Radar active jamming open set identification method and device based on reconstruction-guided hierarchical adversarial learning

Abstract

This invention relates to a method and device for open-set identification of active radar interference based on hierarchical adversarial learning guided by reconstruction. The method includes: performing time-frequency transformation on the received radar signal to generate network input samples; inputting the training samples into a shared feature extraction network to extract features, with the shared feature extraction network outputting two paths; determining the sample to be tested as belonging to a corresponding known interference category when the maximum logit score is greater than or equal to a preset open-set decision threshold; otherwise, determining the sample to be tested as belonging to an unknown interference category. This invention constructs a hierarchical labeling system based on different interference mechanisms and time-frequency structure differences, dividing interference categories into coarse-grained labels and fine-grained labels, thereby enhancing the ability of the shared feature extraction network to express the hierarchical relationship between interference categories; introducing a hierarchical adversarial learning strategy further strengthens the discriminative features of fine-grained interference categories, improves the accuracy of identifying known interference types, and enhances the separability of known and unknown classes in the feature space.

Description

Technical Field

[0001] This invention relates to the field of radar electronic countermeasures and intelligent identification technology, and in particular to a radar active jamming open set identification method and device based on reconstruction-guided hierarchical adversarial learning. Background Technology

[0002] Radar active jamming identification is a crucial prerequisite for threat analysis and anti-jamming decision-making in electronic countermeasures systems. Traditional jamming identification methods typically rely on manually designed features and shallow classifiers, which significantly limit their identification performance and generalization ability under complex electromagnetic environments and diverse jamming conditions. Deep learning-based radar active jamming identification methods, however, eliminate the need for complex manual feature design and can automatically uncover the essential differences between different jamming types. Therefore, they have become a research hotspot for scholars both domestically and internationally in recent years, attracting widespread attention.

[0003] However, most existing deep learning-based radar active jamming identification methods still rely on the closed-set identification assumption, which assumes that the class space is consistent between the training and testing phases. In actual combat environments, new jamming types that did not appear during the training phase may emerge during the testing phase. In such cases, traditional closed-set identification models often misclassify unknown jamming as a known category. To address this issue, some research has begun to introduce open-set identification methods into radar active jamming identification.

[0004] Although existing research has made some progress in open set identification, most methods still rely mainly on classification boundaries for decision-making, and do not make sufficient use of the physical mechanism relationship between interference categories. In particular, when unknown interference and known interference are highly similar at the mechanism level and differ only in fine-grained time-frequency structure, existing methods are still prone to misjudgment. Summary of the Invention

[0005] To address the problem that radar active interference identification models in open environments lack the ability to identify unknown interference types and are prone to misclassifying unknown interference as known interference, the primary objective of this invention is to provide a radar active interference open set identification method based on reconstruction-guided hierarchical adversarial learning that improves the accuracy of identifying known interference types and enhances the ability of shared feature extraction networks to detect and distinguish unknown active interference types in open environments.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: a radar active jamming open set identification method based on reconstruction-guided hierarchical adversarial learning, the method comprising the following sequential steps:

[0007] (1) Perform time-frequency transformation on the radar received signal to generate network input samples. Based on the interference mechanism and time-frequency structure differences of the interference type, input samples are provided to the network. Construct coarse-grained and fine-grained labels and combine them into training samples;

[0008] (2) Input the training samples into the shared feature extraction network to extract features. The shared feature extraction network has two outputs. One output is sent to the fine-grained classification branch and the coarse-grained classification branch respectively. The coarse-grained classification branch forms an adversarial learning process with the shared feature extraction network through the gradient reversal mechanism. The other output is sent to the time-frequency reconstruction branch to reconstruct the training samples and retain the key time-frequency structure information of the interference signal.

[0009] (3) During the testing phase, the maximum logit score output by the fine-grained classification branch is used as the open set decision criterion. When the maximum logit score is greater than or equal to the open set decision threshold, the sample to be tested is determined to be the corresponding known interference category. When the maximum logit score is lower than the open set decision threshold, the sample to be tested is determined to be the unknown interference category.

