Anti-deception interference method and system based on dynamic weight and timing constraint

Abstract

The application belongs to the technical field of signal processing, and specifically discloses an anti-deception interference method and system based on dynamic weight and timing constraints, wherein the method comprises the following steps: using a field programmable logic gate array to perform hardware-level real-time preprocessing on synthetic aperture radar original echo data, and triggering hardware-level adaptive routing truncation according to an amplitude distortion gradient, and outputting preprocessed data after algorithm power deployment; based on the preprocessed data, multi-modal features are extracted in parallel, and the multi-modal features comprise image domain incoherent topology features, local phase dispersion degree features and spatial geometric constraint features; the application aims to solve the problems of single feature source representation, poor environmental adaptability of fusion mechanism, not using deception interference timing physical contradictions and large model calculation overhead in existing synthetic aperture radar anti-interference schemes, which leads to limited real-time performance.

Description

Technical Field

[0001] This invention relates to the technical field of signal processing, specifically to a method and system for resisting deception interference based on dynamic weights and timing constraints. Background Technology

[0002] In the field of synthetic aperture radar (SAR) signal processing and electronic countermeasures, existing anti-jamming schemes face limitations in physical mechanisms and computational efficiency in complex electromagnetic environments in order to counter deception jamming based on digital radio frequency memory (DRFM). Specifically, DRFM jammers can intercept and modulate relayed radar signals to create false targets at the imaging end that are highly similar to the scattering characteristics and spatial distribution of real targets, leading to incorrect situational awareness in the radar system. Currently, most mainstream detection architectures rely on convolutional neural networks to extract amplitude domain features from complex SAR images. However, when facing intelligent and covert deception jamming, a single information source and a static fusion mechanism cannot adapt to the changing environment. The dynamic adjustment of feature weights based on changes in ambient noise and interference intensity can lead to information contamination and reduced detection accuracy due to low-quality feature components during fusion. Furthermore, existing anti-interference discrimination logic is generally based on a single-frame snapshot mode, failing to consider the processing delay introduced by the DRFM jammer in the signal processing flow and the residual error in estimating the motion parameters of the radar carrier. This results in the system being unable to detect the discontinuities in the motion trajectory of false targets and the temporal incoherence such as deviations from Doppler physical laws in continuous observation time series. In addition, on spaceborne or airborne embedded platforms, there is a contradiction between limited computing resources and the large number of parameters and complex global dependency modeling requirements of traditional models, which limits the real-time performance of anti-interference processing. Summary of the Invention

[0003] This invention provides a method and system for resisting deception and interference based on dynamic weights and temporal constraints. The aim is to solve the problems of single feature source representation, poor environmental adaptability of fusion mechanism, failure to utilize the temporal and physical contradictions of deception and interference, and high model computational overhead leading to limited real-time performance in existing synthetic aperture radar anti-jamming schemes.

[0004] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows: A method and system for resisting deception and interference based on dynamic weights and temporal constraints includes: performing hardware-level real-time preprocessing on raw synthetic aperture radar echo data using a field-programmable gate array (FPGA), and triggering hardware-level adaptive routing truncation based on amplitude distortion gradients to output preprocessed data with allocated computing power; extracting multimodal features in parallel based on the preprocessed data, including image domain incoherent topological features, local phase dispersion features, and spatial geometric constraint features; quantifying the multidimensional vector space manifold curvature abrupt change factor of the multimodal features and generating dynamic fusion weights accordingly; performing cross-modal physical evidence chain interaction fusion on the multimodal features to output enhanced feature vectors; performing dual temporal physical verification of the enhanced feature vectors using kinematics and Doppler history to obtain a temporal consistency evaluation factor; inputting the enhanced feature vectors and the temporal consistency evaluation factor into a decision-level recursive Bayesian fusion module, accumulating evidence across multiple frames using Markov transition constraints, and outputting the interference posterior probability; identifying deception and interference targets based on the interference posterior probability, and performing signal suppression on the identified interference regions.

[0005] In one aspect of the invention, the triggering of hardware-level adaptive routing truncation includes: the field-programmable gate array (FPGA) calculates the amplitude distortion gradient of the baseband signal in the fast time dimension in real time; when the amplitude distortion gradient exceeds the upper limit of the threshold set by the ground clutter statistical model, the adaptive logic dimensionality reduction truncation of the system bus is triggered, the graphics processor stops executing the high-energy-consuming cross-resolution unit sampling operation, and forces the output of a minimal constant background vector consistent with the preset dimension to the subsequent steps of cross-modal physical evidence chain interaction fusion; at the same time, the backbone bus bandwidth is released to the digital signal processor, which exclusively uses its resources to perform the extraction of the local phase dispersion features.

[0006] In one aspect of the invention, the extraction of the local phase dispersion features includes: extracting the instantaneous unwrapped phase sequence of the baseband complex signal based on the phase truncation physical mechanism of the digital-to-analog converter inside the digital radio frequency memory; locating the hardware-level phase jump by calculating the second-order differential phase of adjacent sampling points; and calculating the dispersion statistics within a short-time sliding window to identify the microscopic phase distortion introduced by the digital radio frequency memory.

[0007] In one aspect of the present invention, the generation of dynamic fusion weights includes: extracting the manifold curvature abrupt change factor of the multimodal features in the feature space, characterizing the smoothness of the feature distribution by calculating the second-order difference norm between adjacent feature dimensions, and converting the manifold curvature abrupt change factor into the corresponding modal reliability score using a feedforward analytical mapping network.

[0008] In one aspect of the invention, the extraction of incoherent topological features in the image domain includes: constructing a cross-resolution cell sampling matrix for spatial topological mismatch of discrete scattering centers, wherein the sampling step size of the cross-resolution cell sampling matrix is ​​in a stepwise multiple relationship with the spatial resolution of the synthetic aperture radar system, and is used to extract the phase discontinuity topological map of false interference targets in the cross-resolution cell.

[0009] In one aspect of the invention, the cross-modal physical evidence chain interaction fusion includes: generating a query matrix using the incoherent topological features of the image domain to lock the suspected target location, generating a key matrix and a value matrix using the local phase dispersion features, and matching phase dispersion evidence at the suspected target location by scaling dot product operators to achieve a nonlinear correlation between the distribution of physical defects and spatial location.

[0010] In one aspect of the invention, the dual temporal physical verification includes: estimating the continuous motion trajectory of the target based on Kalman filtering iteration, calculating the statistical Mahalanobis distance between the observed position and the physical prediction value as the positional abrupt change factor, and using the imaging geometric model to verify the residual value between the observed Doppler frequency and the theoretical Doppler frequency, so as to determine whether the target signal conforms to the kinematic laws of the ground object.

[0011] In one aspect of the invention, the multi-frame evidence accumulation includes: using recursive Bayesian logic to accumulate the observation features of consecutive imaging frames probabilities, and introducing a Markov transition constraint smoothing operator during the probability transition process, thereby filtering out false posterior probability spikes introduced by flicker interference by limiting the physical maximum step size threshold of state transitions between adjacent frames.

[0012] In one aspect of the invention, the execution signal suppression includes: employing a reconstruction algorithm based on compressed sensing to construct a mask matrix in the complex image domain for regions determined to be deception interference, and using a sparse transform dictionary composed of discrete wavelet bases and an orthogonal matching pursuit algorithm to recover the masked true background based on the statistical characteristics of surrounding uncontaminated pixels.

[0013] In another aspect, the present invention also relates to an anti-spoofing interference system based on dynamic weights and timing constraints, used to implement the above-mentioned anti-spoofing interference method based on dynamic weights and timing constraints, comprising: A heterogeneous computing baseband processing board that integrates a field-programmable gate array, a digital signal processor, and a graphics processor; The field-programmable gate array is used to execute the amplitude distortion monitoring and computing power routing truncation logic of the synthetic aperture radar raw echo data; The graphics processor is used to perform matrix extraction of incoherent topological features in the image domain; The digital signal processor is used to perform local phase divergence features, manifold curvature evaluation, and decision-level recursive Bayesian fusion operations.

