A radar anti-jamming method and device based on double-scale phase perception

Abstract

The application discloses a radar anti-interference method and equipment based on a double-scale phase perception, and the method comprises the following steps: acquiring a single-channel time-domain complex signal, inputting the signal into a trained double-scale high-fidelity phase perception network to filter out interference signals, and obtaining a target echo complex signal; the network is a network based on an encoder-separator-decoder structure, the encoder comprises a decoupling unit for IQ decoupling of the single-channel time-domain complex signal, the separator is used for estimating a target mask for filtering out heterogeneous interference from mixed features according to latent features output by the encoder, and the separator comprises a double-scale space-time joint modeling unit, the unit comprises a macroscopic branch for capturing global timing dependent features of long-range periodic interference and a microscopic branch for local phase detail repair and wideband noise suppression. The application solves the phase distortion pain point, and can adaptively process multiple types of heterogeneous interference with completely different time-frequency characteristics through a single network model.

Description

Technical Field

[0001] This invention belongs to the field of radar anti-jamming technology, specifically relating to a radar anti-jamming method and device based on dual-scale phase sensing. Background Technology

[0002] With the rapid advancements in Digital Radio Frequency Memory (DRFM) and Direct Digital Synthesis (DDS) technologies, modern active radar jamming techniques are evolving rapidly towards greater sophistication, diversification, and intelligence. Jammers can generate deceptive interference highly coherent with radar echoes, as well as broadband suppression jamming, drastically increasing the complexity of the electromagnetic environment faced by airborne radars. Based on the incident angle, radar electromagnetic interference is divided into sidelobe interference and mainlobe interference. For sidelobe interference, mature technologies such as adaptive sidelobe cancellation can effectively suppress it. However, mainlobe interference and target echoes highly overlap in the spatial domain, rendering traditional spatial filtering methods completely ineffective. Its suppression remains a core challenge in the field of radar signal processing.

[0003] Current research on radar main lobe interference suppression mainly falls into three categories: radar system innovation, signal processing algorithm optimization, and deep learning feature extraction. However, existing technologies have the following significant shortcomings in radar main lobe interference suppression: First, the deployment of the new radar system is difficult. The anti-jamming technology of the new radar system requires a large-scale upgrade of the hardware architecture, relies on high-precision system synchronization and multi-channel hardware, which is costly and difficult to deploy on existing airborne radars. It is also highly sensitive to antenna array errors and lacks engineering robustness.

[0004] Second, traditional algorithms are highly dependent on parameters. Traditional signal processing anti-interference algorithms are only suitable for ideal scenarios where the interference parameters are known. In dynamic, non-cooperative environments, the parameters are easily estimated inaccurately, which will not only significantly reduce the performance of the algorithm, but may even cancel out the target signal, making them unsuitable for real-world scenarios.

[0005] Third, existing deep learning methods are detached from the physical mechanisms of radar front-ends and cannot achieve temporal complex baseband waveform reconstruction. Most current deep learning-based radar signal processing solutions (such as existing dual-stream networks and spatiotemporal joint target detection models) focus on the back end of radar signal processing (such as after constant false alarm rate detection or after the range-Doppler domain). Their multi-branch structures are mainly used to fuse amplitude gating features and spatial trajectory representations after detection. These back-end feature extraction networks are essentially feature-level splicing with lost core phase information and cannot be applied to the front end of radar signal processing. When facing heterogeneous main lobe interference with overlapping spatial domains and extremely complex waveforms, these networks cannot perform blind source separation at the complex baseband physical level. Existing front-end deep learning anti-jamming technologies also suffer from multiple bottlenecks: ignoring the complex phase information of radar echoes leads to severe waveform distortion during recovery; designs primarily targeting single interference types result in poor scenario adaptability; high model computational overhead makes it difficult to meet the real-time processing requirements of airborne platforms; and they cannot balance the contradiction between capturing global long-range interference features and restoring local phase details, easily leading to incomplete interference suppression.

[0006] In summary, existing technologies for suppressing radar main lobe interference suffer from problems such as easily causing target phase distortion and difficulty in using a single model to handle heterogeneous interference. Summary of the Invention

[0007] To address the aforementioned problems in the prior art, this invention provides a radar anti-jamming method and device based on dual-scale phase sensing. The technical problem to be solved by this invention is achieved through the following technical solution: In a first aspect, the present invention provides a radar anti-jamming method based on dual-scale phase sensing, comprising: Acquire the single-channel time-domain complex signal to be processed; The single-channel time-domain complex signal to be processed is input into a trained dual-scale high-fidelity position sensing network to filter out interference signals and obtain the target echo complex signal. The dual-scale high-fidelity position sensing network is a radar anti-jamming network based on an encoder-splitter-decoder structure. The encoder includes a decoupling unit for IQ decoupling of a single-channel time-domain complex signal. The splitter utilizes the long-range dependence and phase sensitivity of the radar signal to estimate a target mask for filtering out heterogeneous interference from mixed features based on the latent features output by the encoder. The splitter includes a dual-scale spatiotemporal joint modeling unit, each containing a serial macroscopic branch and a microscopic branch. The macroscopic branch is used to capture the global temporal dependence features of long-range periodic interference; the microscopic branch is used for local phase detail restoration and broadband noise suppression.

[0008] Secondly, the present invention provides a radar anti-jamming device based on dual-scale phase sensing, including a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus; The memory is used to store computer programs; When the processor executes the computer program stored in the memory, it implements the steps of the radar anti-jamming method based on dual-scale phase sensing described above.

[0009] Thirdly, the present invention provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the radar anti-jamming method based on dual-scale phase sensing described above.

[0010] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention, through a micro-branched PReLU phase-preserving activation design and a dual-scale collaborative mechanism, perfectly preserves the fine micro-phase structure of radar echoes, fundamentally solving the phase distortion problem caused by existing deep learning masking methods and ensuring the reliability of subsequent high-precision ranging, velocity measurement, and Doppler processing. Simultaneously, this invention can adaptively handle multiple types of heterogeneous interference with drastically different time-frequency characteristics using a single network model, completely overcoming the core deficiency of poor adaptability in existing technologies. Unlike existing deep learning networks that focus on back-end feature-level target detection or trajectory prediction, this invention directly acts on the complex baseband time-domain signal at the radar's front end. The output of this invention is a pure complex baseband echo, rather than simply a detection probability or target coordinate. The purified complex signal can be seamlessly integrated into all standard radar back-end processing procedures (such as pulse compression, MTI, MTD, CFAR, etc.). This invention solves the more fundamental physical problem of "waveform source reconstruction," which, compared to methods that only perform fusion judgments at the feature layer, has more fundamental anti-interference physical significance and broader engineering application value.