[0010] Step (1) specifically refers to: collecting the radar received signal, assuming the radar received signal is... Then at any time t, we have:

[0011] ;

[0012] in, Indicates the target echo signal. Indicates an active interference signal. Indicates a noise signal;

[0013] To simultaneously characterize the structural features of the interference signal in both the time and frequency domains, the received signal... Perform a short-time Fourier transform to obtain the time-frequency representation. :

[0014] ;

[0015] in, Represents the window function. This indicates the translation position of the window function on the time axis. Represents frequency variables;

[0016] Further take The amplitude constitutes the time-frequency diagram of the interference samples. :

[0017] ;

[0018] right Normalization, cropping, and scaling are performed to obtain the network input samples. ;

[0019] Depending on the different interference mechanisms, input samples are used for each network. Simultaneously, fine-grained and coarse-grained tags are constructed, with the fine-grained tag denoted as... Coarse-grained label is denoted as ; where the superscript 'e' indicates fine-grainedness and the superscript 'c' indicates coarse-grainedness;

[0020] Coarse-grained labels are generated from fine-grained labels through a mapping function. To obtain, that is:

[0021] ;

[0022] Network input samples Fine-grained labels and coarse-grained labels Together they form the training samples, the first Each training sample is represented as .

[0023] In step (2), the output to the fine-grained classification branch and the coarse-grained classification branch respectively specifically refers to: assuming the shared feature extraction network is... Then we have:

[0024] ;

[0025] in, These are deep feature vectors;

[0026] Fine-grained classification branch Output the logit values ​​for all known fine-grained categories. :

[0027] ;

[0028] in, Parameters representing fine-grained classification branches;

[0029] Cross-entropy loss of fine-grained classification branches for:

[0030] ;

[0031] in, Indicates network input sample Predicted probabilities on fine-grained labels; Input samples for the network Quantity; Fine-grained labeling; This represents the fine-grained classification probability of the fine-grained classification branch for the i-th sample;

[0032] coarse-grained classification branches Output the logit values ​​for all known coarse-grained categories. :

[0033] ;

[0034] in, Parameters representing coarse-grained classification branches;

[0035] Cross-entropy loss of coarse-grained classification branches Defined as:

[0036] ;

[0037] in, This represents the predicted probability on the coarse-grained label; This represents the coarse-grained classification probability of the coarse-grained classification branch for the i-th sample;

[0038] A gradient inversion layer is inserted between the shared feature extraction network and the coarse-grained classification branch. Therefore, the coarse-grained branch minimizes... To learn mechanistic semantics, the shared feature extraction network receives the inverted gradient, thus suppressing the encoding of overly dominant coarse-grained common features; correspondingly, the fine-grained classification branch learns through... Optimization is being performed, while the shared feature extraction network is in and Training is conducted under the combined effect of these factors, among which... The temperature coefficient for controlling the gradient reversal intensity.

[0039] In step (2), the other output to the time-frequency reconstruction branch specifically refers to: assuming the last convolutional feature map of the shared feature extraction network is... First, use the bottleneck compression module. Obtaining bottleneck characteristics :

[0040] ;

[0041] Bottleneck characteristics Subsequently, through the bottleneck recovery module and decoder By gradually restoring, a reconstructed time-frequency diagram is obtained. :

[0042] ;

[0043] Mean square error loss function of time-frequency reconstruction branch Defined as:

[0044] ;

[0045] In the formula, Input samples for the network Quantity;

[0046] The time-frequency reconstruction branch includes a bottleneck compression module, a bottleneck recovery module, a decoder, and a mean square error loss function.

[0047] Step (3) specifically refers to: during the training phase, jointly optimizing the fine-grained classification branch, the coarse-grained classification branch, and the time-frequency reconstruction branch, and the overall loss function. Defined as:

[0048] ;

[0049] in, To control the temperature coefficient of gradient reversal intensity; To control the temperature coefficient of reconstruction intensity;

[0050] For the cross-entropy loss of fine-grained classification branches, For the cross-entropy loss of the coarse-grained classification branch, Let be the mean square error loss function for the time-frequency reconstruction branch;

[0051] During the testing phase, for network input samples The logit vector of all known classes output by its fine-grained classification branch. for:

[0052] ;

[0053] in, This represents the number of known fine-grained interference categories. The logit vector representing the Cth fine-grained category;

[0054] Use the maximum logit score As a criterion for decision-making in open sets:

[0055] ;

[0056] In the formula, This represents the maximum value among all known fine-grained categories of logit.