[0014] Compared with the prior art, the present invention has the following beneficial effects: First, this invention achieves multi-dimensional feature coverage of synthetic aperture radar echo and imaging data by constructing three heterogeneous feature extraction branches: image features, temporal features, and auxiliary information. The complementarity of multi-modal features enhances the system's ability to capture interference modes, compensating for the shortcomings of single-source representation. Second, the proposed scheme introduces a dynamic weight adaptive module that adjusts the fusion ratio of each modality based on the real-time signal-to-interference ratio (SIR) and feature completeness of the input data. This automatically suppresses feature branches distorted by complex clutter or malicious interference, ensuring that the fusion result is dominated by high-quality feature components and improving the system's detection stability in complex environments. Third, this scheme introduces a physical mechanism-based temporal consistency constraint into the anti-interference architecture. By utilizing the jammer's processing delay and motion parameter estimation bias, the target is verified from two dimensions: motion continuity and Doppler physical laws. This reduces the judgment error caused by high-fidelity deception interference, maintaining recognition accuracy even under low SIR conditions. Finally, the system employs multiple model compression strategies, including depthwise separable convolution, channel pruning, and quantized perceptual training, to keep the number of parameters and computational load of the detection network at a low level. This enables the system to be deployed on resource-constrained spaceborne or airborne embedded computing platforms, meeting the real-time online processing requirements against deception interference and optimizing the system's engineering application value. Attached Figure Description

[0015] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained from these drawings without creative effort.

[0016] Figure 1 This is a flowchart of an anti-spoofing interference method based on dynamic weights and temporal constraints according to the present invention.

[0017] Figure 2 This is a flowchart of step-by-step steps in S1 of the anti-spoofing interference method based on dynamic weights and timing constraints of the present invention.

[0018] Figure 3 This is a flowchart of step-by-step steps in S2 of the anti-spoofing interference method based on dynamic weights and timing constraints of the present invention.

[0019] Figure 4This is a flowchart of step-by-step steps in S3 of the anti-spoofing interference method based on dynamic weights and timing constraints of the present invention.

[0020] Figure 5 This is a flowchart of step S4 in the anti-spoofing interference method based on dynamic weights and timing constraints of the present invention.

[0021] Figure 6 This is a flowchart of step S5 in the anti-spoofing interference method based on dynamic weights and timing constraints of the present invention.

[0022] Figure 7 This is a flowchart of step S6 in the anti-spoofing interference method based on dynamic weights and timing constraints of the present invention.

[0023] Figure 8 This is a schematic diagram of the composition of an anti-spoofing interference system based on dynamic weights and timing constraints according to the present invention. Detailed Implementation

[0024] The present invention will be further described below with reference to embodiments. These embodiments are merely some, not all, of the embodiments of the present invention. Other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are all within the protection scope of the present invention.

[0025] In the specific implementation of the anti-spoofing jamming method and system based on dynamic weights and timing constraints provided by this invention, it is first necessary to engineering define the physical working environment of synthetic aperture radar, the computing power allocation boundary of hardware platform, and the underlying electromagnetic and digital quantization mechanism of digital radio frequency memory (DRFM) jammer.

[0026] The synthetic aperture radar system selected in this embodiment operates in the X-band, with a center frequency set at 9.6 GHz, a transmit signal bandwidth of 500 MHz, and a pulse repetition frequency configured at 2000 Hz. Within this hardware parameter framework, the system achieves sub-meter level range and azimuth resolution. Furthermore, the system is equipped with a heterogeneous computing baseband processing board containing a field-programmable gate array (FPGA), a digital signal processor (DSP), and a graphics processing unit (GPU), laying the data and computing power foundation for subsequent hardware-level adaptive routing truncation and fine physical feature extraction.

[0027] Please see Figure 1 As shown, this embodiment discloses an anti-spoofing interference method based on dynamic weights and timing constraints. This embodiment provides the top-level execution architecture of the method, which achieves high-precision identification against interference by exploiting the inherent flaws in the physical hardware quantization and real-time computing power allocation of the jamming machine. The method includes: S1. The raw echo data of synthetic aperture radar is preprocessed in real time at the hardware level using a field-programmable gate array, and hardware-level adaptive routing is triggered based on the amplitude distortion gradient to output preprocessed data with computing power allocation.

[0028] Specifically, it includes: S11. The high-frequency analog echo is captured by the receiving front end, low-noise amplification and down-conversion are performed, and the acquired synthetic aperture radar raw echo data is sent to the independent buffer register of the field programmable gate array.

[0029] S12. Inside the field-programmable gate array, the amplitude distortion gradient of the baseband signal in the fast time dimension is calculated in real time, and the amplitude distortion gradient is logically compared with the upper limit of the threshold set by the ground clutter statistical model.

[0030] S13. When the amplitude distortion gradient exceeds the upper limit of the threshold, hardware-level adaptive routing is triggered to physically block the direct memory access channel for transmitting data to the graphics processor, and the digital signal processor exclusively uses the system bus to output preprocessed data with computing power allocation.

[0031] S2. Based on the preprocessed data, multimodal features are extracted in parallel. The multimodal features include image domain incoherent topological features, local phase dispersion features, and spatial geometric constraint features.

[0032] Specifically, it includes: S21. Construct a cross-resolution cell sampling matrix for topological mismatch in the discrete scattering center space, and extract the phase discontinuity topological spectrum of the false interference target in the cross-resolution cell as an incoherent topological feature in the image domain.

[0033] S22. Extract the instantaneous unwrapped phase sequence of the baseband complex signal, calculate the second-order differential positioning hardware-level phase jump of the phase of adjacent sampling points, and calculate the dispersion statistics within a short-time sliding window to generate local phase dispersion features.

[0034] S23. The instantaneous orbital altitude, beam pointing and pulse repetition frequency of the synthetic aperture radar platform are encoded to construct spatial geometric constraint characteristics.

[0035] S3. Quantify the multidimensional vector space manifold curvature mutation factor of the multimodal features, and generate dynamic fusion weights accordingly. Perform cross-modal physical evidence chain interaction fusion on the multimodal features and output an enhanced feature vector.

[0036] Specifically, it includes: S31. Calculate the second-order difference norm between adjacent feature dimensions, and extract the manifold curvature abrupt change factor of each multimodal feature in the feature space to characterize the smoothness of the feature distribution.

[0037] S32. A feedforward analytical mapping network is used to convert the manifold curvature mutation factor into the corresponding modal reliability score, and then normalized to generate dynamic fusion weights.

[0038] S33. Generate a query matrix using incoherent topological features in the image domain to locate suspected target locations. Generate a key matrix and a value matrix using local phase dispersion features. Perform cross-modal physical evidence chain interaction fusion by scaling dot product operators and output an enhanced feature vector.

[0039] S4. Perform dual temporal physical verification of the enhanced feature vector using kinematics and Doppler history to obtain the temporal consistency evaluation factor.

[0040] Specifically, it includes: S41. Based on the Kalman filter state transition equation, iteratively estimate the continuous motion trajectory of the target.

[0041] S42. Calculate the statistical Mahalanobis distance between the observation position and the physical prediction value of the current imaging frame, as the position mutation factor.

[0042] S43. Calculate the residual between the observed Doppler frequency and the theoretical Doppler frequency using the imaging geometry model, and generate a time series consistency evaluation factor by combining the positional abrupt change factor.

[0043] S5. Input the enhanced feature vector and the temporal consistency evaluation factor into the decision-level recursive Bayesian fusion module, use Markov transition constraints to accumulate evidence across multiple frames, and output the interference posterior probability.

[0044] Specifically, it includes: S51. In the recursive Bayesian probability transition process, a Markov transition constraint smoothing operator is introduced to limit the physical maximum step size threshold of state transition between adjacent frames.

[0045] S52. Based on the independence assumption, multiply the historical cumulative probability with the likelihood probability of the current frame to output the posterior probability of the target being a deception interference at the current moment.

[0046] S6. Identify the deceptive interference target based on the interference posterior probability, and perform signal suppression on the identified interference area.

[0047] Specifically, it includes: S61. When the posterior probability of interference is higher than the preset decision threshold, a mask matrix is ​​constructed in the complex image domain for the regions determined to be deception interference.

[0048] S62. A reconstruction algorithm based on compressed sensing is adopted, which uses a sparse transform dictionary composed of discrete wavelet basis and an orthogonal matching pursuit algorithm to recover the occluded true background.