[0011] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0012] Figure 1 This is a flowchart illustrating the radar anti-jamming method based on dual-scale phase sensing provided in an embodiment of the present invention. Figure 2 This is a schematic diagram of the overall architecture of the dual-scale high-fidelity position sensing network provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the architecture of the HyDRA Block provided in an embodiment of the present invention; Figure 4This is a schematic diagram of the architecture of FFConvM provided in an embodiment of the present invention; Figure 5 This is a schematic diagram of the architecture of 2-D Conv provided in an embodiment of the present invention; Figure 6 This is a visualization result of signal recovery under ISRJ interference provided in an embodiment of the present invention; Figure 7 This is a visualization result of signal recovery under CSNJ interference provided in an embodiment of the present invention; Figure 8 This is a visualization result of signal recovery under NFMJ interference provided in an embodiment of the present invention; Figure 9 This is a schematic diagram of the structure of a radar anti-jamming device based on dual-scale phase sensing provided in an embodiment of the present invention. Detailed Implementation

[0013] The present invention will be further described in detail below with reference to specific embodiments, but the implementation of the present invention is not limited thereto.

[0014] In a first aspect, embodiments of the present invention provide a radar anti-jamming method based on dual-scale phase sensing, which is applicable to radar main lobe interference scenarios.

[0015] Please see Figure 1 , Figure 1 This is a flowchart illustrating the radar anti-jamming method based on dual-scale phase sensing provided in an embodiment of the present invention. Figure 1 As shown, the method includes the following steps: S101. Obtain the single-channel time-domain complex signal to be processed.

[0016] Here, the single-channel time-domain complex signal to be processed can be a single single-channel time-domain complex signal or multiple single-channel time-domain complex signals; the present invention does not limit this.

[0017] The single-channel time-domain complex signal at the radar receiver includes the target echo signal, the main lobe interference signal, and additive white Gaussian noise. Its baseband complex signal model is as follows: ; in, This is a single-channel time-domain complex signal received by the radar. For the actual target echo signal, The main lobe interference signal It is additive white Gaussian noise.

[0018] S102. Input the single-channel time-domain complex signal to be processed into the trained dual-scale high-fidelity position sensing network (HyDRA) to filter out interference signals and obtain the target echo complex signal. The dual-scale high-fidelity position sensing network is a radar anti-jamming network based on an encoder-splitter-decoder structure. The encoder contains a decoupling unit for IQ decoupling of the single-channel time-domain complex signal. The splitter is used to estimate the target mask for filtering out heterogeneous interference from the mixed features based on the latent features of the encoder output, utilizing the long-range dependence and phase sensitivity of the radar signal. The splitter contains a dual-scale spatiotemporal joint modeling unit. Each dual-scale spatiotemporal joint modeling unit contains serial macroscopic branches and microscopic branches. The macroscopic branches are used to capture the global temporal dependence features of long-range periodic interference. The microscopic branches are used for local phase detail repair and broadband noise suppression.

[0019] The main lobe interference that this invention can handle includes, but is not limited to, three typical heterogeneous interferences: intermittent sampling forwarding interference (ISRJ), noise frequency modulation interference (NFMJ), and convolutional clever noise interference (CSNJ).

[0020] In this invention, the core objective of the proposed dual-scale high-fidelity position-aware network is to construct an end-to-end deep generative mapping function. With contaminated single-channel time-domain complex signal As input, output a high-fidelity estimate of the echo of the real target. The optimization objective is to minimize Compared with the true echo signal The error between them. The process of filtering out interference signals from the trained dual-scale high-fidelity position sensing network to obtain the target echo complex signal from the single-channel time-domain complex signal to be processed includes: S1021. Perform IQ decoupling on the single-channel time-domain complex signal to be processed to obtain the in-phase I component and the quadrature Q component.

[0021] This invention introduces an IQ decoupling mechanism for single-channel complex signals, strictly preserving the complete complex amplitude and phase physical information of radar signals in the channel dimension, thus avoiding the permanent loss of phase information caused by traditional time-frequency transformation or back-end detection processing. Since the principle of IQ decoupling for single-channel time-domain complex signals is an existing principle, it will not be elaborated upon here.

[0022] S1022. Perform feature encoding on the in-phase I component and the quadrature Q component to obtain the latent features.

[0023] S1023. Utilizing the long-range dependence and phase sensitivity of radar signals, a target mask for filtering out heterogeneous interference from mixed features is estimated based on latent characteristics.

[0024] S1024. Multiply the potential features and the target mask element by element to complete the feature separation of the interference and target signals, and obtain the separated target features.

[0025] For example, when the latent features output by the encoder are represented as And the target mask output by the separator is represented as At that time, the separated target features are represented as .

[0026] S1025. Decode the separated target features to transform them back into time-domain waveforms.

[0027] S1026. Perform IQ component synthesis on the time-domain waveform to obtain the target echo complex signal.

[0028] Since the principle of synthesizing IQ components of time-domain waveforms is an existing principle, it will not be elaborated here.

[0029] For example, Figure 2 This is a schematic diagram of the overall architecture of the dual-scale high-fidelity position-sensing network proposed in this invention. Figure 2 As shown, the network sequentially includes an encoder, a splitter, a first multiplication unit, and a decoder, wherein the first multiplication unit is... Figure 2 The middle part is located between the decoder and the splitter. The decoupling unit in the encoder was not in Figure 2 As shown in the diagram, the input to the decoupling unit is the single-channel time-domain complex signal to be processed. The output of the decoupling unit is connected to the input of the encoder. The output of the encoder is connected to the input of the separator on one hand and to one input of the first multiplication unit on the other. The output of the separator is connected to the other input of the first multiplication unit. The output of the first multiplication unit is connected to the input of the decoder. The output of the decoder serves as the output of the dual-scale high-fidelity position sensing network. The encoder is used to encode the in-phase I-component and quadrature Q-component obtained after decoupling to obtain latent features. The first multiplication unit is used to perform element-wise multiplication of the latent features and the target mask to complete the feature separation of the interference and target signals, obtaining the separated target features. The decoder is used to transform the separated target features back to the time-domain waveform through transposed convolution, and to synthesize the IQ components of the time-domain waveform to obtain the target echo complex signal.

[0030] In this invention, the encoder aims to replace the Short-Time Fourier Transform (STFT) and adaptively learn the basis functions of the signal. To preserve the complete complex information of the radar signal, the received single-channel radar signal is first processed... ( The number of sampling points (e.g., 6400) is decoupled into in-phase (I) and quadrature (Q) components in the channel dimension, and then passed through a one-dimensional convolutional layer (Conv1D) and a rectified linear unit (ReLU) activation function. Assuming the kernel size is L, the stride is L / 2, and the number of output feature channels is... N For example, the encoding process of an encoder can be represented by the following formula: ; ; in, Indicates the in-phase I component. Indicates orthogonal Q components, It is the generated latent feature representation. This refers to the number of feature frames. (Reference) Figure 2 As shown, when the size parameter of the radar signal input to the encoder is B × T (Right now B ×1× T When ), the size parameter of the feature output by the encoder is B × N × S ,in, B This indicates the batch size, representing the number of single-channel radar signals in the input encoder.

