into the radar high-resolution range image classifier to generate corresponding classification results. The radar target recognition model is trained based on the classification results and the hybrid loss function, and the trained radar target recognition model is fine-tuned using standardized data corresponding to the validation dataset and the test dataset to ensure that the radar target recognition model meets the preset performance requirements.

4. A radar target recognition method based on attention networks under non-uniform interference, applied in the online recognition stage, characterized in that, The radar target recognition method based on attention networks under non-uniform interference, as described in any one of claims 1-3, is applied to the offline training phase, wherein the method includes the following steps: The target radar acquires the high-resolution range image of each target under test in the frequency domain, which includes interference. The interference-containing high-resolution range image of the frequency domain radar is input into the trained radar target recognition model, so that the complex attention weights corresponding to the interference-containing high-resolution range image of the frequency domain radar are generated through the complex frequency attention network of the radar target recognition model. The interference-containing high-resolution range image of the frequency domain radar and the complex attention weights are multiplied point by point to obtain the corresponding filtered spectrum features. The filtered spectral features are input into the radar high-resolution range image classifier to perform multi-layer nonlinear transformation on the filtered spectral features, and output the probability distribution corresponding to each target to be tested, so as to obtain the recognition result corresponding to each target to be tested based on the probability distribution.

5. A radar target recognition device under non-uniform interference based on an attention network, applied in the offline training phase, characterized in that, include: The preprocessing module is used to construct a model of the transmitted waveform signal of the target radar and perform signal preprocessing operations on the transmitted waveform signal model to construct the corresponding time-domain radar high-resolution range image. The dataset construction module is used to simulate the echoes of multiple target models to construct a multi-model interference-free simulation dataset. Based on a preset colored noise interference generation algorithm, it generates corresponding colored noise and superimposes the colored noise and the time-domain radar high-resolution range image to obtain corresponding interference-containing data samples. Based on the interference-containing data samples and the multi-model interference-free simulation dataset, a radar high-resolution range image simulation dataset is constructed. The training module is used to construct a complex frequency attention network based on the preset frequency domain non-uniform interference characteristics, and to establish a radar target recognition model based on the complex frequency attention network, the preset radar high-resolution range image classifier and the hybrid loss function. The radar target recognition model is trained through the radar high-resolution range image simulation dataset so that the trained radar target recognition model can be used to perform radar target recognition operations under non-uniform interference environment in the online recognition stage. The dataset construction module includes: The simulation unit is used to perform simulation calculations on different types of aircraft based on preset three-dimensional electromagnetic field simulation software in order to obtain the corresponding radar cross section data. The first construction unit is used to calculate the time-domain radar high-resolution range image of each type of aircraft at different observation angles, so as to construct the multi-aircraft interference-free simulation dataset based on the time-domain radar high-resolution range image. The adjustment unit is used to determine different interference types and the filter transfer function corresponding to each interference type based on a preset colored noise interference generation algorithm, and to adjust the total interference energy based on the different interference types and the filter transfer function corresponding to each interference type at a preset signal-to-interference ratio to generate the colored noise. The superposition unit is used to superimpose the colored noise and the time-domain radar high-resolution range image to generate a corresponding interference-containing signal, and to perform a Fourier transform on the interference-containing signal to obtain the interference-containing data sample. The training module includes: The second construction unit is used to build a shared-weight multilayer perceptron through multiple fully connected layers, and to construct the complex frequency attention network based on a preset dual-path pooling mechanism, activation function and the shared-weight multilayer perceptron. The third construction unit is used to construct the radar high-resolution range image classifier based on a preset convolutional layer, batch normalization layer, activation function layer, pooling layer, Flatten layer and fully connected layer; The establishment unit is used to establish the radar target recognition model based on the complex frequency attention network, the radar high-resolution range image classifier, and a hybrid loss function composed of a preset local MSE loss function and a cross-entropy loss function.

6. A radar target identification device under non-uniform interference based on an attention network, applied in the online identification stage, characterized in that, The radar target recognition device based on an attention network under non-uniform interference, as described in claim 5, is applied in the offline training phase, wherein the device comprises: The acquisition module is used to acquire the high-resolution range image of the frequency domain radar corresponding to each target under test, which is subject to interference, through the target radar. The inference module is used to input the interference-containing frequency domain radar high-resolution range image into the trained radar target recognition model, so as to generate the complex attention weights corresponding to the interference-containing frequency domain radar high-resolution range image through the complex frequency attention network of the radar target recognition model, and to perform a point-by-point multiplication operation on the interference-containing frequency domain radar high-resolution range image and the complex attention weights to obtain the corresponding filtered spectrum features. The identification module is used to input the filtered spectrum features into the radar high-resolution range image classifier to perform multi-layer nonlinear transformation on the filtered spectrum features and output the probability distribution corresponding to each target to be tested, so as to obtain the identification result corresponding to each target to be tested based on the probability distribution.

7. An electronic device, characterized in that, include: The system includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the radar target recognition method under non-uniform interference based on an attention network as described in any one of claims 1-3 or claim 4.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, The program is executed by the processor to implement the radar target recognition method under non-uniform interference based on attention networks as described in any one of claims 1-3 or claim 4.

Images