[0049] Please see Figure 2As shown in this embodiment, it should be noted that in S1, in order to solve the technical problems of computing power idleness and anti-interference response failure caused by the operating system thread scheduling delay in traditional pure software architecture under extreme electromagnetic interference environment, the present invention constructs a hardware-level real-time preprocessing and routing truncation mechanism based on field-programmable gate array (FPGA) through S1.

[0050] Specifically, in S11, the high-frequency echo is quantized by the analog-to-digital converter (ADC) and then directly fed into the field-programmable gate array in a fully pipelined, non-blocking manner.

[0051] In S12, to accurately capture abrupt changes in the external electromagnetic environment, a differential operator array is hardwired and instantiated within the field-programmable gate array (FPGA) to calculate the amplitude distortion gradient of the baseband signal in the fast time dimension in real time. The formula for calculating the amplitude distortion gradient is as follows:

[0052] in, This represents the amplitude distortion gradient of the baseband signal in the fast time dimension, with units of volts per second (V / s). This represents the instantaneous amplitude envelope value of the radar baseband signal, measured in volts (V). The fast time variable representing the synthetic aperture radar echo, in seconds (s); This represents the mathematical operator for taking the absolute value; It represents the differential operator for taking the first derivative with respect to a fast-time variable.

[0053] This differential operator is implemented within a field-programmable gate array (FPGA) via a three-level high-speed adder tree. The system operates at its main frequency, achieving nanosecond-level environmental degradation monitoring. Subsequently, the system needs to calculate... The upper limit of the threshold set by the statistical model of ground clutter Perform a comparison. This threshold upper limit. The value of is not arbitrarily set, but rather determined by statistically fitting the Rayleigh distribution of a large area of ​​interference-free historical clutter data, and taking its confidence level as . The limit value of the envelope transition rate is determined.

[0054] Among them, the upper limit of the threshold set for the ground clutter statistical model. The method for determining the value is based on a finite number of historical data. First, at least 5000 sets of clean synthetic aperture radar baseband echo sequences under interference-free conditions are extracted. The amplitude distortion gradient set of these sequences in the fast time dimension is calculated. Then, Rayleigh distribution statistical fitting is performed on the gradient set, and the gradient envelope limit value corresponding to a confidence level of 99.9% is selected to calibrate the upper limit of the threshold. The value, for example, is obtained by fitting the Rayleigh distribution parameters from a set of measured background clutter data. The value corresponding to the 99.9th percentile is obtained by inversely solving the cumulative distribution function. Therefore, the scenario is set as follows. This ensures the system's robustness to normal clutter fluctuations.

[0055] In S13, when At the time of its establishment, the physical meaning of the system indicated that the current radar receiving channel was under a mixed attack of high-power broadband suppression or high-density pulse deception jamming. Under such a poor signal-to-interference ratio, traditional image texture feature extraction algorithms not only failed to extract effective information, but also exhausted the overall computing power of heterogeneous computing boards due to massive invalid convolution operations, leading to processing latency collapse.

[0056] To resolve this technical contradiction, the routing control register within the field-programmable gate array undergoes a hardware-level toggle within three clock cycles, triggering a hardware-level adaptive routing truncation. The system directly pulls down the enable pin level of the direct memory access channel for transmitting data to the graphics processor, achieving physical-level communication blocking; simultaneously, it releases all the bandwidth resources of the heterogeneous bus to the digital signal processor, allowing the digital signal processor to exclusively occupy the system bus and focus solely on extracting the local phase dispersion features based on quantization errors.

[0057] Using actual test scenario data for deduction: Assume the system's originally calibrated end-to-end full-link anti-interference inference total latency limit is... Without a route truncation mechanism, when facing a JSR (signal-to-interference ratio) of... The combined interference causes the GPU to forcibly process a complex image matrix filled with wideband Gaussian noise, resulting in a surge in processing time. This leads to a significant decrease in the system's real-time protection capabilities. However, after employing the hardware-level adaptive routing truncation of this invention, at the signal input... At what time, the field-programmable gate array (FPGA) measures Much greater than the upper limit of the threshold A hardware interrupt was immediately triggered, physically truncating the time-consuming GPU computation path and routing the preprocessed data solely to the DSP for one-dimensional sequence analysis. Experimental results show that this truncation significantly reduced the system's response time under extreme conditions. sudden drop This effectively avoids the meaningless consumption of computing resources and endows the synthetic aperture radar system with hard real-time survivability in strong confrontation environments.

[0058] Please see Figure 3 As shown in this embodiment, it should be noted that in S21, in order to solve the defect of the existing technology that uses conventional convolutional neural networks to directly extract complex image features and cannot distinguish between real object coherent scattering and jammer deception modulation from the physical cause, this step innovatively constructs a cross-resolution unit sampling matrix based on the spatial resolution of synthetic aperture radar.

[0059] Specifically, while traditional deception jamming can simulate highly realistic outlines of false targets such as aircraft and vehicles in the amplitude domain, it is essentially a discretely modulated signal relayed from a single source within a digital radio frequency memory (RFMemory). When a real three-dimensional physical target is illuminated by radar electromagnetic waves, complex electromagnetic coupling and multipath interactions exist between strong scattering centers at different locations on its surface, resulting in continuous phase transitions and coherence between adjacent radar resolution cells. However, RFMemory struggles to accurately simulate this microscopic electromagnetic coupling in the spatial dimension, causing false targets to exhibit fragmented, discontinuous phase topology when crossing resolution cells.

[0060] To quantify this physical defect, the formula for extracting the sampling step size across the resolution cell is defined as follows:

[0061] in, This represents the physical sampling space step size on the complex image matrix during feature extraction, in meters (m). Representing discrete multiplier coefficients, let the set be... , is a dimensionless constant; This represents the inherent spatial resolution of the current synthetic aperture radar system (taking the maximum value of the range and azimuth resolutions), in meters (m).

[0062] Based on the aforementioned physical step size, the system does not employ the traditional... Instead of compact convolutional kernels, it constructs a cross-resolution unit sampling matrix that includes dilation features. Through The phase gradient change rate of pixels in the complex image is extracted to capture the spatial topological mismatch of discrete scattering centers.

[0063] This computational logic ensures that the extracted incoherent topological features in the image domain are no longer abstract, high-dimensional vectors devoid of physical meaning, but rather directly reflect the target's position across the traverse. times, times, The ability of electromagnetic wave phase to continuously evolve when the physical size is doubled by radar resolution. The characteristic vector of a real target exhibits strong low-frequency clustering in the vector space, while that of a false target exhibits a cluttered white noise spectral topology.

[0064] In this embodiment, it should be noted that in S22, under extreme interference conditions (such as when hardware routing is triggered by S1), the image domain information is completely paralyzed. At this time, this embodiment uses a digital signal processor to deeply mine the hardware physical limits of the jammer's digital-to-analog converter in the baseband one-dimensional time domain, extracts local phase dispersion features based on the quantization error of the digital radio frequency memory, and solves the problem of identifying weak deceptive signals masked by pure noise.

[0065] Specifically, after intercepting radar signals, the digital radio frequency memory (RF memory) needs to reconstruct the deceptive Doppler phase using a lookup table method and a limited-width analog-to-digital / digital-to-analog converter (ADC). Limited by the hardware's effective quantization bit width (ENOB, typically 8-bit or 10-bit), the jammer generates a continuous phase signal... A step-like phase truncation error is unavoidable. This microscopic physical defect is difficult to detect in the conventional short-time Fourier transform (STFT) energy spectrum, but it will expose high-frequency energy spikes in the second-order derivative domain of the unwound phase.

[0066] To accurately capture the digital traces left by hardware quantization, the instantaneous unwrapped phase sequence of the baseband complex signal is first extracted using a phase unwinding algorithm. Subsequently, in the discrete-time domain, the second-order difference of the phase between adjacent sampling points is calculated. The formula is as follows:

[0067] in, Indicates at discrete sampling points The instantaneous second-order phase difference value at the point, in radians (rad). , , These represent the sampling points respectively. , , The instantaneous untangling phase amplitude at the point, in radians (rad). The sequence number of the discrete-time sampling point is a dimensionless integer.