[0031] In this invention, the separator aims to estimate a target mask used to filter out heterogeneous interference from mixed features. The separator's design fully considers the long-range dependence and phase sensitivity of radar signals. Features are first projected through layer normalization and point convolution, and positional encoding for global information is added to the projected features, outputting projected features with positional information. This is crucial for combating periodic slice interference. The sequence then enters a stacked HyDRABlock. The HyDRA Block is a dual-scale spatiotemporal joint modeling unit that innovatively combines "RoPE-enhanced linear attention" for capturing the periodicity of macroscopic long-range signals with "gated dilated convolution" for repairing microscopic local phase details, thereby achieving high-fidelity waveform recovery under heterogeneous main lobe interference. After deep feature extraction by the HyDRA Blocks, the output features are transformed through a series of nonlinear transformations to generate the final target mask. For example, the processing principle of the separator is as follows: After performing layer normalization, point convolution, adding positional encoding for global information, and reshaping on the latent features sequentially, R stacked dual-scale spatiotemporal joint modeling units are used to extract dual-scale features from the reshaped features. Then, a series of nonlinear transformations are performed on the extracted features to obtain the target mask, where R is a preset positive integer greater than 1. It should be noted that the value of R can be set according to actual needs, for example, it can be 10, 12, 24, etc., and this invention does not limit this. For example, see [link to example]. Figure 2 The separator sequentially comprises: a first-layer normalization unit, a first-point convolutional unit, a fixed-position encoding unit, a first concatenation block, a first reshaping unit, R stacked dual-scale spatiotemporal joint modeling units, a second-layer normalization unit, a second reshaping unit, a PReLU activation function, a second-point convolutional unit, a third reshaping unit, a third-point convolutional unit, a fourth-point convolutional unit, a gated linear unit (GLU), a fifth-point convolutional unit, and a ReLU activation function. The third and fourth point convolutional units are implemented in parallel. The input of the first-layer normalization unit serves as the input to the separator, and the output of the ReLU activation function serves as the output of the separator. It should be noted that... Figure 2 In this context, LayerNorm represents layer normalization unit, Sinusoidal Position Encoding represents fixed-position encoding unit, Pointwise Convolution represents point convolution unit, Reshape represents reshaping unit, and HyDRA Block represents dual-scale spatiotemporal joint modeling unit. "" represents a splicing block, and GLU represents a gated linear unit. Since all units or blocks except for the HyDRA Block are existing units or blocks, their specific structures will not be described in detail in this invention. GLU can effectively control the information flow and suppress non-target energy. The ReLU activation function ensures that the target mask is non-negative. Figure 2 In B × N × S, B ×(C× N) × S、 ( B ×C)× N × S, C × B × N × S C represents the size parameter of the output / input characteristics of the corresponding unit or block, and C represents the number of masks output by the separator. It should be noted that the separator separates the mixed signal into the target signal and other signals, thus generating target masks and other masks, so the number of C is 2. However, the focus of this invention is on the target, and other factors, including interference and noise, are not the focus of this application.

[0032] In this invention, the HyDRA Block is the core computational unit of the proposed network, aiming to jointly model the macroscopic long-range dependence and microscopic local details of radar signals. The HyDRA Block includes serial macroscopic and microscopic branches; the macroscopic branch adopts a temporal shift dual-stream hybrid attention mechanism with rotational position coding (RoPE), and performs cross-multiplication between the attention-enhanced feature stream and the original projected feature stream to adaptively adjust the information stream according to the signal-to-interference ratio (SIR) and capture the global temporal dependence features of long-range periodic interference; the microscopic branch adopts a gated dilated convolutional structure, and the convolutional structure embeds instance normalization and PReLU activation functions to adapt to the large dynamic range characteristics of radar signals and retain the negative half-axis phase information, achieving high-fidelity restoration of local waveforms and phases. In this invention, each dual-scale spatiotemporal joint modeling unit is specifically used for: performing temporal shift embedding, offset-scaling-based shared projection, dual-stream hybrid attention-based aggregation, and cross-gated output fusion processing on the input features of the macro-branch to obtain the output feature tensor of the macro-branch; sequentially performing one-dimensional convolution, activation, and layer normalization on the output feature tensor of the macro-branch to obtain the input features of the micro-branch; and then performing feedforward convolution, dilated convolution-based gated convolution, layer normalization, and 1×1 convolution on the input features of the micro-branch to obtain the output features of the dual-scale spatiotemporal joint modeling unit. This dual-branch structure is not simply for concatenating backend detection features, but rather for collaborative decoupling of the physical differences in heterogeneous main lobe interference: the macro-branch captures the global temporal dependence of long-range periodic interference, while the micro-branch achieves local phase detail restoration and broadband noise suppression. Based on this dual-scale feature estimation, a non-negative mask is obtained. This mask is multiplied element-wise with high-dimensional complex features to construct an adaptive time-varying filter in the latent space, thereby achieving hard stripping of complex baseband level interference from the physical components of the target signal.

[0033] In this invention, each HyDRA Block sequentially includes: a macroscopic branch, a first 1×1 convolutional layer, a PReLU activation function, a third normalization unit, and a microscopic branch; the inputs and outputs of the macroscopic branch, the first 1×1 convolutional layer, the PReLU activation function, the third normalization unit, and the microscopic branch are connected sequentially, and the input of the macroscopic branch serves as the input of the HyDRA Block, and the output of the microscopic branch serves as the output of the HyDRA Block. The macroscopic branch includes: a temporal shift unit, a first feedforward convolution unit, a second feedforward convolution unit, a third feedforward convolution unit, a positional encoding scaling and translation operation unit, a local-global hybrid attention unit, a second multiplication unit, a third multiplication unit, a fourth multiplication unit, a sigmoid activation function, a fourth feedforward convolution unit, and a second concatenation block. The output of the temporal shift unit is connected to the inputs of the first, second, and third feedforward convolution units and the second concatenation block, respectively. The input of the temporal shift unit is the input of the macroscopic branch. The output of the first feedforward convolution unit is connected to the input of the positional encoding scaling and translation operation unit, and the output of the second feedforward convolution unit is simultaneously connected to the input of the second concatenation block. The inputs of the position encoding scaling and translation operation unit and the second multiplication unit are connected. The output of the third feedforward convolution unit is connected to both the inputs of the position encoding scaling and translation operation unit and the third multiplication unit. The outputs of the position encoding scaling and translation operation unit are connected to the inputs of the second and third multiplication units, respectively. The output of the second multiplication unit, after passing through a Sigmoid activation function, is connected to the input of the fourth multiplication unit. The output of the third multiplication unit is connected to the input of the fourth multiplication unit. The output of the fourth multiplication unit is connected to the input of the fourth feedforward convolution unit. The output of the fourth feedforward convolution unit is connected to the input of the second concatenation block. The output of the second concatenation block serves as the output of the macro-branch. The micro-branch includes: the fifth feedforward convolution unit, the sixth feedforward convolution unit, the first linear layer, the ReLU activation function, the second linear layer, and 2... L-1 The system consists of a series of deep-hole 2D convolutional blocks, a third concatenation block, a fifth multiplication unit, a fourth concatenation block, a fourth normalization unit, and a second 1×1 convolutional layer. The input of the fifth feedforward convolutional unit is connected to the input of the sixth feedforward convolutional unit and the input of the fourth concatenation block. The input of the fifth feedforward convolutional unit serves as the input to the micro-branch. The output of the fifth feedforward convolutional unit is connected to the input of the fifth multiplication unit. The output of the sixth feedforward convolutional unit is simultaneously connected to the input of the third concatenation block and the input of the first linear layer. The system also includes the first linear layer, the ReLU activation function, the second linear layer, and a second 1×1 convolutional layer. L-1 The inputs and outputs of a series of deeply dilated 2D convolutional blocks are sequentially connected, and there is a residual connection between adjacent deeply dilated 2D convolutional blocks, as well as between the first deeply dilated 2D convolutional block and the second... L-1 Residual connections between two-dimensional convolutional blocks with deep holes, the second L-1The output of the first deep-hole 2D convolutional block is connected to the input of the third concatenation block. The output of the third concatenation block is connected to the input of the fifth multiplication unit. The output of the fifth multiplication unit is connected to the input of the fourth concatenation block. The output of the fourth concatenation block, the input of the fourth normalization unit, and the input of the second 1×1 convolutional layer are sequentially connected. L is a preset positive integer. It should be noted that the value of L can be set according to actual needs, and this invention does not limit it. d This indicates the dilation rate of the dilated convolution. It should be noted that the use of terms like "first" and "second" in this invention is merely for distinguishing objects with the same name and has no other meaning. For example, "first multiplication unit," "second multiplication unit," "third multiplication unit," and "fourth multiplication unit" represent four multiplication units; "first 1×1 convolutional layer" and "second 1×1 convolutional layer" represent two 1×1 convolutional layers, and so on.