[0057] Given an open set decision threshold The final open set decision criterion is:

[0058] ;

[0059] in, This represents the final open set recognition result; when the sample has the maximum logit score. Greater than or equal to the open set decision threshold hour, Take the known category corresponding to the largest logit in the fine-grained classification branch output; otherwise, classify it as an unknown category.

[0060] Another object of the present invention is to provide an electronic device comprising:

[0061] Processor; and

[0062] The memory stores computer program instructions that, when executed by the processor, cause the processor to perform the radar active jamming open set identification method based on reconstruction-guided hierarchical adversarial learning as described above.

[0063] The present invention also provides a computer-readable storage medium having stored thereon computer program instructions, which, when executed by a processor, cause the processor to perform the radar active jamming open set identification method based on reconstruction-guided hierarchical adversarial learning as described above.

[0064] As can be seen from the above technical solution, the present invention has the following beneficial effects: First, the present invention first constructs a hierarchical labeling system based on the different interference mechanisms and time-frequency structure differences, dividing interference categories into coarse-grained labels and fine-grained labels, thereby enhancing the ability of the shared feature extraction network to express the hierarchical relationship between interference categories; Second, by introducing a hierarchical adversarial learning strategy, the shared feature extraction network can further strengthen the discriminative features of fine-grained interference categories while maintaining the consistency of coarse-grained interference features, thus improving the accuracy of identifying known interference types; Third, a reconstruction branch is introduced to reconstruct the time-frequency graph of the input network, enabling the shared feature extraction network to retain the inherent key time-frequency structure features of the interference signal while learning classification and discrimination information, thereby enhancing the separability of known and unknown classes in the feature space. Finally, combined with an open-set inference strategy, open-set inference is performed using the maximum logit score to effectively reject unknown interference samples, thereby improving the ability of the shared feature extraction network to detect and distinguish unknown active interference types in an open environment while ensuring the performance of known interference identification. Attached Figure Description

[0065] Figure 1 This is a diagram of the overall network architecture of the present invention;

[0066] Figure 2 This is a schematic diagram of the time-frequency reconstruction branch in this invention;

[0067] Figure 3 This is a schematic diagram showing the time-frequency plots of the original signal and the interference signal;

[0068] Figure 4This diagram illustrates the performance of the present invention and other published algorithms at different signal-to-noise ratios.

[0069] Figure 5 This is a schematic diagram of the confusion matrix of the present invention and other published algorithms under mixed signal-to-noise ratios;

[0070] Figure 6 This is a schematic diagram comparing the ROC curves of the present invention and the OpenMax method. Detailed Implementation

[0071] like Figure 1 As shown, a radar active jamming open set identification method based on reconstruction-guided hierarchical adversarial learning is presented, which includes the following sequential steps:

[0072] (1) Perform time-frequency transformation on the radar received signal to generate network input samples. Based on the interference mechanism and time-frequency structure differences of the interference type, input samples are provided to the network. Construct coarse-grained and fine-grained labels and combine them into training samples;

[0073] (2) The training samples are input into the shared feature extraction network to extract features. The shared feature extraction network has two outputs. One output is sent to the fine-grained classification branch and the coarse-grained classification branch respectively. The coarse-grained classification branch forms an adversarial learning process with the shared feature extraction network through the gradient inversion mechanism. The other output is sent to the time-frequency reconstruction branch to reconstruct the training samples and retain the key time-frequency structure information of the interference signal. The shared feature extraction network adopts ResNet-50.

[0074] (3) During the testing phase, the maximum logit score output by the fine-grained classification branch is used as the open set decision criterion. When the maximum logit score is greater than or equal to the open set decision threshold, the sample to be tested is determined to be the corresponding known interference category. When the maximum logit score is lower than the open set decision threshold, the sample to be tested is determined to be the unknown interference category.