[0068] To overcome the disturbance of environmental white noise on the single-point second-order difference value, the system further calculates the energy concentration of the second-order difference, i.e., the local phase dispersion statistic, within a short-time sliding window of preset length. The formula is as follows:

[0069] in, This represents the local phase divergence statistics within the sliding window, expressed in radians squared (rad²). This represents the total number of discrete sampling points contained in the short-time sliding window, and is a dimensionless integer. This represents the local summation and integration variable within the sliding window; This represents the integer down operator; This represents a nonlinear amplification operator that takes the absolute value of an internal complex or real variable and then squares it. This represents the discrete summation operator.

[0070] Verification using specific scenario data: In the area of ​​normal ground object echoes without interference, affected by system thermal noise, its second-order phase difference exhibits stable, weak Gaussian white noise. (Using a sliding window...) The local phase dispersion statistics of the real clutter were calculated from each sampling point. When encountering a false target generated by a digital radio frequency memory jammer with an 8-bit quantization width, because its phase quantization step is approximately When a carry is truncated during phase reconstruction, a regular pulse impulse is generated in the second-order difference sequence. Substituting into the same formula, the dispersion statistics of this deceptive interference signal segment surge to... The two differ significantly by two orders of magnitude in numerical magnitude. Specifically, this refers to the total number of discrete sampling points included in the short-time sliding window. The method for determining the values ​​is based on a finite number of historical data sets. First, 1000 known spoofing interference time-domain signal sequences with different quantization bit widths (8-bit to 12-bit) from real digital radio frequency memories are extracted. Then, the sliding window length is traversed in the range of 16 to 256 in powers of 2, and the local phase dispersion statistics are calculated for each sequence. Next, the intersection parameters that minimize the second-order difference variance of the real background clutter and maximize the proportion of interference quantization step pulse energy are extracted for calibration. The optimal size. For example, offline grid search of real data shows that when the window length is less than 32, it is susceptible to random thermal noise, while a length greater than 128 smooths out the instantaneous phase truncation traces of the jammer. When the value is 64, its characteristic discrimination (interference / clutter dispersion ratio) reaches a peak of 153.3, therefore, the value of this scenario is set as follows. =64, thus ensuring the sensitivity and statistical stability of micro-phase distortion capture.

[0071] This computational logic effectively avoids the conventional methods used by existing technologies to identify targets in the image domain. It directly overcomes the underlying hardware bottlenecks of enemy jammers, transforming the inherent quantization errors in the enemy's hardware design into stable decision features for our anti-jamming system, significantly improving the rigidity and robustness of the identification features.

[0072] Please see Figure 4 As shown in this embodiment, it should be noted that in S3, traditional statistical indicators such as image entropy are easily masked by large-area ground clutter and become ineffective, and are difficult to characterize the continuity defects of physical signals. Therefore, this invention innovatively introduces manifold curvature in a high-dimensional feature space to measure the reliability of multimodal features.

[0073] Specifically, in S31, whether it is incoherent topological features in the image domain, local phase dispersion features, or spatial geometric constraint features, the extraction results can all be represented as high-dimensional feature vectors. ( These correspond to three modal branches. (This represents the total dimension of the feature vector).

[0074] After proper feature mapping, a real radar echo should form a smooth, low-dimensional manifold structure across continuous feature dimensions. However, when subjected to high-intensity broadband suppression interference or discrete digital radio frequency memory spoofing interference, this manifold structure is disrupted, manifesting as high-frequency oscillations and curvature spikes in local dimensions.

[0075] To quantify the extent of damage to this physical manifold, the system calculates the manifold curvature abrupt change factor for each modal eigenvector. The calculation formula is as follows:

[0076] in, Indicates the first The manifold curvature abrupt change factor of each modal feature branch, the larger the value, the more serious the incoherent noise or artificial modulation defects contained in the modal feature, the lower the physical reliability, and the unit is related to the square of the feature vector dimension; The total dimension of the feature vector is represented by a dimensionless integer. Indicates the discrete dimension index within the feature vector; , , They represent the eigenvectors respectively. In the , , Numerical components in each dimension; This indicates the square operation after taking the absolute value.

[0077] In S32, to transform the aforementioned physical curvature into normalized weights usable for fusion, the system employs a feedforward analytical mapping network for computation. This network abandons backpropagation-based black-box neural network learning and instead uses a rigorous analytical function for real-time solution, ensuring deterministic responses under extreme adversarial conditions. The mapping formula is as follows:

[0078] in, Indicates assignment to the first The dynamic fusion weights of each modal branch, with values ​​ranging from... And satisfy , is a dimensionless parameter; and They represent the first and the Manifold curvature abrupt change factor for each modal characteristic branch; This represents a preset minimum positive constant (e.g., taking...). This is used to prevent mathematical calculation overflow exceptions when the curvature is extremely small, resulting in a denominator of zero. This indicates the reciprocal operator.

[0079] Through this mapping logic, when a certain modality (e.g., an image domain branch) is... Suppression and interference led to During a surge, its corresponding weight The decay will rapidly approach zero in an inverse proportion. This dynamic evaluation and allocation mechanism based on manifold curvature, starting from the physical manifold smoothness, achieves automatic shielding against deteriorating channels and ensures the purity of multi-source information fusion.

[0080] In this embodiment, it should be noted that in S33, traditional feature concatenation is merely a simple stacking of data, which cannot uncover deep corroborative relationships between heterogeneous physical evidence. This step introduces a scaling dot product operator to forcibly physically bind the macroscopic spatial indications in the image domain with the microscopic phase distortions in the temporal domain.

[0081] Specifically, the system will assign weights Image-domain incoherent topological features are considered as instructions for finding suspicious targets in space, and are used to generate a query matrix. (Query). At the same time, weights were assigned. Local phase dispersion characteristics and weights The spatial geometric constraint features are cascaded and regarded as a base characterizing the distribution of hardware defects of the jammer on the underlying physical properties. These features are then used to generate a key matrix. (Key) and value matrix (Value).

[0082] The core computational formula for cross-modal interaction fusion is as follows:

[0083] in, This represents the enhanced feature vector output after performing cross-modal feature enhancement; This represents the query matrix generated by linear projection of image domain features; This represents the key matrix generated by the concatenated projection of temporal and auxiliary information features; This represents the value matrix containing physical attribute information, corresponding to the key matrix. The mathematical operator for transposing a matrix; This represents the feature dimension of the key vector in the mapping space, used to numerically scale the dot product result and prevent the gradient from vanishing when the dimension is too high and the softmax function becomes saturated. This represents the normalized exponential function operator, used to transform the dot product result into a weighted probability distribution.

[0084] The physical essence of this computational logic lies in the fact that only when the image domain ( When extracting topological mismatch features of suspected false targets at a certain spatial location, the key matrix is ​​retrieved. If this position also exhibits extremely strong phase dispersion distortion in the time domain, the vector dot product of the two... Only then will a peak response be generated. This cross-modal dot product operation forms a physical evidence chain from the spatial target to the underlying hardware defect with physical correlation. Only the energy of the false target that satisfies both of these physical properties will be nonlinearly amplified and output to the subsequent enhanced feature vector.

[0085] Please see Figure 5 As shown in this embodiment, it should be noted that in S4, in order to identify the target spatial coordinate jump caused by the inherent time delay of signal interception, processing and forwarding of the digital radio frequency memory jammer, the system introduces a dual time-series physical verification model for continuous multi-frame imaging results.

[0086] Specifically, in S41, to establish a mathematical description of the continuous motion of a real physical target, the system incorporates a Kalman filter based on the prior assumption of uniform linear motion. Its state prediction equation is defined as follows:

[0087] in, Indicates the first Frame time for the first The prior prediction column vector of the frame target state in the two-dimensional imaging plane (including range position, azimuth position and its first-order velocity component). Indicates the first The posterior state estimate column vector of the frame target; L represents the linear state transition matrix constructed based on the pulse repetition period of synthetic aperture radar imaging; This represents the process noise vector simulating aircraft vibration and random atmospheric disturbances. This indicates the discrete frame time sequence number of the synthetic aperture radar imaging.

[0088] Specifically, the target's state vector in the two-dimensional imaging plane is defined as including its range position, azimuth position, and first-order velocity components, combined with the pulse repetition period. (in This represents the pulse repetition frequency of a synthetic aperture radar system, in units of... The linear state transition matrix L is specifically defined as follows:

[0089] This provides rigorous mathematical support for the iterative estimation of continuous motion trajectories.