[0034] For example, Figure 3 This is a schematic diagram of the architecture of the HyDRA Block provided in an embodiment of the present invention. It should be noted that... Figure 3 In this context, 1×1 Conv represents a 1×1 convolutional layer, FFConvM represents a feedforward convolutional unit, and Rope&Scale&Offset represents a position encoding, scaling, and translation operation unit, which is the unit that performs position encoding, scaling, and translation operations. Joint AttentionLocal&Global represents a local-global hybrid attention unit, Linear represents a linear layer, and 2-D Conv is short for 2-D Conv Block, representing a depthwise dilated 2D convolutional block.

[0035] In this invention, FFConvM is a designed feedforward convolutional module for extracting fine-grained local features. This module constructs a hierarchical feature projection mechanism through sequential normalization, linear transformation, SiLU activation function, and convolution operations. FFConvM sequentially includes: a scaling normalization unit, a third linear layer, a SiLU activation function, a one-dimensional depthwise convolutional unit, a fifth concatenation block, and a regularization layer. The inputs and outputs of the scaling normalization unit, linear layer, SiLU activation function, one-dimensional depthwise convolutional unit, fifth concatenation block, and regularization layer are sequentially connected. The input of the scaling normalization unit serves as the input of FFConvM, and the output of the regularization layer serves as the output of FFConvM. For example, Figure 4This is the architecture diagram of FFConvM. ScaleNorm represents the scaling and normalization unit, Linear Layer represents a linear layer, 1D Depthwise Conv represents a one-dimensional depthwise convolutional unit, and Dropout represents a regularization layer. It should be noted that these layers or units are all existing, so their principles will not be elaborated upon here. Compared to standard linear layers, the convolutional structure introduced by FFConvM can effectively capture the local contextual dependencies of sequential data. Combined with the smooth gating characteristics of the SiLU activation function, it improves the model's expressiveness while maintaining training stability.

[0036] In this invention, the 2-D Conv is also a carefully designed module. The 2-D Conv contains meticulously designed normalization and activation paths to accommodate the large dynamic range of radar signals. The 2-D Conv sequentially includes: a constant-padding block, a two-dimensional dilated convolutional block, an instance normalization layer, a PReLU activation function, and a sixth stitching block. The inputs and outputs of the constant-padding block, the two-dimensional dilated convolutional block, the instance normalization layer, the PReLU activation function, and the sixth stitching block are connected sequentially. The input of the constant-padding block serves as the input of the 2-D Conv, and the output of the sixth stitching block serves as the output of the 2-D Conv. For example, Figure 5 This is a 2-D Conv architecture diagram, where ConstantPad2d represents a constant padding block, Conv2d represents a 2D (two-dimensional) dilated convolution block, InstanceNorm2d represents an instance normalization layer, and Concatenate represents a concatenation block.

[0037] The following is based on Figure 3 This paper introduces the working principle of HyDRA Block. The main task of the macroscopic branch is to establish a long-range dependency model of radar echoes to identify the global periodic characteristics of signals and interference (such as ISRJ). This invention uses linear attention as the core operator, reducing computational complexity from... Reduce to Meanwhile, to compensate for the potential shortcomings of purely linear attention in capturing microscopic phase details, this invention proposes a dual-stream hybrid attention mechanism. As shown in Figure 3, this macroscopic branch consists of a timing shift unit ( Figure 3 It consists of stages such as (not shown in the diagram), shared projection, hybrid attention computation, and gating fusion. For example... Figure 3 As shown: The first stage, timing shift embedding: When the input features of the HyDRA Block are To capture adjacent phase dependencies and reduce the number of parameters, the input features (i.e., the input feature tensor) are first temporally shifted using a temporal shift unit to generate the temporally shifted features. After that, the features Three FFConvMs sharing the same parameters are input separately, where the first FFConvM is based on features. Generate shared features The other two FFConvMs are based on features Each feature stream is generated individually, resulting in a dual feature stream. , ,and, , , The dimensional parameters are respectively B × S × A , B × S × 2N , A The number of channels representing the feature, i.e. , . , , , The expressions are as follows: , , , Representation of features The first half, This indicates misaligned channels, and the two are spliced ​​together to achieve the effect of overlapping adjacent features.

[0038] The second stage involves shared projection based on offset scaling: The first FFConvM will share features After inputting Rope&Scale&Offset, Rope&Scale&Offset first uses a learnable scaling factor. Shared features Scaling, then utilizing learnable offsets. After offsetting, it is divided into 4 blocks to obtain and Based on this, then for all and By applying Rotated Position Encoding (RoPE) to explicitly inject the absolute timing and relative phase information of the radar echo, it is possible to obtain... , These represent local query features, global query features, local key features, and global key features, respectively.