[0075] Step (1) specifically refers to: collecting the radar received signal, assuming the radar received signal is... Then at any time t, we have:

[0076] ;

[0077] in, Indicates the target echo signal. Indicates an active interference signal. Indicates a noise signal;

[0078] To simultaneously characterize the structural features of the interference signal in both the time and frequency domains, the received signal... Perform a short-time Fourier transform to obtain the time-frequency representation. :

[0079] ;

[0080] in, Represents the window function. This indicates the translation position of the window function on the time axis. Represents frequency variables;

[0081] Further take The amplitude constitutes the time-frequency diagram of the interference samples. :

[0082] ;

[0083] right Normalization, cropping, and scaling are performed to obtain the network input samples. ;

[0084] Depending on the different interference mechanisms, input samples are used for each network. Simultaneously, fine-grained and coarse-grained tags are constructed, with the fine-grained tag denoted as... Coarse-grained label is denoted as ; where the superscript 'e' indicates fine-grainedness and the superscript 'c' indicates coarse-grainedness;

[0085] Coarse-grained labels are generated from fine-grained labels through a mapping function. To obtain, that is:

[0086] ;

[0087] Network input samples Fine-grained labels and coarse-grained labels Together they form the training samples, the first Each training sample is represented as By constructing the aforementioned hierarchical labeling system, the shared feature extraction network can simultaneously obtain fine-grained category supervision information and coarse-grained mechanism supervision information during training.

[0088] The coarse-grained tags include suppression interference, deception interference, and composite interference. The fine-grained tags include radio frequency noise interference (RF), frequency modulation noise interference (FM), amplitude modulation noise interference (AM), intermittent sampling and forwarding interference (ISRJ), range gate dragging interference (RGPO), range-dimensional dense decoy interference (RMT), spectrum dispersion interference (SMSP), comb spectrum interference (COMB), noise convolution interference (MNJ), comb spectrum + range-dimensional dense decoy interference (COMB_RMT), intermittent sampling and forwarding + comb spectrum interference (ISRJ_COMB), intermittent sampling and forwarding + spectrum dispersion interference (ISRJ_SMSP), noise convolution + intermittent sampling and forwarding interference (MNJ_ISRJ), range gate dragging + comb spectrum interference (RGPO_COMB), range gate dragging + spectrum dispersion interference (RGPO_SMSP), range-dimensional dense decoy + intermittent sampling and forwarding interference (RMT_ISRJ), and spectrum dispersion + comb spectrum interference (SMSP_COMB).

[0089] In step (2), the output to the fine-grained classification branch and the coarse-grained classification branch respectively specifically refers to: assuming the shared feature extraction network is... Then we have:

[0090] ;

[0091] in, These are deep feature vectors;

[0092] Fine-grained classification branch Output the logit values ​​for all known fine-grained categories. :

[0093] ;

[0094] in, Parameters representing fine-grained classification branches;

[0095] Cross-entropy loss of fine-grained classification branches for:

[0096] ;

[0097] in, Indicates network input sample Predicted probabilities on fine-grained labels; Input samples for the network Quantity; Fine-grained labeling; This represents the fine-grained classification probability of the fine-grained classification branch for the i-th sample;

[0098] coarse-grained classification branches Output the logit values ​​for all known coarse-grained categories. :

[0099] ;

[0100] in, Parameters representing coarse-grained classification branches;

[0101] Cross-entropy loss of coarse-grained classification branches Defined as:

[0102] ;

[0103] in, This represents the predicted probability on the coarse-grained label; This represents the coarse-grained classification probability of the coarse-grained classification branch for the i-th sample;

[0104] A gradient inversion layer is inserted between the shared feature extraction network and the coarse-grained classification branch. Therefore, the coarse-grained branch minimizes... To learn mechanistic semantics, the shared feature extraction network receives the inverted gradient, thus suppressing the encoding of overly dominant coarse-grained common features; correspondingly, the fine-grained classification branch learns through... Optimization is being performed, while the shared feature extraction network is in and Training is conducted under the combined effect of these factors, among which... The temperature coefficient for controlling the gradient reversal intensity. Used to learn fine-grained discriminative relationships between known interference categories. This is used to suppress the over-reliance of shared feature extraction networks on the commonalities of coarse-grained mechanisms.

