A UAV target recognition method based on ISAC signal decomposition and background contrast attention fusion

Abstract

This invention relates to a UAV target recognition method based on ISAC signal decomposition and background contrast attention fusion, belonging to the field of radar signal processing. The method includes: configuring the transmitted signal of an ISAC base station based on a 5G NR frame structure; receiving and establishing a CSI mixed signal model; constructing a parameterized operator; using the rmD-NSP algorithm to iteratively decompose the mixed signal, separating the rotor micro-Doppler component, and converting the separated signal into a micro-Doppler time-frequency spectrum; using the RMBL module to model multi-scale background features of the time-frequency spectrum; using the BCA module to calculate the difference weights between target features and background features, generating a background-suppressed micro-motion feature vector; generating an HRRP sequence using the CSI mixed signal; using an LSTM network to extract the target's geometric structure feature vector; and weighted concatenating and fusing the micro-motion feature vector and the geometric structure feature vector and inputting them into a classifier to output the UAV target's category probability.

Description

Technical Field

[0001] This invention belongs to the field of radar signal processing technology and relates to a method for UAV target recognition based on ISAC signal decomposition and background contrast attention fusion. Background Technology

[0002] With the rapid development of drone technology, its application in aerial photography, logistics, agriculture, and other fields is becoming increasingly widespread. However, the safety hazards posed by "unauthorized" drone flights are also becoming increasingly prominent, posing a serious challenge to low-altitude security systems. Drones typically have the characteristics of low flight altitude, slow speed, and small radar cross-section (RCS) (i.e., "low, slow, and small"), making them difficult to effectively identify using traditional detection methods.

[0003] Existing drone detection and identification technologies mainly include radio monitoring, photoelectric detection, audio detection, and radar detection. Among them, radar detection has become the mainstream method due to its all-weather and all-time operation advantages. However, in urban or complex field environments, radar echoes are often mixed with strong ground clutter, building reflections, and interfering targets such as birds.

[0004] Current radar-based drone identification methods mainly face the following technical challenges:

[0005] On the one hand, the rotor micro-Doppler signal of a UAV is a key feature for identifying its identity, but this signal has extremely weak energy and is easily masked by translational echoes, fuselage vibrations, and strong environmental clutter. Traditional time-frequency analysis methods (such as short-time Fourier transform) struggle to extract clear micro-motion features under strong interference, resulting in blurred features.

[0006] On the other hand, existing deep learning recognition methods often directly input radar maps into convolutional neural networks (CNNs). The networks tend to learn the dominant background textures in the image rather than the subtle target features, resulting in a high false alarm rate (FAR) under low signal-to-noise ratio (Low SCR) conditions, and lack explicit modeling and suppression mechanisms for background distribution characteristics.

[0007] Furthermore, single-dimensional features (such as using only micro-Doppler spectra or only high-resolution range images) are insufficient to fully characterize the physical properties of the target, resulting in insufficient model generalization ability.

[0008] Therefore, there is an urgent need for a fusion recognition method that can effectively separate weak rotor signals from strongly interfering mixed signals and suppress feature-level interference by learning the background distribution, so as to improve the recognition accuracy of UAV targets in complex scenarios. Summary of the Invention

[0009] In view of this, the purpose of this invention is to address the technical pain points of existing radar target recognition technology in complex urban and field environments, such as weak echoes from "low, slow, and small" UAV targets, strong environmental clutter, large airframe vibration interference, and high false alarm rate. The invention proposes a UAV target recognition method based on ISAC signal decomposition and background contrast attention fusion.

