Radar active slice forwarding interference resisting method based on high-order statistics

Abstract

The invention discloses a radar anti-active slice forwarding interference method based on high-order statistics, and the method comprises the steps: S10, carrying out the preprocessing of a radar receiving signal, carrying out the high-order cumulant construction of the radar receiving signal, completing the signal preprocessing, the calculation of a fourth-order cumulant, and the construction of a feature tensor, the high-order cumulant feature tensor of each time window corresponds to the corresponding time domain signal according to the time dimension and serves as a source of subsequent training input; s20, generating a label mask required by supervision training, and constructing an input feature and a supervision label so as to form a data structure in which the input and the label are paired; s30, constructing a MoE deep neural network based on high-order cumulant characteristics; s40, constructing a loss function and an optimization target, and calculating results of all loss items; s50, constructing an end-to-end blind interference suppression framework; s60, performing model training and parameter optimization; and S70, completing interference suppression and target output generation.

Description

Technical Field

[0001] This invention belongs to the field of radar signal processing and electronic countermeasures technology, specifically relating to a radar anti-active slice forwarding interference method based on high-order statistics. Background Technology

[0002] In response to the increasingly rampant active coherent deception jamming in modern electronic warfare environments, although radar systems can play a crucial role in sensing complex adversarial fields, their robustness is significantly constrained by slice-and-forward jamming. This type of jamming, by performing high-fidelity segmented sampling of the radar's transmitted signal and selectively forwarding it, can create a large number of false targets after pulse compression that are completely identical to the echo amplitude, Doppler, and modulation slope of the real target. This directly causes the failure of traditional constant false alarm rate (CFAR) detection and the complete submersion of the tracking gate, greatly weakening the radar's information acquisition capabilities.

[0003] Existing anti-jamming methods can be categorized into traditional methods and emerging anti-jamming approaches based on their application scenarios. Traditional methods often rely on frequency-domain adaptive filtering, time-frequency analysis, or prior knowledge of interference parameters. However, they struggle to effectively separate the real target from the interference component when facing highly realistic, low-interference-to-signal-ratio (INR) and unknown waveform parameters in slice-based forwarding interference. Especially when the interference signal and the real echo highly overlap in the time-frequency domain, conventional processing methods based on second-order statistics become completely ineffective due to their inability to capture the non-Gaussian nature and phase abrupt changes of the signal. While emerging methods have demonstrated some anti-jamming capabilities in specific scenarios, they still have significant shortcomings in complex real-world electromagnetic environments. On the one hand, most still require relatively accurate estimations of the interference sampling period, forwarding frequency, or INR; performance deteriorates sharply once parameters are unknown or waveforms change rapidly. On the other hand, they generally rely on the separability of the signal and interference in the time-frequency domain; when they completely overlap, the separation effect is significantly limited. Therefore, they perform poorly against highly realistic, parameter-unknown, and time-frequency overlapping slice-based forwarding interference, severely limiting the radar's target detection probability and measurement accuracy in highly contested environments. Summary of the Invention

[0004] In view of the above problems, the present invention provides a radar anti-active slice forwarding jamming method based on high-order statistics, so as to improve the problem that the performance of existing technologies is severely limited when facing slice forwarding jamming that is highly realistic, has unknown parameters, and completely overlaps with the time and frequency of the real target echo.

[0005] To solve the above-mentioned technical problems, the present invention adopts the following technical solution: A radar anti-active slice forwarding jamming method based on high-order statistics includes the following steps: S10, preprocessing the radar received signal and constructing the radar received signal using higher-order cumulants, is expressed as: (1) in, For a point in time, For the true target echo, For active slice forwarding of interference signals, For additive white Gaussian noise, signal preprocessing, fourth-order cumulant calculation, and feature tensor construction are performed to form the basic statistical input data for subsequent processing. Specifically, the higher-order cumulant feature tensor for each time window... With the corresponding time domain signal Corresponding to the time dimension, it serves as the source of subsequent training inputs; S20, based on the input feature tensor Construct the training sample set and the higher-order cumulant label space, and generate the label mask required for supervised training. We construct input features and supervision labels to form a data structure in which inputs and labels are paired. S30, construct a MoE deep neural network based on high-order cumulant features, with the input of each sub-expert network derived from the feature tensor. The subset of features, while the gating network simultaneously uses and For input; S40, Input label mask The probability of each sample point being identified as a real target. Recalculate higher-order cumulants Based on the statistical consistency constraints among the three, a loss function and an optimization objective are constructed, and the results of all loss terms are calculated. S50 constructs an end-to-end blind interference suppression framework, inputting normalized features. The mask generated by the model's forward propagation and As a result of subsequent interference suppression; S60, Model Training and Parameter Optimization: The parameter optimization process uses the loss function as the objective to iteratively update the parameters of the expert network and the gating network, ultimately outputting an anti-interference model. ,and It fully inherits the input and output format from the training phase at runtime; S70, input the preprocessing result r(t) into the trained mask generation model. and output the interference-suppressed This completes interference suppression and target output generation.