[0105] like Figure 2 As shown, in step (2), the other output to the time-frequency reconstruction branch specifically refers to: assuming the last convolutional feature map of the shared feature extraction network is... First, use the bottleneck compression module. Obtaining bottleneck characteristics :

[0106] ;

[0107] Bottleneck characteristics Subsequently, through the bottleneck recovery module and decoder By gradually restoring, a reconstructed time-frequency diagram is obtained. :

[0108] ;

[0109] Mean square error loss function of time-frequency reconstruction branch Defined as:

[0110] ;

[0111] In the formula, Input samples for the network Quantity;

[0112] The time-frequency reconstruction branch includes a bottleneck compression module, a bottleneck recovery module, a decoder, and a mean square error loss function.

[0113] Step (3) specifically refers to: during the training phase, jointly optimizing the fine-grained classification branch, the coarse-grained classification branch, and the time-frequency reconstruction branch, and the overall loss function. Defined as:

[0114] ;

[0115] in, To control the temperature coefficient of gradient reversal intensity; To control the temperature coefficient of reconstruction intensity;

[0116] For the cross-entropy loss of fine-grained classification branches, For the cross-entropy loss of the coarse-grained classification branch, Let be the mean square error loss function for the time-frequency reconstruction branch;

[0117] During the testing phase, for network input samples The logit vector of all known classes output by its fine-grained classification branch. for:

[0118] ;

[0119] in, This represents the number of known fine-grained interference categories. The logit vector representing the Cth fine-grained category;

[0120] Use the maximum logit score As a criterion for decision-making in open sets:

[0121] ;

[0122] In the formula, This represents the maximum value among all known fine-grained categories of logit.

[0123] Given an open set decision threshold The final open set decision criterion is:

[0124] ;

[0125] in, This represents the final open set recognition result; when the sample has the maximum logit score. Greater than or equal to the open set decision threshold hour, Take the known category corresponding to the largest logit in the fine-grained classification branch output; otherwise, classify it as an unknown category.

[0126] like Figure 1 As shown, during training, the network is first fed with input samples. The input is a shared feature extraction network for feature extraction. Then, the extracted deep feature vectors are... A fine-grained classification branch is input for fine-class identification, and the cross-entropy loss function is used for supervised training of the fine-grained classification branch. Simultaneously, the deep feature vector... The input coarse-grained classification branch performs coarse class discrimination, and backpropagation is performed through a gradient reversal layer, thereby achieving hierarchical adversarial learning. Furthermore, the last convolutional feature map of the shared feature extraction network is used... Input time-frequency reconstruction branch, the time-frequency reconstruction branch structure is as follows Figure 2 As shown, the time-frequency graph is reconstructed from the input time-frequency graph and the reconstructed time-frequency graph is constrained by the mean square error loss function. With network input samples The differences between them are used to reconstruct guided learning.

[0127] Example 1

[0128] This invention constructs two open-set datasets for 17 types of interference. One is a mixed interference-to-noise ratio (JNR) dataset ranging from -5dB to 5dB, generating 50 time-frequency samples per class at each JNR level, divided into training, validation, and test sets in a 6:2:2 ratio. The other is an independent JNR dataset ranging from -15dB to 7dB, generating 200 samples per class at each JNR level. The parameters for the interference simulation are shown in Table 1 below. The parameters for the independent interference constituting the composite interference are also consistent with those in Table 1. The final time-frequency plots generated for each category are shown below. Figure 3 As shown, Figure 3 In the diagram, (a) LFM is the original signal, and the others are interference signals.

[0129] Table 1

[0130]

[0131] The known classes used for training and testing are RF, AM, ISRJ, COMB, RMT, ISRJ_COMB, SMSP_COMB, COMB_RMT, and ISRJ_SMSP. Unknown classes are not involved in training but are introduced only during the testing phase; these include FM, RGPO, SMSP, MNJ, RMT_ISRJ, RGPO_SMSP, RGPO_COMB, and MNJ_ISRJ. Table 2 below details the hierarchical categories and the division between known and unknown classes.