[0039] The third stage, dual-stream hybrid attention convergence: In obtaining and dual-path feature flow , ,Will as well as , Input Joint AttentionLocal&Global. To strike a balance between "macro-global modeling" and "micro-detail preservation," Joint AttentionLocal&Global executes two attention paths in parallel: Path A: local secondary attention, and Path B: global linear attention. In Path A, standard dot-product attention is computed within the local window G, aiming to capture high-frequency waveform details to obtain local value features. and local value features , and The calculation formulas are as follows: , , Where T is the transpose symbol. This represents the local attention weight matrix. This is the scaling factor for the attention mechanism, typically taken as the dimension of the attention head. In path B, the global context matrix is ​​first aggregated to capture the signal periodicity of the entire sequence, yielding global value features. and global value features , and The expressions are as follows: , , , ,in, Indicates the first n Global key features at each position, Indicates the first n Value characteristics of each position, Indicates the characteristic of the value at the second nth position, Represents the first global context matrix, This represents the second global context matrix. n Indicates the first n The position (i.e., the first position) n (each time step), the symbols in these formulas This represents the outer product used to compute the global context matrix. Finally, by directly superimposing the local and global components, the attention-enhanced two-stream features can be obtained. and , .

[0040] Phase 3, cross-gated output fusion: Unlike the standard GAU, this invention uses attention-enhanced feature streams ( and ) and the original projected feature flow ( Cross-multiplication is performed, a design that allows the network to adaptively adjust the information flow based on the signal-to-interference ratio (SIIR), and then output the result via residual connections. , ,in, Represents the element-wise product of a matrix or vector (Hadamard Product). It is the Sigmoid activation function. This represents the output feature tensor, i.e. This represents the output feature tensor of the macroscopic branch. This indicates the cross-gated output characteristic.

[0041] The macro branch constructs global temporal dependencies through temporal shift embedding, shared projection, and local-global dual-stream attention. Furthermore, it introduces a cross-gated fusion output mechanism, which cross-multiplies the attention-enhanced feature stream with the original projected feature stream to adaptively adjust the information stream according to the signal-to-interference ratio (SJR) of the mixed echo. This enables the accurate identification and elimination of global features of periodic coherent interference such as intermittent sample-and-forward interference (ISRJ).

[0042] Continue to refer to the above. Figure 3 The 1×1 Conv, PReLU, and LayerNorm layers following the macro branch constitute the bottleneck layer. The main purpose of the bottleneck layer is to reduce the embedding dimension while preserving key features. The output of the macro branch... First, a 1×1 convolutional layer is applied, followed by a parameterized correction unit activating PReLU and a layer normalization layer to obtain the dimensionality-reduced features. ,in This indicates the dimension of the dimensionality reduction embedding.

[0043] Continue to refer to the above. Figure 3 While macroscopic branches can effectively capture the global periodicity of signals and interference, the microscopic waveform structure of radar echoes is often severely disrupted by interference such as CSNJ and NFMJ, leading to phase discontinuities and spectral dispersion. To address this microscopic challenge, this invention introduces a cyclic module based on a finite state machine network (FSMN) without loop connections as the microscopic branch. Figure 3 As shown, the micro-branchings include FFConvM, Linear, ReLU, 2-D Conv, and " "and" "This forms a gated convolutional unit (GCU), and LayerNorm and 1×1 Conv form the output layer. The GCU mainly processes the input features..." Mapped to output The core module consists of stacked 2D Convs. For example... Figure 5As shown, unlike standard convolutions, each 2D Conv contains carefully designed normalization and activation paths to accommodate the large dynamic range of radar signals. The two FFConvMs in the GCU are based on the input features... The features output separately and The expression, and the characteristics of the GCU output. The expressions are as follows: , , , ,in, This represents a dilated dense convolutional network. Specifically, it represents network input characteristics, Indicates gating features, This represents the gated output characteristics. After obtaining the gated output characteristics... Then, the features Input to output layer, features The final output is obtained by passing the material through a normalization layer and a 1×1 convolutional layer. , This refers to the output characteristics of the micro-branch.

[0044] In addition to expanding the receptive field through a densely connected exponentially dilated deep convolutional structure, the micro-branch specifically introduces instance normalization (InstanceNorm) and PReLU activation functions into its network nodes. This structure is specifically adapted to the large dynamic range characteristics of radar radio frequency signals and completely preserves the negative half-axis phase information of complex signals during nonlinear mapping, ensuring high-fidelity restoration of local radar waveforms and micro-phase.

[0045] In this invention, the decoder performs the inverse process of encoding. For example... Figure 3 As shown, the decoder uses transposed convolution (Transposed Conv1D) to convert the features output by the multiplication unit after the separator into features. Transform back to the time domain waveform to obtain the recovered target time domain waveform. , After that, then... IQ component synthesis is performed to obtain the target echo complex signal.

[0046] The core problem that HyDRA, proposed in this invention, aims to solve can be formally defined as: designing a deep generative mapping function. The function receives a heavily contaminated signal. As input, output is the echo of the real target. A high-fidelity estimate ,Right now: ,in, This represents the learnable parameters of HyDRA. To achieve high-fidelity waveform reconstruction and effectively suppress nonlinear phase distortion, the optimization objective is to maximize the similarity in waveform structure between the estimated echo and the true echo. This invention uses the scale-invariant signal-to-noise ratio (SI-SNR) as the loss function for HyDRA. Specifically, end-to-end training is performed by minimizing the negative SI-SNR: The calculation process for SI-SNR is as follows: First, the estimated echo... Projected onto real echo To obtain scale-invariant target components in the direction of... Then, the energy ratio of the target component to the residual noise component is calculated: ; ; ; in, Denotes the square of the L1 norm. This represents the inner product of two signals. Through the strict constraints of the SI-SNR loss function, HyDRA can focus on extracting the coherent manifold of the target in complex high-entropy interference backgrounds, preserving the local phase information of the original echo to the greatest extent.

[0047] To systematically evaluate the signal separation performance and waveform recovery fidelity of the proposed HyDRA framework in complex heterogeneous electromagnetic environments, a series of comprehensive simulation experiments were designed.

[0048] 1) Dataset Considering the non-cooperative nature of radar main lobe interference, we constructed a high-fidelity simulation dataset based on radar echo equations and interference mechanisms. The dataset covers linear frequency modulated (LFM) target echoes and three types of heterogeneous main lobe interference (ISRJ, NFMJ, CSNJ).

[0049] The simulation parameters are shown in Table 1. We generated a total of 9600 pulse samples (corresponding to 150 coherent processing intervals, with each CPI containing 64 pulses). To simulate realistic dynamic battlefield scenarios, the jamming-to-signal ratio (JSR) was randomly sampled from 0 dB to 30 dB, and the signal-to-noise ratio (SNR) varied from -5 dB to 15 dB. The dataset was randomly divided into training, validation, and test sets in a 7:2:1 ratio.

[0050] Table 1 Dataset Parameter Configuration

[0051] 2) Evaluation Indicators The signal-to-interference-plus-noise ratio (SINR) improvement factor (SINRIF) is the most intuitive indicator of interference suppression capability, defined as the output signal-to-interference-plus-noise ratio after separation. ) and input signal-to-interference-plus-noise ratio (SINR) The difference is ). Its mathematical expression is: , in, For a pure target echo, The signal recovered by HyDRA. This represents a mixed observation signal containing interference. A higher SINRIF value indicates more thorough interference suppression.