[0090] In S42, the system calculates the current number of... Mahalanobis distance between the actual observed target position in the frame and the physical predicted position by the Kalman filter. This directly transforms the hidden processing delay of the jammer into a quantifiable discriminant factor (position change factor). The calculation formula is as follows:

[0091] in, This represents the Mahalanobis distance between the calculated target observation position and the physical predicted trajectory, and is a non-negative scalar. Indicates the first The frame is actually the column vector of the target centroid observation coordinates calculated from the radar image; This represents the observation transformation matrix that maps the high-dimensional state space to the imaging coordinate observation space; This represents the inverse of the information covariance matrix of the predicted residuals; This represents the matrix transpose operator.

[0092] like If the chi-square distribution signal threshold set based on the system ranging accuracy is exceeded, it proves that the target has undergone a sudden change in instantaneous displacement between adjacent frames that exceeds the physical inertial constraints. This is a typical spatial characteristic of deception interference.

[0093] Wherein, the new information covariance matrix The calculation model is ,in Characterizing the first The one-step prediction covariance matrix of the frame state. The measurement noise covariance matrix characterizes the radar system, and together they determine the statistical distribution boundary for Mahalanobis distance calculation.

[0094] Specifically, the confidence threshold for the chi-square distribution set based on the system ranging accuracy is determined using a finite number of historical data. First, the position prediction residual data of 10,000 consecutive frames of observations of a known static cooperative target by the carrier aircraft under normal uniform straight flight conditions is extracted. The squared Mahalanobis distance values ​​of these true residual sequences are calculated. Then, the set of distances is fitted with a chi-square distribution with 4 degrees of freedom. The critical value corresponding to the cumulative probability density reaching 0.99 is selected as the confidence threshold. For example, historical data statistics show that 99% of the squared Mahalanobis distance values ​​of normal observations are distributed between 0 and 7.8. Therefore, the confidence threshold of 7.8 is selected, which can effectively distinguish and identify abnormal position jumps caused by the processing delay of the jammer.

[0095] In S43, the phase history of the real target must strictly match the instantaneous relative motion between the radar platform and the ground. To prevent high-order jammers from smoothly camouflaging the trajectory, the system further performs a rigid check of the center Doppler frequency shift. The Doppler frequency shift residual... The calculation model is as follows:

[0096] in, This represents the absolute value of the residual between the observed Doppler frequency and the theoretical Doppler frequency, expressed in Hertz (Hz). This indicates that by analyzing the first The real-time observation center Doppler frequency extracted from the phase gradient evolution in the enhanced feature vector of frame synthetic aperture radar is expressed in Hertz (Hz). This indicates the radial relative flight speed of the radar-carrying aircraft with respect to the target observation area, expressed in meters per second (m / s). This represents the instantaneous angle of view of the target relative to the center of the radar antenna beamforming, expressed in radians (rad). The wavelength of the electromagnetic waves emitted by the radar system is expressed in meters (m). This represents absolute value operations.

[0097] The Mahalanobis distance calculated in the above steps and Doppler residuals Together, they constitute the time-series consistency evaluation factor. If Exceeding the tolerance limit set based on synthetic aperture radar Doppler resolution (e.g.) If the phase center evolution of the target signal deviates physically from the actual geometric mapping, the system will determine it as a physical mismatch caused by spatial geometric parameter estimation errors in the digital radio frequency memory, rather than a real ground object echo.

[0098] Specifically, the upper tolerance limit set based on the synthetic aperture radar Doppler resolution is determined by a finite number of historical data sets. First, raw radar echo data is extracted from the aircraft under different flight attitude angles (yaw, pitch, roll) and atmospheric turbulence disturbances. Then, clutter adaptive estimation technology is used to extract the estimation error set of the true Doppler center frequency. Finally, statistical variance analysis is performed on this error set to extract the mean plus three standard deviations. The physical deviation of the extreme value is used to calibrate the upper limit of the tolerance to 50Hz, so as to prevent the slight attitude jitter of the carrier aircraft itself from being misjudged as deceptive interference.

[0099] Please see Figure 6 As shown in this embodiment, it should be noted that in S5, in order to solve the problem of false alarms caused by the instantaneous scintillation interference of synthetic aperture radar or strong ground clutter disturbance in single-frame decision-making, this invention constructs a recursive Bayesian probabilistic fusion framework with Markov properties, which filters logic spikes through the "physical inertia" of the time dimension.

[0100] Specifically, in S51, the system first defines the state transition logic of the target's physical attributes between discrete imaging frames. According to first principles, the electronic countermeasures behavior of a digital radio frequency memory jammer exhibits statistical continuity over a nanosecond timescale; that is, the jamming signal will not randomly change between the real target and the deceptive jamming state within two adjacent radar imaging frames. To quantify this physical constraint, the system introduces a Markov transition-restricted smoothing operator to truncate and restrict the transition rate of the posterior probability. The modified formula for calculating the jamming posterior probability is as follows:

[0101] in, Indicates the first The perturbation posterior probability after state transition constraint correction at frame time is a dimensionless parameter; This represents the raw posterior probability calculated directly from the multimodal physical evidence of the current frame using the Bayesian criterion; Indicates the previous imaging frame (the first one) The final decision probability of the frame output; The preset physical maximum step size threshold for state transition between adjacent frames (set to 0.15 in this embodiment) is used to limit the amplitude of a single probability fluctuation. This represents a mathematical boundary truncation function that forcibly restricts the value of the first variable to the closed interval defined by the latter two variables. Indicates the first A collection of multimodal physical evidence acquired from frames; This represents the set of physical evidence acquired in the previous frame.

[0102] Among them, the preset physical maximum step size threshold for inter-frame state transition is... The process of determining the value is based on a finite number of historical data sets. First, at least 5,000 sets of continuous multi-frame temporal imaging feature sequences of normal maneuvering targets and deceptive jamming targets in a real radar system are extracted. Then, the first derivative of the natural fluctuation of the Bayesian posterior probability of these sequences within the continuous radar observation period is calculated. Finally, the amplitude of the jump in these 5,000 sets of posterior probabilities is fitted with a normal distribution, and the upper limit of the maximum probability increment corresponding to its 95% confidence interval is selected to obtain the physical maximum step size threshold. This ensures that the threshold can truly reflect the continuous laws governing state transitions in the physical world.

[0103] In S52, the system performs recursive accumulation of evidence across multiple frames. Given the known... Given the likelihood probability distribution of each piece of physical evidence in the frame (image topological mismatch, phase dispersion pulse, Mahalanobis distance anomaly), a recursive update is performed using the law of total probability:

[0104] in, : indicates the current number as of now The cumulative posterior probability of the target being deception and interference, given the sequence of all historical physical observation evidence for the frame. : Indicates that, under the premise that the true state of the target is deception and interference, the current i-th The conditional likelihood probability of the multimodal feature evidence vector observed in the frame; : indicates that after the previous imaging frame (the first 100 frames) The perturbation posterior probability after Markov transition constraint smoothing operator correction (frame) is used as the prior probability in the current frame computation; and : These represent two mutually exclusive physical states that the target may exist. To deceive and interfere with the category state, This represents the actual category status of ground features; : Represents the traversal state variable of the target's category, and its set of values ​​is ; : Represents a normalization mathematical operator that sums over all possible target class states (disturbance and real), used to form the denominator of the full probability distribution; : Indicates that the target state is (i.e., the state is) or Under the assumption that ), the current number The conditional likelihood probability of actually observing the multimodal feature evidence in a frame; : indicates the previous imaging frame (the first 1000 frames) The historical posterior probability of the target belonging to state category c is obtained by cumulative calculation of frames.

[0105] Verification using specific scenario data: Assume that in the first... In the first frame, the radar experienced a momentary multipath scintillation interference, causing a transient jump in the manifold curvature abrupt change calculated in the image domain, thus altering the original Bayesian calculation. From the previous frame Rising sharply to Without Markov constraints, the system will immediately trigger a false alarm. However, by using the smoothing operator of this invention, the following settings are made: The maximum probability increase for the current frame to be allowed is only By Cut off to The system successfully suppressed the logical discrimination oscillations introduced by the non-stationary fluctuations of the physical environment. Only when subsequent continuous Only when all frames above exhibit consistent interference features will the posterior probability steadily increase to a certain level. This achieves high-confidence interference detection.