[0010] To achieve the above objectives, the present invention provides the following technical solution: A UAV target recognition method based on ISAC signal decomposition and background contrast attention fusion includes the following steps: S1: Based on the 5G NR frame structure, configure the transmission signal of the integrated sensing ISAC base station, receive and establish a channel state information (CSI) hybrid signal model that includes UAV target echo, fuselage vibration and environmental clutter; S2: Construct a parameterized operator and use the rotor micro-Doppler null space tracking rmD-NSP algorithm to iteratively decompose the mixed signal in S1, separate the rotor micro-Doppler component, and use synchronous extraction transform SET to convert the separated signal into a micro-Doppler time-frequency spectrum. S3: Design a micro-motion feature extraction network based on background contrast attention. This network includes a random mask multi-scale background learner (RMBL) module and a background contrast attention (BCA) module. The RMBL module is used to model the multi-scale background features of the time-frequency map generated in S2, and the BCA module is used to calculate the difference weights between the target features and the background features to generate a micro-motion feature vector after background suppression. S4: Use the signal in S1 to generate a high-resolution range profile HRRP sequence, input it into a long short-term memory network LSTM, and use a gating mechanism to mine the temporal correlation within the HRRP sequence to extract the geometric feature vector of the target. S5: The micro-motion feature vector obtained in S3 and the geometric structure feature vector obtained in S4 are weighted and concatenated and fused, and then input into the fully connected layer classifier. The class probability of the UAV target is output through the Softmax function.

[0011] Furthermore, in step S1, the mixed-signal model is represented as:

[0012] in, This represents a UAV signal vector that includes fuselage translation, fuselage vibration, and rotor rotation. This represents the clutter and noise vector.

[0013] Furthermore, in step S2, the rmD-NSP algorithm is used to extract the rotor micro-Doppler signal. The optimization objective function is:

[0014] in, Let be the real part of the mixed signal. For residual signals, A second-order differential parameterization operator designed for rotor signals. Let be the operator parameter vector, where Operator parameters characterizing the instantaneous frequency of a signal, Operator parameters characterizing the instantaneous bandwidth and amplitude changes of a signal. It is a second-order difference operator. For Lagrange multipliers, Leakage parameters for controlling zero-space retention.

[0015] Furthermore, in step S2, the instantaneous frequency is calculated using the synchronous extraction transform SET. And generate time-frequency spectra The formula is:

[0016]

[0017] in, The result is the short-time Fourier transform (STFT) of the separated rotor signal. For frequency index, For time indexing, For the Dirac function, This represents a discrete approximation of the partial derivative with respect to time. Represents the set of real numbers. Represents the imaginary unit ( ).

[0018] Furthermore, step S3, which involves using the RMBL module to model the multi-scale background features of the time-frequency spectrum generated in step S2, includes: inputting the micro-Doppler time-frequency spectrum obtained in step S2 into the RMBL module; employing a random masking strategy to randomly occlude pixel blocks of the input image at a certain ratio; then using multi-scale convolutional layers to extract background features under different receptive fields; and finally concatenating the multi-scale features to output a background feature vector. .

[0019] Furthermore, step S3, which involves using the BCA module to calculate the difference weights between the target features and the background features to generate a background-suppressed micro-motion feature vector, includes: The BCA module calculates the target feature map. With background feature map The difference is used to generate attention weights, and the specific process is as follows: First, the background features extracted by the RMBL module are expanded to match the target feature dimension, that is, the background feature vector is copied and expanded to obtain... Target features are downsampled to obtain The joint features are obtained by splicing. :

[0020] Subsequently, a target attention weight map is generated using pixel-level max pooling and convolution operations. And act on the original target features. :

[0021]

[0022] in, Indicates feature concatenation operation; Indicates the convolution operation; This represents a pixel-level max pooling operation; Indicates an upsampling operation; This represents the Sigmoid activation function; The reshaped target features; This is the final output micro-motion feature vector after background suppression.

[0023] Furthermore, in step S4, the LSTM network utilizes its gating mechanism to capture long-term dependencies in the HRRP sequence. The specific update formula is as follows:

[0024]

[0025]

[0026]

[0027]

[0028]

[0029] in, For time step HRRP sample amplitude input; and They are time steps and The hidden state; and They are time steps and cellular state; These represent the forget gate, input gate, and output gate, respectively. Candidate cell state; These correspond to the forget gate weight matrix, input gate weight matrix, cell state weight matrix, and output gate weight matrix, respectively. These correspond to the forget gate bias, input gate bias, cell state bias, and output gate bias, respectively. The Sigmoid activation function is used. This represents element-wise multiplication. The final output is the hidden state. This refers to the robust geometric structure feature vector that incorporates temporal attitude change information.