[0006] In one possible implementation, S10 specifically includes: S101 samples, aligns, and normalizes the energy of the original received signal, performs L2 norm normalization on each echo pulse, and ensures that the start time of different pulses is consistent through pulse alignment operation. S102, Construct a fourth-order cumulant FOC sequence to pair the preprocessed signal. Constructing fourth-order cumulants: (2) in This represents a fourth-order cumulant operator; S103, constructing fourth-order cumulant diagonal slices or 1.5D spectra for direct detection and selection of interference fragments: (3) Alternatively, a 1.5-dimensional cumulative spectrum can be constructed, and by analyzing the position of the sequence peaks, blind detection and coarse localization of the interference slice boundaries can be achieved, thus obtaining a coarse detection sequence for the interference fragment. ; S104, based on multiple latency dimensions Constructing multi-channel high-order statistical feature tensors This is used as input for subsequent neural networks.

[0007] In one possible implementation, S20 specifically includes: S201, Constructing training input features. The training input consists of two parts: the first is a higher-order cumulant tensor. The second is the corresponding time-domain signal window. These two form a joint input as training input features: (4) S202, Construct output labels, temporal interference suppression mask, and construct for each temporal sampling point. mask ,in, This indicates that the sample belongs to an interference segment. Represents the real target or noise, used to train the neural network to output the optimal interference suppression mask.

[0008] In one possible implementation, S30 specifically includes: S301, firstly, constructs high-order cumulative quantum experts (HOC-Experts), including: constructing A lightweight sub-expert network, where each expert processes a different cumulant slice or a different latency depth, where the input of the k-th expert is... Each expert outputs an interference probability curve: (5) S302, Next, a gating network is constructed. The gating network calculates the weights of each expert based on the non-Gaussianity and higher-order spectral distribution of the current input: (6) And ensure that the sum of the weights is 1: (7) S303, the expert output fusion yields the final output of the interference mask prediction: (8) in, This represents the probability that each sample point is identified as a real target.

[0009] In one possible implementation, S40 includes: S401, construct the interference suppression mask supervision loss, which takes the form of a binary classification cross-entropy loss function: (9) S402, constructing a higher-order statistical consistency loss: To ensure that the suppressed signal remains consistent with the higher-order statistical characteristics of the true LFM echo, a fourth-order cumulant residual constraint is introduced: (10) in: (11) S403, Construct the fusion loss function. The fusion formula for the total loss function is as follows: (12) in To balance the hyperparameters.

[0010] In one possible implementation, S50 includes: S501, Set the higher-order cumulant feature input, and convert the higher-order cumulant tensor... Input the MoE neural network model; S502 achieves zero-prior-interference detection through forward propagation of a neural network; S503, obtaining the mask probability sequence Then, it can be directly used as a soft mask to receive signals using formula (11). The target echo estimate after interference suppression is obtained, and finally the target echo after interference suppression is obtained.

[0011] In one possible implementation, S60 includes: S601 uses stochastic gradient descent to optimize the loss function, and uses the Adam optimizer to update the model parameters on the training data to minimize the total loss function; S602, train until convergence to obtain the final robust neural network model. When the loss function converges or the performance on the validation set stabilizes, the final robust network model is obtained. .

[0012] In one possible implementation, S70 includes: S701 performs interference suppression on actual radar data, converting the original received signal... Input the trained model The time-domain signal after interference suppression is obtained: (14) S702 outputs the range-dimensional or time-domain signal after interference suppression, which is used for subsequent range compression, pulse compression imaging, target detection, and tracking tasks, ensuring that the real target is still completely preserved under strong slice relay interference.