[0106] Please see Figure 7 As shown in this embodiment, it should be noted that in S6, in order to solve the technical problem that the traditional amplitude zeroing suppression method will leave obvious black hole regions in the synthetic aperture radar complex image, thereby destroying the subsequent target automatic recognition algorithm's semantic understanding of the surrounding environment, this invention introduces a refined signal suppression and background restoration mechanism based on compressed sensing.

[0107] Specifically, in S61, once the target area is determined to be a deception interference area, the system generates a mask matrix in the complex domain. Set the weight of the coordinate pixels occupied by the interference target to The remaining background areas are set to In S62, a reconstruction optimization model is constructed using the sparsity principle of ground background in the discrete wavelet transform domain. Its sparse reconstruction cost function is defined as follows:

[0108] in, Represents the sparse projection coefficient vector of Norm, with units related to signal strength, is used to induce sparse solutions of signals in the transform domain; Represents the projection coefficients of the ground background under the discrete wavelet dictionary; This represents a complex image observation vector containing interfering holes after being processed by a physical mask; This represents the measurement sampling matrix determined by the field-programmable gate array (FPGA) preprocessing stage; This represents the sparse transform dictionary matrix composed of discrete wavelet bases; This represents the upper limit of the statistical noise error tolerance allowed by the algorithm, in watts (W). This represents an optimization operator that finds the global minimum value. express Norm Euclidean distance metric.

[0109] Specifically, this refers to the upper limit of the statistical noise error tolerance allowed by the algorithm. The method for determining the values ​​is based on a finite number of historical data. First, 5000 sets of real synthetic aperture radar ground object echo complex image data under interference-free environment are extracted. Discrete wavelet transform is performed on these clean background data and reconstruction is carried out using a sparse dictionary. The mean square error between the reconstructed image and the original image is calculated. These 5000 sets of mean square error data are fitted with a Gaussian distribution, and the mean plus two standard deviations are selected as the extreme value of the noise floor fluctuation. This extreme value is used to calibrate the upper limit of the statistical noise error tolerance. The value is 0.05W, thus ensuring that the compressed sensing reconstruction algorithm can strip away external deception interference energy while tolerating the inherent background thermal noise of the system.

[0110] The system employs an orthogonal matching pursuit algorithm to iteratively solve the aforementioned model within a digital signal processor. Each iteration searches for the dictionary atom most orthogonal to the residual signal, and progressively infers and fills in the original ground features of the disturbed area through the statistical correlation of surrounding uncontaminated pixels.

[0111] This computational logic solves the problem of secondary distortion in anti-interference processing. (The signal-to-interference ratio is...) Under extreme conditions, experiments show that the reconstructed background image maintains a high degree of correlation with the original pristine landscape. The above not only physically removes the deceptive energy injected by the digital radio frequency memory, but also preserves the integrity of the battlefield situation to the greatest extent, providing a reliable data foundation for subsequent precision strike decisions.

[0112] Please see Figure 8 As shown, this invention also provides an anti-spoofing interference system based on dynamic weights and timing constraints. This system is physically designed as a heterogeneous computing unit integrated within the signal processor of a satellite or spacecraft, used to execute the steps of the methods described in the preceding claims. The system includes: The field-programmable gate array (FPGA) monitoring module is directly connected to the front end of the receiver's analog-to-digital converter (ADC) via a high-speed bus. Internally, it uses hard-wired logic to implement real-time pipelined calculation of amplitude distortion gradients and maintains a hardware state machine for adaptive route truncation. This module overcomes the technical bottleneck of software detection failing to achieve microsecond-level environmental response.

[0113] The graphics processor spatial topology extraction module is responsible for performing two-dimensional matrix operations to extract incoherent topological features in the image domain using an anti-spoofing interference method based on dynamic weights and temporal constraints. It outputs an incoherent topological feature vector in the image domain by parallel computation of phase maps at different dilation rates. Specifically, this module is primarily responsible for performing computationally intensive two-dimensional matrix operations, namely, the cross-resolution unit sampling described in step S21. It utilizes thousands of stream processors to parallel compute phase maps at different dilation rates, outputting an incoherent topological feature vector in the image domain.

[0114] The digital signal processor's physical verification and decision-making core is responsible for executing the sequence analysis logic in the anti-spoofing interference method based on dynamic weights and temporal constraints. This logic involves extracting local phase dispersion features, generating dynamic fusion weights and performing cross-modal physical evidence chain interaction fusion, performing dual temporal physical verification, and accumulating multi-frame evidence. Specifically, as the system's main control center, this core is responsible for executing the sequence analysis logic in steps S22, S3, S4, and S5. Because these steps involve complex double-precision floating-point operations (such as Mahalanobis distance inversion, Bayesian probability recursion, and compressed sensing iteration), the digital signal processor's strong scalar computational capabilities ensure the rigor of the physical verification.

[0115] The zero-copy data bus scheduling module is used to schedule data transmission between internal system components in accordance with the high-speed serial computer extended bus standard, and drives the switching of computing power routes through hardware interrupt signals. Specifically, data transmission between internal system components follows the high-speed serial computer extended bus standard (PCIe 4.0), and the switching of computing power routes is driven by hardware interrupt signals. This design eliminates redundant latency caused by multiple data transfers between different processors, supporting the aforementioned... Hard real-time processing performance.

[0116] In this embodiment, it should be noted that future adversarial environments are not limited to monostatic radar; the system will evolve towards distributed multistatic / bistatic radar. A digital radio frequency memory jammer can only accurately simulate the Doppler phase in the main lobe direction of a single, specific radar node at any given time. For other nodes with bidirectional non-transmitter-receiver co-location geometry, the phase history of their forwarded signals easily reveals significant physical inconsistencies. This embodiment further expands upon Embodiment Seven (dual temporal physical verification) by introducing three-dimensional spatial consistency verification under a bistatic / multistatic detection system.

[0117] Specifically, the system's main control processing center aggregates the multimodal features extracted by each radar node within the network via a broadband data link. For the bistatic radar model, the bistatic Doppler frequency shift of the real target... The physical composition relationship of the velocity vectors of the transmitting and receiving machines must be strictly satisfied. The constraint model is defined as follows:

[0118] The characters in the above formula are defined as follows: : Represents the actual theoretical value of bistatic Doppler frequency shift, in Hertz (Hz); : Indicates the electromagnetic wave wavelength of the transmitter, in meters (m). : Represents the instantaneous velocity vector of the transmitter in three-dimensional space, with units of meters per second (m / s); : Represents the instantaneous velocity vector of the receiver in three-dimensional space, with units of meters per second (m / s); : Represents the unit direction vector (dimensionless) pointing from the target's physical center of mass to the transmitter; : Represents the unit direction vector (dimensionless) pointing from the target's physical centroid to the receiver; : Represents the inner product (dot product) operator of two three-dimensional spatial vectors.

[0119] The system calculates the deviation between the measured Doppler frequency shift at each receiving node and the theoretical value mentioned above. If the echo Doppler frequency observed by a certain space node deviates significantly from the bistatic physical constraint (i.e., the residual exceeds the synthetic Doppler resolution of the bistatic system), the system will trigger a strong weight reduction logic in the decision-level Bayesian fusion module. Using DS evidence theory, the basic probability of the confidence level that the target submitted by the suspicious node is true is assigned a weight reduction value.

[0120] This three-dimensional physical locking based on the geometric topology of multiple base stations in space makes any deception jammer that relies on single-point interception and forwarding mechanisms impossible to hide at the algorithm level, greatly improving the system's survivability in system-wide confrontation.

[0121] To further demonstrate the inventiveness and significant progress of this invention in a real electronic warfare environment, the following analysis combines a set of specific battlefield electromagnetic data to conduct a full-link extrapolation.

[0122] Bring in a working A spaceborne synthetic aperture radar reconnaissance scenario. The vehicle target is located at coordinates... It travels at a constant speed. Assume that... At that moment, the enemy's digital radio frequency memory jammer was activated, at a distance from the actual vehicle. A highly realistic deception target was generated, and the texture details of the radar image were masked by injecting broadband Gaussian noise.

[0123] Execute S1-S2 phases: Field-programmable gate arrays (FPGAs) at fast sampling points An instantaneous amplitude jump was detected, and the amplitude distortion gradient was calculated. This value far exceeds the upper limit of the threshold. The system immediately determines that the environment has deteriorated. The adaptive routing truncation mechanism is triggered, physically cutting off the data flow to the GPU.