[0030] Furthermore, in step S5, the Jensen-Shannon divergence loss function is introduced during the training phase. The BCA module in constraint S3 is used to maximize the difference between the target feature distribution and the background feature distribution, and the total loss function is... Represented as:

[0031]

[0032] in, Cross-entropy classification loss, For balance coefficient, It is a non-negative constant. , , The relative entropy. By minimizing , making By maximizing the divergence, the target features can be separated from the background features.

[0033] The beneficial effects of this invention are as follows: (1) Signal recovery capability under extremely low signal-to-noise ratio: Unlike traditional methods that directly process noisy signals, this invention utilizes the rmD-NSP algorithm, based on the physical priors of the sparsity and periodicity of rotor signals, to achieve effective decoupling of "fuselage-rotor-clutter" at the signal level. Combined with SET transform, even in environments with a signal-to-noise ratio as low as -10dB, it can still generate clear and energy-concentrated rotor micro-Doppler time-frequency ridges, significantly reducing the difficulty of subsequent identification.

[0034] (2) Feature-level intelligent background suppression: The RMBL and BCA modules are innovatively introduced, and the mask learning mechanism is used to force the network to understand the background distribution. The feature difference between the target and the background is maximized by constraining the JS divergence loss function. This enables the model to automatically focus on weak rotor feature regions and shield the interference caused by strong ground clutter or bird wing flapping, which significantly reduces the false alarm rate (FAR) in complex scenes.

[0035] (3) Multidimensional feature complementarity enhances robustness: The system integrates micro-Doppler features reflecting motion characteristics and HRRP structural features reflecting physical dimensions, and utilizes LSTM to mine temporal information. This multidimensional fusion strategy solves the problem of decreased recognition rate of single features under specific perspectives (such as radial flight causing micro-Doppler to disappear), enabling the system to have extremely high recognition accuracy and generalization ability for UAVs of different models and flight attitudes.

[0036] Other advantages, objectives, and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination, or may be learned from practice of the invention. The objectives and other advantages of the invention can be realized and obtained through the following description. Attached Figure Description

[0037] To make the objectives, technical solutions, and advantages of the present invention clearer, the preferred embodiments of the present invention will be described in detail below with reference to the accompanying drawings, wherein: Figure 1 This is a flowchart of the overall process for a UAV target recognition method based on ISAC signal decomposition and background contrast attention fusion. Figure 2 This is a flowchart illustrating the UAV target recognition method based on ISAC signal decomposition and background contrast attention fusion. Figure 3 Here is a diagram of the RMBL network structure; Figure 4 This is a diagram of the network structure for micro-motion feature extraction based on background contrast attention (BCA). Detailed Implementation

[0038] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. Unless otherwise specified, the following embodiments and features can be combined with each other.

[0039] It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. Therefore, the drawings only show the components related to the present invention and are not drawn according to the actual number, shape and size of the components in the actual implementation. In the actual implementation, the form, quantity and proportion of each component can be arbitrarily changed, and the layout of the components may also be more complex.

[0040] In the following description, numerous details are explored to provide a more thorough explanation of embodiments of the invention. However, it will be apparent to those skilled in the art that embodiments of the invention may be practiced without these specific details. In other embodiments, well-known structures and devices are shown in block diagram form rather than in detail to avoid obscuring embodiments of the invention.

[0041] Example 1: like Figure 1-4 As shown, this invention provides a UAV target recognition method based on ISAC signal decomposition and background contrast attention fusion, which includes the following steps: S1: Construct a complex target echo signal model in an integrated sensory communication (ISAC) environment.