[0013] The present invention has the following beneficial effects: (1) Even under extreme conditions where radar signals and interference time and frequency highly overlap, efficient and reliable interference and signal separation can still be achieved. Taking advantage of the essential statistical difference that the theoretical value of the fourth-order cumulant of the linear frequency modulated real target echo is zero, while the slice relay interference has a significant non-zero peak in the fourth-order cumulant due to the phase discontinuity introduced by the hard slice, by constructing a diagonal slice or 1.5-dimensional spectrum of the fourth-order cumulant, the precise blind detection and localization of the interference segment boundary can be achieved directly in the high-order statistical domain. This completely eliminates the dependence on time and frequency separability. Even in the case of complete time and frequency overlap where all existing methods have completely failed, it can still robustly achieve effective suppression of interference and complete preservation of the real target.

[0014] (2) Achieving zero-priority blind anti-interference effect in complex electromagnetic environments with completely unknown interference parameters, waveform agility, and multi-source interference. A pure data-driven end-to-end high-order cumulant feature extraction and interference suppression joint optimization framework was constructed. The core adopts a Mixture-of-Experts (MoE) architecture, which uses multiple lightweight high-order cumulant quantum experts to mine different time delay dimensions and high-order spectral features in parallel, and then dynamically fuses them according to the non-Gaussian distribution of the current signal by an adaptive gating network, finally outputting the optimal time-domain interference suppression mask in one go. Without any explicit estimation or prior assumptions about the interference sampling period, number of forward passes, interference-to-signal ratio, and waveform parameters, the interference can be accurately eliminated in a single forward propagation, completely breaking through the dependence of existing methods on parameter estimation, and enabling the radar to maintain extremely high detection and tracking performance in scenarios with unknown parameters, waveform agility, and multi-source complex interference. Attached Figure Description

[0015] Figure 1This is a flowchart illustrating the steps of a radar anti-active slice forwarding interference method based on higher-order statistics according to an embodiment of the present invention. Detailed Implementation

[0016] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0017] See Figure 1 The diagram shows a flowchart of a radar anti-active slice forwarding interference method based on higher-order statistics according to an embodiment of the present invention, including the following steps: S10, preprocessing the radar received signal and constructing the radar received signal using higher-order cumulants, is expressed as: (1) in, For a point in time, For the true target echo, For active slice forwarding of interference signals, For additive white Gaussian noise, signal preprocessing, fourth-order cumulant calculation, and feature tensor construction are performed to form the basic statistical input data for subsequent processing. Specifically, the higher-order cumulant feature tensor for each time window... With the corresponding time domain signal Corresponding to the time dimension, it serves as the source of subsequent training inputs; S20, based on the input feature tensor Construct the training sample set and the higher-order cumulant label space, and generate the label mask required for supervised training. We construct input features and supervision labels to form a data structure in which inputs and labels are paired. S30, construct a MoE deep neural network based on high-order cumulant features, with the input of each sub-expert network derived from the feature tensor. The subset of features, while the gating network simultaneously uses and For input; S40, Input label mask The probability of each sample point being identified as a real target. Recalculate higher-order cumulants Based on the statistical consistency constraints among the three, a loss function and an optimization objective are constructed, and the results of all loss terms are calculated. S50 constructs an end-to-end blind interference suppression framework, inputting normalized features. The mask generated by the model's forward propagation and As a result of subsequent interference suppression; S60, Model Training and Parameter Optimization: The parameter optimization process uses the loss function as the objective to iteratively update the parameters of the expert network and the gating network, ultimately outputting an anti-interference model. ,and It fully inherits the input and output format from the training phase at runtime; S70, input the preprocessing result r(t) into the trained mask generation model. and output the interference-suppressed This completes interference suppression and target output generation.

[0018] In a specific application example, S10 specifically includes: S101 samples, aligns, and normalizes the energy of the original received signal, performs L2 norm normalization on each echo pulse, and ensures that the start time of different pulses is consistent through pulse alignment operation. S102, Construct a fourth-order cumulant FOC sequence to pair the preprocessed signal. Constructing fourth-order cumulants: (2) in, This is the discrete-time lag, expressed in units of the number of sampling points; This represents the fourth-order cumulant operator, which can be calculated using sample moment estimation. S103, constructing fourth-order cumulant diagonal slices or 1.5D spectra for direct detection and selection of interference fragments: (3) in, For a single time delay variable, the range of values ​​is... Consistent.