[0052] 3) Simulation Results and Analysis I. Quantitative Evaluation of Anti-interference Performance To comprehensively evaluate the HyDRA framework's ability to suppress heterogeneous main lobe interference, we conducted comparative experiments with five representative benchmark methods. These five methods cover traditional time-frequency analysis (STFT filtering, STFRFT filtering), deep recurrent networks (DPIndRNN), complex convolutional networks (CVDPCS-TssNet), and generative adversarial networks (GAN).

[0053] Table 2 Average values ​​for various methods

[0054]

[0055] Table 2 quantitatively presents the SINRIF performance of each method under different interference scenarios. The results show that HyDRA achieves significantly optimal performance in all test scenarios. Compared with the suboptimal GAN ​​method, HyDRA achieves performance gains of 7.78 dB, 16.64 dB, and 12.79 dB in ISRJ, NFMJ, and CSNJ scenarios, respectively. This significant performance leap is mainly attributed to HyDRA's unique dual-scale architecture: the macroscopic branch captures the long-distance, coarse-grained dependencies of the signal, accurately identifying and eliminating the periodic slice features of ISRJ; while the microscopic branch focuses on modeling fine-grained phase and spectral details, effectively suppressing noise-like interference with complex time-frequency characteristics such as NFMJ and CSNJ.

[0056] II. Qualitative Visualization and Target Detection Performance Analysis To visually verify the reconstruction effectiveness of the HyDRA framework in the multidimensional signal domain, Figures 6 to 7 are provided. Figure 8 The results of signal recovery visualization are presented under three typical interference scenarios (JSR = 20 dB).

[0057] First, for intermittent sampling-forwarding interference (ISRJ, Fig. 6) with periodic modulation characteristics, the time-frequency plot of the mixed echo (Fig. 6, plot (d)) and the pulse compression result (Fig. 6, plot (g)) show severe energy leakage, resulting in a dense group of false targets on the range-Doppler (RD) spectrum. Figure 6 The (i) image in the image obscures the real target. After HyDRA processing, the missing slices in the temporal domain are accurately filled in. Figure 6 In Figure (c), false targets on the RD spectrum were completely removed, leaving only the single true target peak. Figure 6 (See Figure (k)). This result strongly demonstrates that the macro-branch, through a global attention mechanism, successfully captured and stripped away the long-range periodicity of the ISRJ, thus eliminating the false alarm risk of the radar system.

[0058] Secondly, for noise frequency modulation interference (NFMJ) with broadband characteristics, Figure 8 ) and Convolutional Smart Noise (CSNJ, Figure 7 The interference energy covers the entire frequency band in the time-frequency domain (Figure 7(d) and Figure 8(d)), causing the linear frequency modulation trajectory of the LFM signal to be completely submerged, resulting in an extremely low signal-to-interference-plus-noise ratio. HyDRA demonstrates excellent feature decoupling and denoising capabilities: in the recovered time-frequency diagram ( Figure 7 Figure (f) in the middle, Figure 8 In Figure (f), complex background noise is effectively filtered out, clearly restoring the sparse "diagonal" time-frequency manifold of the target signal. This indicates that micro-branching, utilizing dilated dense convolution, can extract deterministic signal structures from high-entropy noise backgrounds, achieving blind source separation in complex electromagnetic environments.

[0059] Finally, to verify the phase consistency of the signal recovery, we focused on analyzing the fidelity check results of pulse compression (PC) (Figure (h)). Pulse compression is extremely sensitive to phase errors; any nonlinear phase distortion will cause main lobe widening or sidelobe elevation. Figure 6 As shown in Figures (h) in sections 7, 7, and 8, the pulse compression curves of the HyDRA-recovered signal highly overlap with the true values ​​in terms of both main lobe width and side lobe level. The magnified views further confirm that the recovered signal does not exhibit asymmetrical side lobe elevation or main lobe splitting. This high waveform fidelity demonstrates that HyDRA not only recovers the signal amplitude but also perfectly preserves the fine microscopic phase structure of the radar echo, ensuring the reliability of subsequent high-precision ranging and Doppler processing.

[0060] III. Computational Complexity Analysis Table 3 compares the performance of HyDRA and the baseline model in terms of computational complexity (FLOPs), space complexity (model size), and inference latency. The experimental platform was a single NVIDIA 5070ti. Compared to DPIndRNN based on recurrent neural networks, HyDRA's inference speed is improved by approximately 60 times (22.91 ms vs. 1320.66 ms). Compared to generative adversarial networks (GANs), HyDRA maintains state-of-the-art performance while reducing the model size by 88% (13.28 MB vs. 114.10 MB). This means that HyDRA has extremely low storage overhead and is easily deployed on resource-constrained edge computing devices or airborne radar platforms. Although CVDPCS-TssNet has a slight advantage in pure inference speed, its computational cost (FLOPs) is 2.2 times that of HyDRA (3260 M vs. 1460 M), which means higher energy consumption. HyDRA achieves the significant performance improvement (SINRIF improvement of 10dB+) mentioned above with moderate computing resource consumption, achieving the best balance between computing efficiency, storage cost and reconstruction accuracy.

[0061] Table 3 Comparison of Complexity of Different Models

[0062] Compared with existing technologies in the field of radar main lobe interference suppression (including traditional algorithms such as STFT / STFRFT filtering, and deep learning algorithms such as DPIndRNN, CVDPCS-TssNet, GAN, and various multi-branch target detection networks), this invention has the following significant technical advantages: 1) Overcoming the limitations of backend detection algorithms to achieve a fundamental purification of the frontend signal: Unlike existing deep learning networks that focus on backend feature-level target detection or trajectory prediction, this invention directly acts on the complex baseband time-domain signal at the radar's front end. The output of this invention is a pure complex baseband echo, rather than simply detection probability or target coordinates. The purified complex signal can be seamlessly integrated into all standard radar backend processing procedures (such as pulse compression, MTI, MTD, CFAR, etc.). This invention addresses the more fundamental physical problem of "waveform source reconstruction," which, compared to methods that only perform fusion and judgment at the feature layer, has more fundamental anti-interference physical significance and broader engineering application value.

[0063] 2) Comprehensive superior anti-interference performance, achieving unified and efficient suppression of heterogeneous interference: In three typical heterogeneous main lobe interference scenarios—ISRJ, NFMJ, and CSNJ—this invention achieves average signal-to-interference-plus-noise ratio (SINRIF) improvements of 47.60 dB, 52.73 dB, and 51.02 dB, respectively. Compared with the best-performing existing GAN methods, this invention achieves significant performance gains of 7.78 dB, 16.64 dB, and 12.79 dB, respectively. Furthermore, this invention can adaptively handle multiple types of heterogeneous interference with drastically different time-frequency characteristics using a single network model, completely overcoming the core deficiency of poor scenario adaptability in existing technologies.