[0042] Specifically, the transmission signal of the ISAC base station is configured based on the frame structure of the 5G NR (New Radio) standard. In the downlink time slot, the Physical Downlink Shared Channel (PDSCH) is used as a sensing and detection signal to ensure a sufficient pulse repetition frequency (PRF) to capture the characteristics of the high-speed rotating rotor.

[0043] After the received baseband signal undergoes quadrature demodulation and cyclic prefix removal (CP) processing, a hybrid signal model containing channel state information (CSI) with multiple physical components is constructed. This model not only considers the target's translational motion, but also provides a refined model of the rotor's dynamic radar cross-section (RCS) and fuselage micro-motions:

[0044] in: The phase change of the fuselage translational echo is determined by the Doppler frequency shift. Decision, among which Radial velocity, Wavelength; The interference signal generated by fuselage vibration is modeled as a non-stationary phase-modulated signal with a micro-Doppler frequency. With vibration frequency and amplitude Related, usually manifested as low-frequency sinusoidal modulation; The rotor micro-Doppler signal to be extracted is derived from the helicopter rotor model. Each rotor, every rotor Composed of blades, its echo not only includes the rotational speed The phase modulation also includes dynamic RCS amplitude modulation determined by the blade material and angle. The signal exhibits periodic flickering in the time domain and a broadband frequency-modulated signal in the frequency domain. This includes additive white Gaussian noise and environmental clutter (such as ground clutter and multipath effects).

[0045] S2: Signal separation and enhancement based on rotor micro-Doppler null space tracking (rmD-NSP) and synchronous extraction transform (SET).

[0046] This step aims to extract from mixed signals The rotor signal was "purified" in the middle. .

[0047] First, a second-order differential parameterization operator for the characteristics of rotor signal waveforms is constructed. This operator utilizes the physical characteristics of the instantaneous frequency and bandwidth of the rotor signal to design a system where the pure rotor signal falls into the operator's null space. .

[0048] The following optimization problem is constructed to solve for the rotor signal components. (i.e., estimated) ):

[0049] objective function Specifically, it can be elaborated as follows:

[0050] in: Let be the real part vector of the mixed signal; The residual signal to be optimized (i.e., the background and fuselage signals after removing the rotor signal); That is, the estimated target rotor signal ; Let be the operator parameter vector, where Operator parameters characterizing the instantaneous frequency of a signal; It is a second-order difference matrix used to constrain the smoothness of operator parameters; For Lagrange multipliers, This is a leakage parameter used to control the aggressiveness of signal separation.

[0051] The above non-convex optimization problem was solved using the Alternating Direction Method, ultimately yielding a pure rotor micro-Doppler signal. .

[0052] Subsequently, in order to address the issues of energy divergence and low resolution in the time-frequency graph of traditional Short Time Fourier Transform (STFT), Synchronous Extraction Transform (SET) was employed to separate the signal. The SET operator processes the data by retaining energy near the instantaneous frequency trajectory and eliminating fuzzy energy. Its mathematical expression is as follows:

[0053] in For STFT results, To estimate the instantaneous frequency using the phase derivative, The Dirac function is used. The final result is a micro-Doppler time-frequency spectrum (Spectrogram) with highly concentrated energy and clear feature texture.

[0054] S3: Construct a micro-motion feature extraction network based on background perception and contrastive attention.

[0055] To address the potential interference from random noise or non-cooperative targets such as birds that may remain in time-frequency maps under low signal-to-noise ratio environments, a deep neural network is designed for secondary feature enhancement. This network comprises two parallel branches: Random Mask Multi-Scale Background Learner (RMBL) Branch: This module is specifically designed to "learn what the background looks like." The input is a time-frequency map generated by S2. First, a random masking strategy is employed to randomly mask pixel blocks of the input image at a certain percentage (e.g., 50%-75%). This operation forces the network to rely on contextual information rather than local target features, and instead reconstruct and understand the global background distribution.