[0019] Alternatively, a 1.5-dimensional cumulative spectrum can be constructed, and by analyzing the position of the sequence peaks, blind detection and coarse localization of the interference slice boundaries can be achieved, thus obtaining a coarse detection sequence for the interference fragment. ; S104, based on multiple latency dimensions Constructing multi-channel high-order statistical feature tensors This is used as input for subsequent neural networks.

[0020] More specifically, for each aligned echo pulse Energy normalization is performed to make the normalized signal... satisfy: (13) Among them, subscript Indicates the pulse number. Let be the energy of the pulse in the time dimension. The above normalization operation can eliminate the influence of the overall energy difference between different pulses on the estimation of higher-order statistics.

[0021] The normalized received signal obtained through preprocessing Based on this, its fourth-order cumulant is calculated using formula (2). For each echo pulse, a segment of length is selected. Using the time window signal as the calculation object, the corresponding fourth-order cumulative value is calculated within the preset time delay range. According to the statistical characteristics of the true target echo of linear frequency modulation, its fourth-order cumulative value theoretically approaches zero; while the slice forwarding interference, due to the phase discontinuity introduced by hard slicing and time-selective forwarding, will produce obvious non-zero spikes in the fourth-order cumulative value domain, providing high-order statistical features for the detection and location of subsequent interference segments.

[0022] Through calculation By analyzing the amplitude sequence and combining it with simple threshold detection or peak detection algorithms, higher-order cumulant peaks corresponding to the locations of interference slices can be obtained. Alternatively, a 1.5-dimensional spectrum of the fourth-order cumulant can be constructed, i.e., for... By summing or transforming along a partial time delay dimension, a spectral sequence for a single time delay variable is obtained. Analyzing the peak positions in this spectral sequence allows for coarse detection and boundary localization of interference segments. Based on the detection results, a coarse detection sequence representing the location of interference is formed. ,in It is a time-varying binary or multi-level sequence used to indicate the time interval where slice forwarding interference is suspected.

[0023] In practical implementation, in order to effectively input higher-order statistical information into subsequent neural network models, constructing higher-order cumulants into structured feature tensors specifically includes: in multiple combinations of different time delays Below, for fourth-order cumulants The real, imaginary, or amplitude values ​​are iterated and rearranged to form a tensor containing multi-channel high-order statistical features. This feature tensor can be written as a tensor of size . The matrix, where, This represents the number of samples or windows along the time dimension. This represents the high-order cumulative feature dimension extracted at each time point. In practical implementation, different diagonal slices, different fixed-delay planes, or different statistical transformation results can be regarded as feature channels and combined in series or parallel to form the high-order statistical feature tensor of the input deep neural network. .

[0024] In a specific application example, S20 specifically includes: S201, Constructing training input features. The training input consists of two parts: the first is a higher-order cumulant tensor. The second is the corresponding time-domain signal window. , The sampled sequence is a discrete-time domain sequence, which can be a local window within a single pulse or a time period resulting from the superposition of multiple pulses. This implementation concatenates the two feature components along the feature dimension, forming a joint input that serves as the training input feature: (4) This represents a set of input samples used for model training, containing high-order statistical information and temporal amplitude information, providing a sufficient representation of the statistical differences between the neural network learning interference and the real target.

[0025] S202, To achieve interference suppression based on supervised learning, output labels and a temporal interference suppression mask are constructed, and a method is created for each temporal sampling point. mask ,in, This indicates that the sample belongs to an interference segment. Representing real targets or noise, this is used to train a neural network to output the optimal interference suppression mask. By recording the start and end times of interference forwarding in simulation or labeled data, a label mask sequence corresponding to each sampling point can be automatically generated. This mask sequence serves as the supervisory output of the neural network, guiding the network to learn where to suppress interfering components and where to preserve the real target and noisy background, thereby training the network to output the optimal temporal interference suppression mask.