[0132] Table 2

[0133]

[0134] Training settings

[0135] The experiment was conducted on a computer equipped with a GeForce RTX 5090 graphics card, running Windows 11, and using PyTorch as the deep learning framework. The shared feature extraction network used ResNet50, with an input image size of 224×224, a batch size of 32, and a maximum training epoch of 50. The SGD optimizer was used to update parameters, with a learning rate of 0.01 and a weight decay coefficient of [missing value]. , Set to 0.1, Set it to 0.7.

[0136] Performance comparison:

[0137] This invention employs five metrics: Known Class Accuracy (KC), Unknown Class Accuracy (UC), Open Set Recognition Accuracy (OSRACC), Area Under the Receiver Operating Characteristic (AUROC) Curve, and Area Under the Precision-Recall Curve (AUPR). KC represents the classification accuracy for known class test samples, UC represents the rejection accuracy for unknown class test samples, and OSRACC represents the overall accuracy across the entire test set, where a correct score is only counted when a known sample is correctly classified or an unknown sample is correctly rejected. AUROC and AUPR are calculated by iterating through the open set decision threshold across the entire score range. To match the open set decision threshold An irrelevant approach to assess the separability between known and unknown samples.

[0138] Figure 4 The diagram shows the open set recognition accuracy (OSRACC) of different methods as a function of JNR, where the horizontal axis represents JNR and the vertical axis represents the open set recognition accuracy for all types of interference. Figure 4 In this invention, the proposed method is described. Figure 4As can be seen, the recognition accuracy of all methods generally increases with the increase of JNR, indicating that a higher JNR helps to highlight the time-frequency characteristics of the interference signal, thereby improving the recognition performance of the model. In contrast, the present invention achieves the best performance under most JNR conditions, and the overall curve changes more smoothly, showing better recognition stability. In the low JNR range, the performance differences between different methods are more significant. The present invention can still maintain a high recognition accuracy and shows strong noise resistance; in contrast, ARPL and OOD-Aug show a more significant performance decline in this range, while MOEP-AE, although close to the present invention at individual points, has greater overall fluctuations. With further increases in JNR, the performance of all methods gradually improves, but the present invention still maintains its leading position overall. These results show that the present invention not only achieves good recognition performance under medium-high JNR conditions, but also maintains strong robustness in low JNR scenarios.

[0139] from Figure 5 It can be seen that under hybrid JNR conditions, the differences between the methods mainly lie in the handling of a few confused categories and unknown category samples. The comparative methods maintain relatively high recognition rates for most known categories, but their stability in distinguishing interference types with highly similar time-frequency structures is insufficient, especially for interference categories such as MNJ and SMSP, where performance significantly degrades. For example, in MOEP-AE and OOD-Aug, MNJ is largely misclassified as COMB. In ARPL, SMSP is severely misclassified as SMSP_COMB. In contrast, this invention maintains high recognition accuracy for most known categories while significantly mitigating confusion between similar categories and providing a more sufficient rejection for unknown samples such as SMSP and MNJ. This indicates that under hybrid JNR conditions, this invention can more effectively suppress unknown categories from entering the known category space, thereby achieving more stable open set recognition performance.

[0140] Table 3 below shows the recognition results of each known class and each unknown class on different methods, and also shows the comparison of indicators when unaffected by thresholds. The method proposed in Table 3 is the present invention. Table 3 shows the recognition results of each method for the known class on the hybrid JNR dataset:

[0141] Table 3

[0142]

[0143] Table 4 shows the identification results of various methods for the unknown class on the hybrid JNR dataset. The method mentioned in Table 4 is the method of this invention.

[0144] Table 4

[0145]

[0146] Table 5 shows the recognition results of various methods on the AUROC and AUPR metrics for the hybrid JNR dataset. The method proposed in Table 5 is the present invention.

[0147] Table 5

[0148]

[0149] The experimental results show that the present invention achieves optimal performance under most JNR conditions, and the overall curve changes more smoothly, demonstrating better recognition stability. At the same time, it achieves the best performance in both AUROC and AUPR indicators, demonstrating that the superiority of the present invention does not depend on the selection of the open set decision threshold, and it can achieve good results under different open set decision threshold conditions.

[0150] This invention and the OpenMax method were compared using ROC curve analysis under mixed signal-to-noise ratio conditions. The analysis results are as follows: Figure 6 As shown, the curve of this invention is closer to the top left corner than that of OpenMax, which means that it can achieve a higher true positive rate at the same false positive rate.