[0056] Next, multi-scale convolutional layers (with kernel sizes of [missing information]) are used. Background features are extracted from different receptive fields. Small convolutional kernels capture high-frequency noise and texture, while large convolutional kernels capture low-frequency environmental distribution. Finally, the multi-scale features are concatenated to output a background feature vector. .

[0057] Background Contrast Attention (BCA) module branch: This module utilizes the background knowledge learned by RMBL to cleanse target features. The input is the original, unmasked time-frequency map features. and background features .

[0058] First, through Convolution incorporates background features Extend to target features The two features are concatenated along the same dimension and then concatenated along the channel dimension to obtain joint features. .

[0059] Then, attention weight maps are generated using pixel-wise max pooling and convolution operations. Pixel-wise max pooling extracts the spatially strongest response regions (usually the target region), as shown in the following formula:

[0060] By using the Sigmoid activation function Mapped to Weighted graph of intervals, and compared with original target features Perform element-wise product:

[0061] This yields a micro-motion feature vector that suppresses background noise and enhances rotor micro-motion texture. .

[0062] S4: Geometric structure temporal feature extraction based on Long Short-Term Memory (LSTM) network.

[0063] In addition to micro-motion characteristics, the size and shape of the drone are also important identification criteria.

[0064] Using the original echo signal in S1, a high-resolution range profile (HRRP) is generated through FFT transformation along the fast time dimension. The HRRP sequence reflects the projection distribution of the target scattering center along the radar line-of-sight direction.

[0065] Considering that changes in the drone's attitude angles (Pitch / Roll / Yaw) during flight cause fluctuations in HRRP, the amount of HRRP information in a single frame is limited. Therefore, a continuous time-step HRRP sequence is used. Input LSTM network.

[0066] LSTM networks utilize their gating mechanism to capture long-term dependencies in HRRP sequences:

[0067]

[0068]

[0069]

[0070]

[0071]

[0072] in, For time step HRRP sample amplitude input; and They are time steps and The hidden state; and They are time steps and cellular state; These represent the forget gate, input gate, and output gate, respectively. Candidate cell state; These correspond to the forget gate weight matrix, input gate weight matrix, cell state weight matrix, and output gate weight matrix, respectively. These correspond to the forget gate bias, input gate bias, cell state bias, and output gate bias, respectively. The Sigmoid activation function is used. Represents element-wise multiplication; the final output is the hidden state. This refers to the robust geometric structure feature vector that incorporates temporal attitude change information.

[0073] S5: Multidimensional feature cascade, JS divergence constraint and classification recognition.

[0074] Background suppression micro-motion features output by S3 Geometric features of S4 output Weighted concatenation is performed to form a multidimensional joint feature vector. This vector is then input into a fully connected layer (MLP) for dimensionality reduction and mapping, and finally outputs the class probability of the UAV target (e.g., quadcopter, hexacopter, helicopter, bird, etc.) through a Softmax layer.

[0075] To ensure that the BCA module in S3 can truly distinguish between the "target" and the "background" rather than simply learning ordinary texture features, this invention introduces Jensen-Shannon (JS) divergence constraints into the loss function during the training phase.

[0076] target features and background features By normalizing using Softmax, we can treat them as two probability distributions. and JS divergence is used to measure the difference between two distributions:

[0077] The total loss function is defined as:

[0078]

[0079] in Used to ensure classification accuracy. Note the negative sign before the JS divergence term, which means that the training objective is to maximize the difference between the target features and the background features (i.e., the larger the JS divergence, the better), thereby forcing the network to completely separate the features containing rotor micro-motion from the pure background clutter features.

[0080] Example 2: An electronic device, comprising a memory and a processor; The memory is used to store computer programs; The processor is configured to implement the method described in Embodiment 1 when executing the computer program.

[0081] Example 3: A computer-readable storage medium storing a computer program that, when executed by a processor, implements the method described in Embodiment 1.

[0082] Example 4: A computer program product includes a computer program that, when executed by a processor, implements the method described in Example 1.