[0026] In a specific application example, S30 specifically includes: S301, firstly, constructs high-order cumulative quantum experts (HOC-Experts), including: constructing A lightweight sub-expert network is used, where each expert processes a different cumulant slice or a different time delay depth. The sub-expert network can consist of several fully connected layers or one-dimensional convolutional layers, and its input is the corresponding high-order cumulant quantum feature. The input of the k-th expert is... Each expert outputs an interference probability curve: (5) in, Indicates the first Nonlinear mapping operators for expert networks, For sampling points The output above is the probability prediction value that the point belongs to the real target.

[0027] S302, to adaptively combine different high-order statistical feature subspaces, a gating network is constructed to weight the experts. The input to the gating network is the joint features. It includes high-order cumulant features and the original time-domain signal, and outputs the weight coefficients corresponding to each expert. Gated networks can employ a set of fully connected layers to map input features to weights. Their outputs are normalized using the SoftMax function. The gated network calculates the weights of each expert based on the non-Gaussian nature and high-order spectral distribution of the current input. (6) in, In order to be with the first The weight vector corresponding to each expert For the number of experts, For the first Normalized weighting coefficients for each expert; And ensure that the sum of the weights is 1: (7) Through the aforementioned gating network, the judgment results of different experts can be adaptively selected and fused based on the non-Gaussian and high-order spectral distribution of the current input signal.

[0028] S303, after obtaining the output of all experts and the weights of the gating network Then, the expert outputs are fused using a weighted summation method to obtain the final interference suppression mask probability sequence. The final output of the expert output fusion is: (8) in, Indicates the time sampling point Above, the predicted probability that this point belongs to the true target echo and does not belong to interference is... This represents the probability that each sample point is identified as a real target. Subsequent thresholding or direct... When used as a soft mask, it can suppress interference from slice forwarding and preserve the real target.

[0029] In a specific application example, S40 includes: S401, to make the network output interference suppression mask It can approximate the real label mask as closely as possible. An interference suppression mask supervision loss is constructed, which takes the form of a binary classification cross-entropy loss function: (9) in, For real labels, This predicts probabilities for the model. The loss function is used to constrain the network to output the correct interference suppression mask in the time domain.

[0030] S402, constructing a higher-order statistical consistency loss: To ensure that the suppressed signal remains consistent with the higher-order statistical characteristics of the true LFM echo, a fourth-order cumulant residual constraint is introduced: (10) in, Indicates based on The recalculated fourth-order cumulant sequence or diagonal slice vector This represents the Euclidean L2 norm. By minimizing this loss, the suppressed signal can be made to approximate the ideal LFM echo in terms of higher-order statistical properties.

[0031] After obtaining the predicted mask Then, it is applied to the original received signal. The estimated target echo after interference suppression is obtained. : (11) S403, Construct the fusion loss function, incorporating the mask supervision loss. loss of consistency with higher-order statistics The weighted fusion is performed, and the total loss function fusion formula is as follows: (12) in, The weighting coefficients, which balance the importance of mask accuracy and higher-order statistical consistency constraints, are non-negative hyperparameters. These are adjusted... The value of can achieve an appropriate balance between interference suppression strength and the fidelity of the true target waveform.

[0032] In a specific application example, S50 includes: S501: Set the higher-order cumulant feature input. Based on the processing results of S10, obtain the higher-order cumulant feature tensor corresponding to each time segment. and the original time-domain signal window And construct joint input features according to S20. , to the higher-order cumulant tensor The input MoE neural network model is processed by various sub-expert networks for the high-order cumulant features, while the time-domain signal part participates in the weight calculation of the gating network.

[0033] S502 achieves zero-priority interference detection through forward propagation of a neural network. The MoE neural network performs a single forward propagation of the input features and outputs the corresponding interference suppression mask probability sequence. This process does not rely on any prior knowledge about the interference sampling period, forwarding times, interference-to-signal ratio, or specific waveform parameters. The network adaptively distinguishes between slice-forwarding interference and the real target by learning higher-order statistical structures and non-Gaussian features, thereby achieving zero-priority blind detection and mask estimation of the interference location.

[0034] S503, obtaining the mask probability sequence Then, it can be directly used as a soft mask to receive signals using formula (11). The above steps yield the target echo estimate after interference suppression, ultimately resulting in the target echo after interference suppression. When a binary mask is required, it can also be obtained by... Thresholding is performed, classifying points above a certain threshold as real targets and points below the threshold as interference, thus forming a hard mask. Regardless of whether a soft or hard mask is used, the final output of this implementation is... All of these are time-domain signals after suppressing slice forwarding interference, which can be directly used in subsequent radar processing steps such as range compression, pulse compression imaging, target detection and tracking.