[0124] At this point, the digital signal processor intervenes to extract the baseband unwound phase. The local phase divergence statistics within the current sliding window are calculated. Because this statistic is relative to the ground background. Presenting near With a multiplier of 10, the system has initially locked in the digital fingerprint of the digital radio frequency memory hardware quantization.

[0125] Execute S3-S4 phases: Due to noise covering the image domain, the manifold curvature abrupt change factor Extremely high (approximately) The system generates dynamic fusion weights. Automatically compressed to The time-domain branch weights Raised to The cross-modal interactive fusion module utilizes the residual weak spatial pointing information to perform a dot product with strong physical phase evidence, outputting a physical evidence vector with a high signal-to-interference ratio.

[0126] Kalman filter in continuous Frame observations revealed that the Mahalanobis distance between the predicted and observed positions of the suspected target remained consistently [value missing]. It exceeds the number of degrees of freedom. Chi-square distribution Confidence threshold. The calculated physical jump rate is equivalent to... .

[0127] Execution of S5-S6 phases: The recursive Bayesian probability flow is initiated. Due to the presence of the Markov smoothing factor, the posterior probability is changed from the initial... After accumulating through three frames, they evolved into , , until the first At frame rate, the joint likelihood value generated by positional abrupt changes and phase dispersion characteristics will Push towards .

[0128] The judgment was triggered. The system activated compressed sensing reconstruction. Utilizing... The Hamming window sampling and orthogonal matching pursuit algorithm for points, in The background filling has been completed.

[0129] Quantitative test conclusions: Through the Regression analysis of simulated adversarial examples in this invention, based on JSR. In extreme occlusion scenarios: Interference identification accuracy: determined by the traditional fixed-weight algorithm Upgraded to ; False Alarm Rate (FAR): Thanks to Markov probability smoothing, the false alarm rate is suppressed to... the following; System inference latency: After FPGA physical routing optimization, the average response time is... It meets the standard requirements for real-time performance.

[0130] This invention offsets the computational overload caused by electromagnetic suppression through physical hardware routing; effectively overcomes the hardware defects of digital radio frequency memory by utilizing quantization dispersion characteristics; and addresses the vulnerability of timing judgments through Markov Bayes recursion. These technologies work together to construct a logically rigorous, physically robust, and highly adaptive anti-spoofing interference technology system.

[0131] To implement the method proposed in this invention, this embodiment also provides an anti-spoofing interference system based on dynamic weights and timing constraints, which can be deployed on airborne or spaceborne platforms.

[0132] The system's hardware topology includes: Data preprocessing and monitoring module: This module physically consists of a field-programmable gate array (FPGA) and its peripheral high-speed analog-to-digital converter. It is configured to receive high-frequency echoes in real time and perform the internal calibration and amplitude distortion gradient calculations described in Example 2. It calculates and includes a hardware state machine to trigger a physical truncation of the direct memory access (DMA) channel when distortion exceeds the limit.

[0133] Multimodal Feature Parallel Extraction Module: This module consists of a graphics processing unit (GPU) and a digital signal processor (DSP). The GPU is configured to perform the cross-resolution unit sampling matrix operation of Embodiment 3 to extract incoherent topological features in the image domain; the DSP is configured to exclusively occupy the bus after the FPGA triggers route truncation to extract the local phase dispersion features of Embodiment 4. .

[0134] Dynamic weighting and interactive fusion module: This module serves as the core computing unit, executing the manifold curvature mutation factor in Example 5. Calculate and feedforward weight mapping, and perform cross-modal scaled dot product interactive fusion as described in Example 6 to output an enhanced physical evidence chain.

[0135] Timing Verification and Decision Control Module: This module includes a cache and a Bayesian inference accelerator. It is configured to perform Kalman filter prediction, Mahalanobis distance and Doppler history verification as described in Example 7, and recursive Bayesian cumulative calculation as described in Example 8. It ultimately outputs an interference determination conclusion and triggers compressed sensing signal suppression as described in Example 9.

[0136] The modules mentioned above are interconnected using a high-speed serial expansion bus (such as PCIe Gen4 / Gen5) to ensure low-latency, zero-copy data transfer between different heterogeneous computing nodes.

[0137] Embodiments of the present invention also provide an electronic device (such as a radar central processing computer), the electronic device comprising: at least one processor (such as a multi-core CPU co-processor DSP); and a memory communicatively connected to the at least one processor; wherein the memory stores computer program instructions executable by the at least one processor, the computer program instructions being executed by the at least one processor to enable the electronic device to execute the anti-spoofing interference method based on dynamic weights and timing constraints as described in any one of Embodiments 1 to 12 above.

[0138] Embodiments of the present invention also provide a non-volatile computer-readable storage medium storing computer program instructions thereon. When the program instructions are executed by the processor of a radar signal processing terminal, the method described in any one of Embodiments 1 to 12 is implemented. The storage medium may include, but is not limited to, solid-state drives (SSDs), read-only memory (ROM), random access memory (RAM), etc.

[0139] In some alternative embodiments, it is assumed that in a real-world adversarial mission, a synthetic aperture radar (SAR) system is tracking a vehicle target and encounters sophisticated digital radio frequency memory (DRFM) spoofing jamming. The system's operating parameters are set as follows: carrier wavelength... (i.e., X-band), pulse duration sampling frequency Radar incident angle .

[0140] Within the k-th imaging frame of the current analysis, the system extracts a suspected vehicle historical trajectory from the radar continuous observation sequence (or target historical point set), which contains N=10 historical observation points.

[0141] First, the field-programmable gate array (FPGA) detects instantaneous amplitude jumps at fast-time sampling points and calculates the amplitude distortion gradient. Extract the upper threshold set by the statistical model for ground clutter. .because The system immediately triggers hardware-level adaptive routing cutoff, physically severing the DMA transfer channel to the graphics processor. The digital signal processor takes over the bus, in a sliding window... Extract the instantaneous unwrapped phase sequence of the baseband complex signal and calculate the local phase divergence statistic. Assuming the calculation yields... However, due to the lack of quantization cutoff, the real background clutter has a reference frequency dispersion of only [missing information]. The system initially determined that the signal had microscopic phase distortion.

[0142] Subsequently, the system calculates the manifold curvature abrupt change factor. Due to noise suppression in the image domain, the manifold curvature of incoherent topological features in the image domain is... =15.8; Manifold curvature of local phase divergence characteristics =0.2; Manifold curvature of spatial geometric constraints =0.15. System default minimum positive constant. Substituting the values ​​into the feedforward analytical mapping network, the dynamic fusion weights in the image domain are calculated. Time-domain dynamic fusion weights Spatial geometric weights This indicates that the system adaptively masks the damaged image channel, with the underlying physical and geometric properties dominating the subsequent cross-modal interaction fusion, outputting an enhanced feature vector.

[0143] Next, the system performs a dual timing-physical verification. Assume the... The Mahalanobis distance between the observed centroid coordinates of the target calculated from the radar image and the physical predicted position using Kalman filtering. The chi-square distribution with 4 degrees of freedom is set with a 0.99 confidence threshold of [value missing]. .because This indicates that the target experienced a sudden change in displacement. Simultaneously, the radial velocity of the radar-launched aircraft was measured. Target perspective The theoretical Doppler frequency is If the actual observation center Doppler frequency is Then the residual greater than the upper tolerance limit The physical verification was deemed unsuccessful.

[0144] After obtaining the aforementioned likelihood evidence, the system inputs it into the decision-level recursive Bayesian fusion module. Assuming the interference posterior probability P(S_J|E_{k-1}) of the previous frame output is 0.10, the posterior probability calculated from the original likelihood of this frame suddenly increases to... The system reads the preset maximum physical step size threshold for Markov transitions. After being truncated by the posterior probability of the current frame, the corrected probability is... After accumulating probability flow across 5 consecutive frames, the posterior probability gradually stabilizes at... .because Greater than the preset decision threshold The system officially confirmed that the region was a deception interference, and constructed a mask matrix in the complex image domain, using discrete wavelet basis and orthogonal matching pursuit algorithm to perform sparse reconstruction of the real background.