[0083] In the above embodiments, the reference to "this embodiment" in the specification indicates that a specific feature, structure, or characteristic described in connection with the embodiment is included in at least some embodiments, but not necessarily all embodiments. Multiple appearances of "this embodiment" do not necessarily refer to the same embodiment.

[0084] In the above embodiments, although the invention has been described in conjunction with specific embodiments thereof, many substitutions, modifications, and variations of these embodiments will be apparent to those skilled in the art from the foregoing description. For example, other memory structures (e.g., dynamic RAM (DRAM)) may be used with the embodiments discussed. The embodiments of the invention are intended to cover all such substitutions, modifications, and variations falling within the broad scope of the appended claims.

[0085] As will be understood by those skilled in the art, the computer-readable storage medium described in this embodiment allows for the implementation of all or part of the steps in the above method embodiments by computer program-related hardware. The aforementioned computer program can be stored in a computer-readable storage medium. When executed, the program performs the steps of the above method embodiments; and the aforementioned storage medium includes various media capable of storing program code, such as ROM, RAM, magnetic disks, or optical disks.

[0086] The electronic terminal provided in this embodiment includes a processor, a memory, a transceiver, and a communication interface. The memory and the communication interface are connected to the processor and the transceiver and complete communication between them. The memory is used to store computer programs, the communication interface is used to perform communication, and the processor and the transceiver are used to run the computer programs, so that the electronic terminal performs the various steps of the above method.

[0087] In this embodiment, the memory may include random access memory (RAM) and may also include non-volatile memory, such as at least one disk storage device.

[0088] 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.

[0089] This invention can be used in a wide range of general-purpose or special-purpose computing system environments or configurations. Examples include: personal computers, server computers, handheld or portable devices, tablet devices, multiprocessor systems, microprocessor-based systems, set-top boxes, programmable consumer electronics, network PCs, minicomputers, mainframe computers, and distributed computing environments including any of the above systems or devices, etc.

[0090] This invention can be described in the general context of computer-executable instructions, such as program modules, that are executed by a computer. Generally, program modules include routines, programs, objects, components, data structures, etc., that perform a specific task or implement a specific abstract data type. This invention can also be practiced in distributed computing environments where tasks are performed by remote processing devices connected via a communication network. In distributed computing environments, program modules can reside in local and remote computer storage media, including storage devices.

[0091] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for UAV target recognition based on ISAC signal decomposition and background contrast attention fusion, characterized in that: Includes the following steps: S1: Based on the 5G NR frame structure, configure the transmission signal of the integrated sensing ISAC base station, receive and establish a channel state information (CSI) hybrid signal model that includes UAV target echo, fuselage vibration and environmental clutter; S2: Construct a parameterized operator and use the rotor micro-Doppler null space tracking rmD-NSP algorithm to iteratively decompose the mixed signal in S1, separate the rotor micro-Doppler component, and use synchronous extraction transform SET to convert the separated signal into a micro-Doppler time-frequency spectrum. S3: Design a micro-motion feature extraction network based on background contrast attention. This network includes a random mask multi-scale background learner (RMBL) module and a background contrast attention (BCA) module. The RMBL module is used to model the multi-scale background features of the time-frequency spectrum generated by S2, and the BCA module is used to calculate the difference weights between the target features and the background features to generate the micro-motion feature vector after background suppression. S4: Use the signal in S1 to generate a high-resolution range profile HRRP sequence, input it into a long short-term memory network LSTM, and use a gating mechanism to mine the temporal correlation within the HRRP sequence to extract the geometric feature vector of the target. S5: The micro-motion feature vector obtained in S3 and the geometric structure feature vector obtained in S4 are weighted and concatenated and fused, and then input into the fully connected layer classifier. The class probability of the UAV target is output through the Softmax function.

2. The UAV target recognition method based on ISAC signal decomposition and background contrast attention fusion according to claim 1, characterized in that: In step S1, the mixed signal model is represented as: in, This represents a UAV signal vector that includes fuselage translation, fuselage vibration, and rotor rotation. This represents the clutter and noise vector.