[0064] 3) Extremely enhanced waveform and phase fidelity, ensuring core radar detection performance: This invention perfectly preserves the fine microscopic phase structure of the radar echo through a micro-branched PReLU phase-preserving activation design and a dual-scale collaborative mechanism. In pulse compression processing, the main lobe width and side lobe level of the physically reconstructed signal highly coincide with the real target, without any asymmetric increase in side lobes or main lobe splitting. This fundamentally solves the phase distortion problem caused by existing deep learning masking methods, ensuring the reliability of subsequent high-precision ranging, velocity measurement, and Doppler processing by the radar.

[0065] 4) Lightweight design to meet the real-time deployment requirements of airborne platforms: This invention achieves extreme lightweight design through the optimization of linear attention and parallel convolutional structures: the model size is only 13.28MB, which is 88% smaller than the GAN model, greatly reducing storage overhead; the single-pulse inference latency is only 22.91ms, which is about 60 times better than the DPIndRNN recurrent network, fully meeting the stringent constraints of pulse-level real-time processing in airborne radar systems.

[0066] 5) The technical solution has strong scalability: The dual-scale feature decoupling architecture of the present invention can not only handle the three typical types of interference described in this embodiment, but also can be quickly adapted to more unknown interference types such as new smart interference and composite interference through dataset expansion. It has strong technical scalability and iteration space, and can cope with the evolution and upgrading of interference technology in electronic warfare for a long time.

[0067] Secondly, embodiments of the present invention provide a radar anti-jamming device based on dual-scale phase sensing.

[0068] Please see Figure 9 , Figure 9This is a schematic diagram of the structure of a radar anti-jamming device based on dual-scale phase sensing provided in an embodiment of the present invention. The device includes: a processor 901, a communication interface 902, a memory 903, and a communication bus 904. The processor 901, communication interface 902, and memory 903 communicate with each other via the communication bus 904. The memory 903 stores computer programs. When the processor 901 executes the computer program stored in the memory 903, it implements the steps of the aforementioned radar anti-jamming method based on dual-scale phase sensing.

[0069] The communication bus mentioned in the above electronic devices can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. This communication bus can be divided into address bus, data bus, control bus, etc. For ease of illustration, only one thick line is used to represent it in the diagram, but this does not mean that there is only one bus or one type of bus.

[0070] The communication interface is used for communication between the aforementioned electronic devices and other devices.

[0071] The memory may include random access memory (RAM) or non-volatile memory (NVM), such as at least one disk storage device. Optionally, the memory may also be at least one storage device located remotely from the aforementioned processor.

[0072] The processors mentioned above can be general-purpose processors, including central processing units (CPUs), network processors (NPs), etc.; they can also be digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.

[0073] The method provided in this invention can be applied to electronic devices. Specifically, the electronic device can be a desktop computer, a portable computer, a smart mobile terminal, a server, etc. No limitation is made herein; any electronic device that can implement this invention falls within the protection scope of this invention.

[0074] For the embodiments of the device / electronic device / storage medium, since they are basically similar to the method embodiments, the description is relatively simple, and relevant parts can be referred to in the description of the method embodiments.

[0075] It should be noted that the electronic device and storage medium in the embodiments of the present invention are respectively electronic devices and storage media that apply the above-mentioned radar anti-jamming method based on dual-scale phase sensing. Therefore, all embodiments of the radar anti-jamming method based on dual-scale phase sensing are applicable to the device, electronic device and storage medium, and can achieve the same or similar beneficial effects.

[0076] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0077] 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 the invention. 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. In addition, those skilled in the art can combine and integrate the different embodiments or examples described in this specification.

[0078] Those skilled in the art will understand that embodiments of this application can be provided as methods, apparatus (devices), or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects, all of which are collectively referred to herein as "modules" or "systems." Furthermore, this application can take the form of a computer program product implemented on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code. The computer program may be stored / distributed in a suitable medium, provided with or as part of other hardware, or may take other distribution forms, such as via the Internet or other wired or wireless telecommunications systems.

[0079] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (devices), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0080] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0081] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0082] The above description, in conjunction with specific preferred embodiments, provides a further detailed explanation of the present invention. It should not be construed that the specific implementation of the present invention is limited to these descriptions. For those skilled in the art, various simple deductions or substitutions can be made without departing from the concept of the present invention, and all such modifications and substitutions should be considered within the scope of protection of the present invention.

Claims

1. A radar anti-jamming method based on dual-scale phase sensing, characterized in that, include: Acquire the single-channel time-domain complex signal to be processed; The single-channel time-domain complex signal to be processed is input into a trained dual-scale high-fidelity position sensing network to filter out interference signals and obtain the target echo complex signal. The dual-scale high-fidelity position sensing network is a radar anti-jamming network based on an encoder-splitter-decoder structure. The encoder includes a decoupling unit for IQ decoupling of a single-channel time-domain complex signal. The splitter utilizes the long-range dependence and phase sensitivity of the radar signal to estimate a target mask for filtering out heterogeneous interference from mixed features based on the latent features output by the encoder. The splitter includes a dual-scale spatiotemporal joint modeling unit, each containing a serial macroscopic branch and a microscopic branch. The macroscopic branch is used to capture the global temporal dependence features of long-range periodic interference; the microscopic branch is used for local phase detail restoration and broadband noise suppression.

2. The radar anti-jamming method based on dual-scale phase sensing according to claim 1, characterized in that, The process by which the trained dual-scale high-fidelity position sensing network filters out interference signals from the single-channel time-domain complex signal to be processed, obtaining the target echo complex signal, includes: The single-channel time-domain complex signal to be processed is decoupled by IQ to obtain in-phase I components and quadrature Q components; The in-phase I component and the orthogonal Q component are feature-encoded to obtain latent features; By utilizing the long-range dependence and phase sensitivity of radar signals, a target mask for filtering out heterogeneous interference from mixed features is estimated based on the latent features. The potential features and the target mask are multiplied element-wise to separate the interference and target signals, resulting in the separated target features. The separated target features are decoded to transform them back into time-domain waveforms; The time-domain waveform is synthesized using IQ components to obtain the target echo complex signal.

3. The radar anti-jamming method based on dual-scale phase sensing according to claim 1, characterized in that, The input of the decoupling unit is the single-channel time-domain complex signal to be processed. The output of the decoupling unit is connected to the input of the encoder. The output of the encoder is connected to the input of the separator on one hand and to one input of the first multiplication unit on the other hand. The output of the separator is connected to the other input of the first multiplication unit. The output of the first multiplication unit is connected to the input of the decoder. The output of the decoder is used as the output of the dual-scale high-fidelity position sensing network. The encoder is used to encode the in-phase I component and quadrature Q component obtained after decoupling to obtain latent features; The first multiplication unit is used to perform element-wise multiplication of the latent features and the target mask to complete the feature separation of the interference and target signals, and obtain the separated target features; The decoder is used to transform the separated target features back into a time-domain waveform through transposed convolution, and to synthesize the time-domain waveform using IQ components to obtain the target echo complex signal.