[0035] In a specific application example, S60 includes: S601 uses stochastic gradient descent to optimize the loss function, and uses the Adam optimizer to update the model parameters on the training data to minimize the total loss function; S602, train until convergence to obtain the final robust neural network model. When the loss function converges or the performance on the validation set stabilizes, the final robust network model is obtained. .

[0036] More specifically, regarding the total loss function To minimize the loss, the Adam optimizer incorporates first- and second-order momentum estimations, which adaptively adjust the learning rate to improve convergence speed and stability. In each iteration, several sets of input features and corresponding labels are selected in batches from the training sample set, and the total loss for the current batch is calculated. Then, the gradients of the parameters of each layer of the network are calculated based on the loss, and the parameters are updated using the Adam optimizer to gradually reduce the loss function.

[0037] As the number of training iterations increases, the model can be considered converged when the overall loss on the validation set no longer decreases significantly and the interference suppression performance index tends to stabilize. At this point, the trained network parameters are fixed to form the final anti-interference neural network model. .Model It records the nonlinear mapping relationship from high-order cumulant characteristics and time-domain signals to interference suppression masks, which can be used for online interference suppression during actual radar operation.

[0038] In a specific application example, S70 includes: S701 performs interference suppression on actual radar data, converting the original received signal... Input the trained model The time-domain signal after interference suppression is obtained: (14) S702 outputs the range-dimensional or time-domain signal after interference suppression, which is used for subsequent range compression, pulse compression imaging, target detection, and tracking tasks, ensuring that the real target is still completely preserved under strong slice relay interference.

[0039] More specifically, the raw received signal acquired by the radar receiver First, preprocessing and higher-order cumulants are calculated according to S10, and then input features are constructed based on S20 and S50 and input into the trained anti-interference model. .Model The model performs forward inference on the input signal and outputs the corresponding masked or suppressed signal. If the model directly outputs the mask probability sequence... According to: (15) The time-domain signal after interference suppression is obtained; if the mask application component has been integrated into the model, the output is the target echo estimate after interference suppression.

[0040] The time-domain signal after interference suppression The signal is then output to the subsequent radar signal processing link. Further adjustments can be made based on the radar architecture. The system employs four-way filtering, pulse compression, range Doppler processing, constant false alarm rate (CFAR) detection, and tracking filtering. Even when the time-frequency of the slice-based relay interference and the true target echo completely overlap and the interference parameters are unknown, it effectively eliminates interference segments while preserving the true target energy. Therefore, even in environments with strong active slice-based relay interference, it can significantly improve radar target detection probability and tracking accuracy.

[0041] It should be understood that the exemplary embodiments described herein are illustrative and not restrictive. Although one or more embodiments of the invention have been described in conjunction with the accompanying drawings, those skilled in the art will understand that various changes in form and detail may be made without departing from the spirit and scope of the invention as defined by the appended claims.

Claims

1. A radar method for countering active slicing and repeating jamming based on high order statistics, characterized in that, The method comprises the following steps: S10, pre-processing and high-order cumulant construction of radar receiving signals, denoted as: (1) wherein, is a time point, is a real target echo, is an active slice forwarding jamming signal, is additive white Gaussian noise, completing signal preprocessing, fourth-order cumulant calculation and feature tensor construction, forming the basis statistical input data for subsequent processing, wherein the high-order cumulant feature tensor of each time window corresponds to the corresponding time domain signal corresponds in time dimension, serving as the source of the required training input in the subsequent. S20, according to the input feature tensor Constructing a training sample set and a high-order cumulant label space to generate a label mask required for supervised training , constructing input features and supervised labels, thereby forming an input and label paired data structure; S30, constructing a MoE deep neural network based on high-order cumulant features, the input of each sub-expert network is derived from the feature tensor of the sub-feature set, while the gating network takes and as input; S40, input label mask probability that each sampling point is determined as a real target , recalculate high-order cumulants , according to the statistical consistency constraint among the three, construct the loss function and optimization target, and calculate the results of all loss terms S50, construct an end-to-end blind interference suppression framework, input normalized features , mask generated by model forward propagation and as subsequent interference suppression results; S60, model training and parameter optimization, wherein the parameter optimization process targets a loss function, iteratively updates the expert network and gating network parameters, and finally outputs an anti-interference model However Inherit the input and output formats of the training phase completely at runtime; S70, inputting the preprocessed result r(t) into the trained mask generation model and outputting the interference-suppressed , completing interference suppression and target output generation.