[0151] The detection performance of the present invention and the initial baseline on the mixed dataset with the open set decision threshold set to 15% is shown in Table 6 below. The initial baseline is the first row, i.e., the row without a checkmark:

[0152] Table 6

[0153]

[0154] As shown in Table 6, the model performance is significantly improved after introducing hierarchical semantic modeling, especially in the identification of unknown classes. Further integration of hierarchical adversarial learning results in the highest KC value. Adding reconstruction constraints leads to the best overall open-set performance for the complete model, although KC decreases slightly. This is because while reconstruction retains more input information, it also slightly weakens the separation of known class boundaries. However, the overall results demonstrate that reconstruction-guided learning helps retain more structural information and improves the separability between known and unknown classes, thus achieving a better trade-off for open-set recognition.

[0155] Experiments demonstrate that the present invention can identify active radar interference in open-set scenarios, verifying the effectiveness of the present invention.

[0156] In summary, this invention first constructs a hierarchical labeling system based on different interference mechanisms and time-frequency structure differences, classifying interference categories into coarse-grained and fine-grained labels, thereby enhancing the ability of the shared feature extraction network to express the hierarchical relationships between interference categories. By introducing a hierarchical adversarial learning strategy, the shared feature extraction network maintains consistency in coarse-grained interference features while further strengthening the discriminative features of fine-grained interference categories, improving the accuracy of identifying known interference types. A reconstruction branch is introduced to reconstruct the time-frequency graph of the input network, enabling the shared feature extraction network to retain the inherent key time-frequency structural features of the interference signal while learning classification and discrimination information, enhancing the separability of known and unknown classes in the feature space. Finally, combined with an open-set inference strategy, open-set inference is performed using the maximum logit score to effectively reject unknown interference samples, thereby improving the shared feature extraction network's ability to detect and distinguish unknown active interference types in an open environment while ensuring the performance of known interference identification.

[0157] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention. The scope of protection claimed by the appended claims and their equivalents is defined.

Claims

1. A radar active jamming open set identification method based on reconstruction-guided hierarchical adversarial learning, characterized in that: The method includes the following steps in sequence: (1) Perform time-frequency transformation on the radar received signal to generate network input samples. Based on the interference mechanism and time-frequency structure differences of the interference type, input samples are provided to the network. Construct coarse-grained and fine-grained labels and combine them into training samples; (2) Input the training samples into the shared feature extraction network to extract features. The shared feature extraction network has two outputs. One output is sent to the fine-grained classification branch and the coarse-grained classification branch respectively. The coarse-grained classification branch forms an adversarial learning process with the shared feature extraction network through the gradient reversal mechanism. The other output is sent to the time-frequency reconstruction branch to reconstruct the training samples and retain the key time-frequency structure information of the interference signal. (3) During the testing phase, the maximum logit score output by the fine-grained classification branch is used as the open set decision criterion. When the maximum logit score is greater than or equal to the open set decision threshold, the sample to be tested is determined to be the corresponding known interference category. When the maximum logit score is lower than the open set decision threshold, the sample to be tested is determined to be the unknown interference category.

2. The radar active jamming open set identification method based on reconstruction-guided hierarchical adversarial learning according to claim 1, characterized in that: Step (1) specifically refers to: collecting the radar received signal, assuming the radar received signal is... Then at any time t, we have: ； in, Indicates the target echo signal. Indicates an active interference signal. Indicates a noise signal; To simultaneously characterize the structural features of the interference signal in both the time and frequency domains, the received signal... Perform a short-time Fourier transform to obtain the time-frequency representation. : ； in, Represents the window function. This indicates the translation position of the window function on the time axis. Represents frequency variables; Further take The amplitude constitutes the time-frequency diagram of the interference samples. : ； right Normalization, cropping, and scaling are performed to obtain the network input samples. ; Depending on the different interference mechanisms, input samples are used for each network. Simultaneously, fine-grained and coarse-grained tags are constructed, with the fine-grained tag denoted as... Coarse-grained label is denoted as ; where the superscript 'e' indicates fine-grainedness and the superscript 'c' indicates coarse-grainedness; Coarse-grained labels are generated from fine-grained labels through a mapping function. To obtain, that is: ； Network input samples Fine-grained labels and coarse-grained labels Together they form the training samples, the first Each training sample is represented as .