4. The radar anti-jamming method based on dual-scale phase sensing according to claim 1, characterized in that, The separator is specifically used to: sequentially perform layer normalization, point convolution, add position encoding for global information, and reshape the potential features, then use R stacked dual-scale spatiotemporal joint modeling units to extract dual-scale features from the reshaped features, and then perform a series of nonlinear transformations on the extracted features to obtain the target mask, where R is a preset positive integer greater than 1.

5. The radar anti-jamming method based on dual-scale phase sensing according to claim 1 or 4, characterized in that, The separator sequentially comprises: a first-layer normalization unit, a first-point convolutional unit, a fixed-position encoding unit, a first splicing block, a first reshaping unit, R stacked dual-scale spatiotemporal joint modeling units, a second-layer normalization unit, a second reshaping unit, a PReLU activation function, a second-point convolutional unit, a third reshaping unit, a third-point convolutional unit, a fourth-point convolutional unit, a gated linear unit, a fifth-point convolutional unit, and a ReLU activation function. The third-point convolutional unit runs in parallel with the fourth-point convolutional unit. The input of the first-layer normalization unit serves as the input of the separator, and the output of the ReLU activation function serves as the output of the separator.

6. The radar anti-jamming method based on dual-scale phase sensing according to claim 1 or 4, characterized in that, Each of the dual-scale spatiotemporal joint modeling units is specifically used to: perform temporal shift embedding, offset scaling-based shared projection, two-stream hybrid attention-based aggregation, and cross-gated output fusion processing on the input features of the macro-branch to obtain the output feature tensor of the macro-branch; perform one-dimensional convolution, activation, and layer normalization processing on the output feature tensor of the macro-branch in sequence to obtain the input features of the micro-branch; and perform feedforward convolution, dilated convolution-based gated convolution, layer normalization, and 1×1 convolution processing on the input features of the micro-branch to obtain the output features of the dual-scale spatiotemporal joint modeling unit.

7. The radar anti-jamming method based on dual-scale phase sensing according to claim 1 or 4, characterized in that, Each of the dual-scale spatiotemporal joint modeling units sequentially includes: a macro-branch, a first 1×1 convolutional layer, a PReLU activation function, a third normalization unit, and a micro-branch. The inputs and outputs of the macro-branch, the first 1×1 convolutional layer, the PReLU activation function, the third normalization unit, and the micro-branch are connected sequentially. The input of the macro-branch serves as the input of the dual-scale spatiotemporal joint modeling unit, and the output of the micro-branch serves as the output of the dual-scale spatiotemporal joint modeling unit. The macroscopic branch includes: a temporal shift unit, a first feedforward convolution unit, a second feedforward convolution unit, a third feedforward convolution unit, a positional encoding scaling and translation operation unit, a local-global hybrid attention unit, a second multiplication unit, a third multiplication unit, a fourth multiplication unit, a sigmoid activation function, a fourth feedforward convolution unit, and a second concatenation block. The output of the temporal shift unit is connected to the inputs of the first, second, and third feedforward convolution units and the second concatenation block, respectively. The input of the temporal shift unit is the input of the macroscopic branch. The output of the first feedforward convolution unit is connected to the input of the positional encoding scaling and translation operation unit, and the output of the second feedforward convolution unit is simultaneously connected to the input of the macroscopic branch. The input of the position encoding scaling and translation operation unit is connected to the input of the second multiplication unit. The output of the third feedforward convolution unit is connected to both the input of the position encoding scaling and translation operation unit and the input of the third multiplication unit. The output of the position encoding scaling and translation operation unit is connected to the input of the second multiplication unit and the third multiplication unit respectively. The output of the second multiplication unit is connected to the input of the fourth multiplication unit after passing through the Sigmoid activation function. The output of the third multiplication unit is connected to the input of the fourth multiplication unit. The output of the fourth multiplication unit is connected to the input of the fourth feedforward convolution unit. The output of the fourth feedforward convolution unit is connected to the input of the second concatenation block. The output of the second concatenation block serves as the output of the macroscopic branch. The micro-branch includes: the fifth feedforward convolutional unit, the sixth feedforward convolutional unit, the first linear layer, the ReLU activation function, the second linear layer, and 2... L-1 The system consists of a series of deep-hole 2D convolutional blocks, a third concatenation block, a fifth multiplication unit, a fourth concatenation block, a fourth normalization unit, and a second 1×1 convolutional layer. The input of the fifth feedforward convolutional unit is connected to the input of the sixth feedforward convolutional unit and the input of the fourth concatenation block. The input of the fifth feedforward convolutional unit serves as the input to the micro-branch. The output of the fifth feedforward convolutional unit is connected to the input of the fifth multiplication unit. The output of the sixth feedforward convolutional unit is simultaneously connected to the input of the third concatenation block and the input of the first linear layer. The system also includes the first linear layer, the ReLU activation function, the second linear layer, and a second 1×1 convolutional layer. L-1 The inputs and outputs of a series of deeply dilated 2D convolutional blocks are sequentially connected, and there is a residual connection between adjacent deeply dilated 2D convolutional blocks, as well as between the first deeply dilated 2D convolutional block and the second... L-1 Residual connections between two-dimensional convolutional blocks with deep holes, the second L-1 The output of a deep-hole 2D convolutional block is connected to the input of the third splicing block. The output of the third splicing block is connected to the input of the fifth multiplication unit. The output of the fifth multiplication unit is connected to the input of the fourth splicing block. The output of the fourth splicing block, the input and output of the fourth normalization unit, and the second 1×1 convolutional layer are connected sequentially. L is a preset positive integer.

8. The radar anti-jamming method based on dual-scale phase sensing according to claim 7, characterized in that, Each feedforward convolutional unit sequentially includes: a scaling and normalization unit, a linear layer, a SiLU activation function, a one-dimensional depthwise convolutional unit, a fifth concatenation block, and a regularization layer. The inputs and outputs of the scaling and normalization unit, the third linear layer, the SiLU activation function, the one-dimensional depthwise convolutional unit, the fifth concatenation block, and the regularization layer are connected sequentially. The input of the scaling and normalization unit serves as the input of the feedforward convolutional unit, and the output of the regularization layer serves as the output of the feedforward convolutional unit. Each deep-hole 2D convolutional block sequentially includes: a constant padding block, a 2D dilated convolutional block, an instance normalization layer, a PReLU activation function, and a sixth concatenation block. The inputs and outputs of the constant padding block, the 2D dilated convolutional block, the instance normalization layer, the PReLU activation function, and the sixth concatenation block are connected sequentially. The input of the constant padding block serves as the input of the deep-hole 2D convolutional block, and the output of the sixth concatenation block serves as the output of the deep-hole 2D convolutional block.

9. A radar anti-jamming device based on dual-scale phase sensing, characterized in that, It includes a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus; The memory is used to store computer programs; When the processor executes the computer program stored in the memory, it implements the steps of the radar anti-jamming method based on dual-scale phase sensing as described in any one of claims 1-8.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, which, when executed by a processor, implements the steps of the radar anti-jamming method based on dual-scale phase sensing as described in any one of claims 1-8.

Images