[0145] The preset decision threshold is determined based on a limited number of historical data sets. First, 10,000 real and fake target sample features are collected, covering different signal-to-noise ratios (-20dB to 10dB) and various interference patterns (suppression and deception). Then, these sample features are input into the decision-level recursive Bayesian fusion module to calculate the final converged posterior probability set. Finally, the receiver operating characteristic (ROC) curve of false alarm rate versus detection probability is plotted to find the tangent probability value that maximizes the detection probability while keeping the false alarm rate below 1%. The preset decision threshold is thus determined to be 0.85, ensuring the objectivity and reliability of the interference identification boundary.

[0146] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for resisting deception interference based on dynamic weights and temporal constraints, characterized in that, include: The raw echo data of synthetic aperture radar is preprocessed in real time at the hardware level using field-programmable gate arrays, and hardware-level adaptive routing is triggered based on amplitude distortion gradient to output preprocessed data with computing power allocation. Multimodal features are extracted in parallel based on the preprocessed data. These multimodal features include image domain incoherent topological features, local phase dispersion features, and spatial geometric constraint features. The multidimensional vector space manifold curvature mutation factor of the multimodal features is quantified, and dynamic fusion weights are generated accordingly. Cross-modal physical evidence chain interaction fusion is performed on the multimodal features to output an enhanced feature vector. The multidimensional vector space manifold curvature mutation factor quantifies the multimodal features and generates dynamic fusion weights accordingly. Cross-modal physical evidence chain interaction fusion is performed on the multimodal features to output an enhanced feature vector, including: Calculate the second-order difference norm between adjacent feature dimensions and extract the manifold curvature abrupt change factor of each multimodal feature in the feature space to characterize the smoothness of the feature distribution; A feedforward analytical mapping network is used to convert the manifold curvature mutation factor into the corresponding modal reliability score, and then normalization is performed to generate dynamic fusion weights. A query matrix is ​​generated using incoherent topological features in the image domain to locate suspected targets. Key and value matrices are generated using local phase divergence features. Cross-modal physical evidence chain interaction fusion is performed using a scaling dot product operator to output an enhanced feature vector. Perform dual temporal-physical verification of the enhanced feature vector using both kinematics and Doppler history to obtain a temporal consistency evaluation factor; The process of performing dual temporal-physical verification of the enhanced feature vector using both kinematics and Doppler history yields temporal consistency evaluation factors, including: The continuous trajectory of the target is estimated iteratively based on the state transition equation of the Kalman filter. Calculate the statistical Mahalanobis distance between the observed position and the physical prediction value of the current imaging frame, as the position mutation factor; The residual values ​​between the observed Doppler frequency and the theoretical Doppler frequency are calculated using the imaging geometry model, and a temporal consistency evaluation factor is generated by combining the positional abrupt change factor. The enhanced feature vector and the temporal consistency evaluation factor are input into the decision-level recursive Bayesian fusion module, and Markov transition constraints are used to accumulate evidence across multiple frames, outputting the interference posterior probability. Based on the interference posterior probability, the deceptive interference target is identified, and signal suppression is performed on the identified interference area.

2. The anti-spoofing interference method based on dynamic weights and time constraints according to claim 1, characterized in that, The triggering of hardware-level adaptive routing truncation includes: the field-programmable gate array (FPGA) calculates the amplitude distortion gradient of the baseband signal in the fast time dimension in real time; when the amplitude distortion gradient exceeds the upper limit of the threshold set by the ground clutter statistical model, the adaptive logic dimensionality reduction truncation of the system bus is triggered, the graphics processor stops executing the high-energy-consuming cross-resolution unit sampling operation, and forces the output of a minimal constant background vector consistent with the preset dimension to the subsequent steps of cross-modal physical evidence chain interaction fusion. At the same time, the backbone bus bandwidth is released to the digital signal processor, which exclusively uses its resources to extract the local phase dispersion features.

3. The anti-spoofing interference method based on dynamic weights and temporal constraints according to claim 1, characterized in that, The extraction of the local phase dispersion features includes: extracting the instantaneous unwrapped phase sequence of the baseband complex signal based on the phase truncation physical mechanism of the digital-to-analog converter inside the digital radio frequency memory; locating the hardware-level phase jump by calculating the second-order differential of the phase of adjacent sampling points; and calculating the dispersion statistics within a short-time sliding window to identify the microscopic phase distortion introduced by the digital radio frequency memory.

4. The anti-spoofing interference method based on dynamic weights and time constraints according to claim 1, characterized in that, The generation of dynamic fusion weights includes: extracting the manifold curvature mutation factor of the multimodal features in the feature space, characterizing the smoothness of the feature distribution by calculating the second-order difference norm between adjacent feature dimensions, and using a feedforward analytical mapping network to convert the manifold curvature mutation factor into the corresponding modal reliability score.

5. The anti-spoofing interference method based on dynamic weights and temporal constraints according to claim 1, characterized in that, The extraction of incoherent topological features in the image domain includes: constructing a cross-resolution cell sampling matrix for spatial topological mismatch of discrete scattering centers. The sampling step size of the cross-resolution cell sampling matrix is ​​in a step-multiple relationship with the spatial resolution of the synthetic aperture radar system, and is used to extract the phase discontinuity topological map of false jamming targets in the cross-resolution cell.

6. The anti-spoofing interference method based on dynamic weights and time constraints according to claim 1, characterized in that, The cross-modal physical evidence chain interaction fusion includes: generating a query matrix using the incoherent topological features of the image domain to lock the suspected target location, generating a key matrix and a value matrix using the local phase dispersion features, and matching phase dispersion evidence at the suspected target location by scaling dot product operators to achieve a nonlinear correlation between the distribution of physical defects and spatial location.

7. The anti-spoofing interference method based on dynamic weights and temporal constraints according to claim 1, characterized in that, The dual temporal physical verification includes: estimating the continuous motion trajectory of the target based on Kalman filtering iteration, calculating the statistical Mahalanobis distance between the observed position and the physical prediction value as the position change factor, and using the imaging geometric model to verify the residual value between the observed Doppler frequency and the theoretical Doppler frequency, so as to determine whether the target signal conforms to the kinematic law of the ground object.

8. The anti-spoofing interference method based on dynamic weights and time constraints according to claim 1, characterized in that, The multi-frame evidence accumulation includes: using recursive Bayesian logic to accumulate the observation features of consecutive imaging frames probabilities, and introducing a Markov transition constraint smoothing operator during the probability transition process. By limiting the physical maximum step size threshold of state transition between adjacent frames, false posterior probability spikes introduced by flicker interference are filtered out.

9. The anti-spoofing interference method based on dynamic weights and temporal constraints according to claim 1, characterized in that, The execution signal suppression includes: using a reconstruction algorithm based on compressed sensing to construct a mask matrix in the complex image domain for regions determined to be deception interference; using a sparse transform dictionary composed of discrete wavelet bases and an orthogonal matching pursuit algorithm to restore the masked true background based on the statistical characteristics of surrounding uncontaminated pixels.

10. A deception-resistant interference system based on dynamic weights and temporal constraints, characterized in that, The system is used to implement the anti-spoofing interference method based on dynamic weights and timing constraints as described in any one of claims 1 to 9, and the system comprises: The field-programmable gate array monitoring module is directly connected to the front end of the receiver's analog-to-digital converter via a high-speed bus. Internally, it implements real-time pipelined calculation of amplitude distortion gradient through hard-wired logic and maintains a hardware state machine for performing adaptive route truncation. The graphics processor spatial topology extraction module performs two-dimensional matrix operations to extract incoherent topological features in the image domain using an anti-spoofing interference method based on dynamic weights and temporal constraints. It outputs incoherent topological feature vectors in the image domain by parallel computation of phase maps under different dilation rates. The digital signal processor physical verification and decision core executes the sequence analysis logic of the anti-spoofing interference method based on dynamic weights and time constraints, which includes extracting local phase dispersion features, generating dynamic fusion weights and performing cross-modal physical evidence chain interaction fusion, performing dual time-series physical verification, and accumulating multi-frame evidence, in order to perform double-precision floating-point operations. The zero-copy data bus scheduling module is used to schedule data transmission between internal system components in accordance with the high-speed serial computer extended bus standard, and drives the switching of computing power routes through hardware interrupt signals.

Images