2. The radar anti-active-jammer method based on high-order statistics according to claim 1, wherein, The S10 specifically comprises: S101, sampling, pulse alignment and energy normalization of the original receiving signal, L2 norm normalization of each echo pulse, and pulse alignment operation to ensure the consistency of the starting time of different pulses; S102, construct a fourth-order cumulant (FOC) sequence pair for the preprocessed signal constructing a fourth-order cumulant (2) wherein denotes the fourth-order cumulant operator; S103, constructing a fourth-order cumulant diagonal slice or a 1.5-dimensional spectrum for interference segment direct detection selection: (3) Or construct its 1.5-dimensional cumulative spectrum, by analyzing the sequence peak position, realize the blind detection of interference slice boundary and rough positioning, get the interference segment rough detection sequence ; S104, based on the multi-delay dimension Constructing multi-channel high-order statistical feature tensors for subsequent neural network input.

3. The radar anti-active-jammer method based on high-order statistics according to claim 2, wherein, The S20 specifically comprises: S201, construct training input features, the training input consists of two parts, the first is the high-order cumulant tensor , the second is the corresponding time domain signal window , the two form a joint input as a training input feature: (4) S202, construct an output label, a time domain interference suppression mask, construct for each time domain sampling point mask wherein, indicates that the sample belongs to an interference segment, indicates a real target or noise, used to train a neural network to output an optimal interference suppression mask.

4. The radar anti-active-slice-repeater-jamming method based on high-order statistics according to claim 3, characterized in that, The S30 specifically comprises: S301, first construct high-order cumulant sub-experts HOC-Experts, including: constructing a lightweight sub-expert network, each expert processing different cumulant slices or different delay depths, wherein the input of the kth expert is , and each expert outputs an interference probability curve: (5) S302, secondly constructing a gating network, the gating network calculating the weight of each expert according to the current input non-higher order statistics and higher order spectrum distribution: (6) And the weight sum is 1: (7) S303, expert output fusion to obtain a final interference mask prediction final output: (8) wherein, represents the probability that each sample point is determined to be a real target.

5. The radar anti-active-jammer method based on high-order statistics according to claim 4, characterized in that, The S40 comprises: S401, constructing an interference suppression mask supervision loss, which is in the form of adopting a binary classification cross-entropy loss function: (9) S402, constructing a high-order statistical consistency loss, in order to ensure that the suppressed signal remains consistent with the high-order statistical characteristics of the real LFM echo, a fourth-order cumulant residual constraint is introduced: (10) Wherein: (11) S403, constructing a fusion loss function, and the total loss function fusion formula is as follows: (12) wherein is a balancing hyperparameter.

6. The radar anti-active-jammer method based on high-order statistics according to claim 5, wherein, The S50 comprises: S501, set high-order cumulant feature input, set high-order cumulant tensor input MoE neural network model; S502, realizing zero-prior interference detection through forward propagation of the neural network; S503, obtaining the mask probability sequence After that, it can be directly used as a soft mask on the received signal After that, it can be directly used as a soft mask on the received signal After that, it can be directly used as a soft mask on the received signal 7. The radar anti-active-slice-repeater-jamming method based on high-order statistics according to claim 6, characterized in that, The S60 comprises: S601, using a stochastic gradient descent method to optimize the loss function, using an Adam optimizer to update the model parameters on the training data, and minimizing the total loss function; S602, training to convergence obtains the final anti-interference neural network model, when the loss converges or the verification set performance is stable, the final anti-interference network model is obtained .

8. The radar anti-active-slice-repeater-jamming method based on high-order statistics according to claim 7, characterized in that, The S70 comprises: S701, interference suppression is performed on the actual radar data, and the original received signal inputting the trained model to obtain a time-domain signal after interference suppression: (14) S702, outputting a distance dimension time domain signal after interference suppression, and outputting for subsequent distance compression, pulse compression imaging, target detection and tracking tasks, to ensure that the real target is still complete under strong slice forwarding interference.

Images