3. The UAV target recognition method based on ISAC signal decomposition and background contrast attention fusion according to claim 1, characterized in that: In step S2, the rmD-NSP algorithm is used to extract the rotor micro-Doppler signal. The optimization objective function is: in, Let be the real part of the mixed signal. For residual signals, A second-order differential parameterization operator designed for rotor signals. Let be the operator parameter vector, where Operator parameters characterizing the instantaneous frequency of a signal, Operator parameters characterizing the instantaneous bandwidth and amplitude changes of a signal. It is a second-order difference operator. For Lagrange multipliers, Leakage parameters for controlling zero-space retention.

4. The UAV target recognition method based on ISAC signal decomposition and background contrast attention fusion according to claim 3, characterized in that: In step S2, the instantaneous frequency is calculated using the synchronous extraction transform SET. And generate time-frequency spectra The formula is: in, The result is the short-time Fourier transform (STFT) of the separated rotor signal. For frequency index, For time indexing, For the Dirac function, This represents a discrete approximation of the time partial derivative. Represents the set of real numbers. Represents the imaginary unit. .

5. The UAV target recognition method based on ISAC signal decomposition and background contrast attention fusion according to claim 1, characterized in that: Step S3, which involves using the RMBL module to model the multi-scale background features of the time-frequency spectrum generated in S2, includes: inputting the micro-Doppler time-frequency spectrum obtained in step S2 into the RMBL module; employing a random masking strategy to randomly occlude pixel blocks of the input image at a certain ratio; then using multi-scale convolutional layers to extract background features under different receptive fields; and finally concatenating the multi-scale features to output a background feature vector. .

6. The UAV target recognition method based on ISAC signal decomposition and background contrast attention fusion according to claim 5, characterized in that: Step S3, which involves using the BCA module to calculate the difference weights between the target features and the background features to generate a background-suppressed micro-motion feature vector, includes: The BCA module calculates the target feature map. With background feature map The difference is used to generate attention weights, and the specific process is as follows: First, the background features extracted by the RMBL module are expanded to match the target feature dimensions, resulting in a background feature vector that is then copied and expanded. The target features are downsampled to obtain The joint features are obtained by splicing. : Subsequently, pixel-level max pooling and convolution operations are used to generate a target attention weight map. And act on the original target features. : in, Indicates feature concatenation operation; Indicates the convolution operation; This represents a pixel-level max pooling operation; Indicates an upsampling operation; This represents the Sigmoid activation function; The reshaped target features; This is the final output micro-motion feature vector after background suppression.

7. The UAV target recognition method based on ISAC signal decomposition and background contrast attention fusion according to claim 1, characterized in that: In step S4, the LSTM network uses its gating mechanism to capture long-term dependencies in the HRRP sequence. The specific update formula is as follows: in, For time step HRRP sample amplitude input; and They are time steps and The hidden state; and They are time steps and cellular state; These represent the forget gate, input gate, and output gate, respectively. Candidate cell state; These correspond to the forget gate weight matrix, input gate weight matrix, cell state weight matrix, and output gate weight matrix, respectively. These correspond to the forget gate bias, input gate bias, cell state bias, and output gate bias, respectively. The Sigmoid activation function is used. Represents element-wise multiplication; the final output is the hidden state. This refers to the robust geometric structure feature vector that incorporates temporal attitude change information.

8. The UAV target recognition method based on ISAC signal decomposition and background contrast attention fusion according to claim 1, characterized in that: In step S5, the Jensen-Shannon divergence loss function is introduced during the training phase. The BCA module in constraint S3 is used to maximize the difference between the target feature distribution and the background feature distribution, and the total loss function is... Represented as: in, Cross-entropy classification loss, For balance coefficient, It is a non-negative constant. , , For relative entropy; by minimizing , making By maximizing the divergence, the target features can be separated from the background features.

Images