3. The radar active jamming open set identification method based on reconstruction-guided hierarchical adversarial learning according to claim 1, characterized in that: In step (2), the output to the fine-grained classification branch and the coarse-grained classification branch respectively specifically refers to: assuming the shared feature extraction network is... Then we have: ； in, These are deep feature vectors; Fine-grained classification branch Output the logit values ​​for all known fine-grained categories. : ； in, Parameters representing fine-grained classification branches; Cross-entropy loss of fine-grained classification branches for: ； in, Indicates network input sample Predicted probabilities on fine-grained labels; Input samples for the network Quantity; Fine-grained labeling; This represents the fine-grained classification probability of the fine-grained classification branch for the i-th sample; coarse-grained classification branches Output the logit values ​​for all known coarse-grained categories. : ； in, Parameters representing coarse-grained classification branches; Cross-entropy loss of coarse-grained classification branches Defined as: ； in, This represents the predicted probability on the coarse-grained label; This represents the coarse-grained classification probability of the coarse-grained classification branch for the i-th sample; A gradient inversion layer is inserted between the shared feature extraction network and the coarse-grained classification branch. Therefore, the coarse-grained branch minimizes... To learn mechanistic semantics, the shared feature extraction network receives the inverted gradient, thus suppressing the encoding of overly dominant coarse-grained common features; correspondingly, the fine-grained classification branch learns through... Optimization is being performed, while the shared feature extraction network is in and Training is conducted under the combined effect of these factors, among which... The temperature coefficient for controlling the gradient reversal intensity.

4. The radar active jamming open set identification method based on reconstruction-guided hierarchical adversarial learning according to claim 1, characterized in that: In step (2), the other output to the time-frequency reconstruction branch specifically refers to: assuming the last convolutional feature map of the shared feature extraction network is... First, use the bottleneck compression module. Obtaining bottleneck characteristics : ； Bottleneck characteristics Subsequently, through the bottleneck recovery module and decoder By gradually restoring, a reconstructed time-frequency diagram is obtained. : ； Mean square error loss function of time-frequency reconstruction branch Defined as: ； In the formula, Input samples for the network Quantity; The time-frequency reconstruction branch includes a bottleneck compression module, a bottleneck recovery module, a decoder, and a mean square error loss function.

5. The radar active jamming open set identification method based on reconstruction-guided hierarchical adversarial learning according to claim 1, characterized in that: Step (3) specifically refers to: during the training phase, jointly optimizing the fine-grained classification branch, the coarse-grained classification branch, and the time-frequency reconstruction branch, and the overall loss function. Defined as: ； in, To control the temperature coefficient of gradient reversal intensity; To control the temperature coefficient of reconstruction intensity; For the cross-entropy loss of fine-grained classification branches, For the cross-entropy loss of the coarse-grained classification branch, Let be the mean square error loss function for the time-frequency reconstruction branch; During the testing phase, for network input samples The logit vector of all known classes output by its fine-grained classification branch. for: ； in, This represents the number of known fine-grained interference categories. The logit vector representing the Cth fine-grained category; Use the maximum logit score As a criterion for decision-making in open sets: ； In the formula, This represents the maximum value among all known fine-grained categories of logit. Given an open set decision threshold The final open set decision criterion is: ； in, This represents the final open set recognition result; when the sample has the maximum logit score. Greater than or equal to the open set decision threshold hour, Take the known category corresponding to the largest logit in the fine-grained classification branch output; otherwise, classify it as an unknown category.

6. An electronic device, comprising: processor; as well as A memory storing computer program instructions that, when executed by the processor, cause the processor to perform the radar active jamming open set identification method based on reconfiguration-guided hierarchical adversarial learning as described in any one of claims 1-5.

7. A computer-readable storage medium having stored thereon computer program instructions, which, when executed by a processor, cause the processor to perform the radar active jamming open set identification method based on reconfiguration-guided hierarchical adversarial learning as described in any one of claims 1-5.

Images