Radar composite interference identification method, system, medium and equipment based on incremental learning

Abstract

The invention discloses a radar composite interference identification method, system, medium and device based on incremental learning, which obtains multiple radar interference time-frequency image data sets with interference type labels; constructs a basic training data set, an incremental training data set and an unbalanced test set; uses the basic training data set and the incremental training data set to perform meta-learning base class training and incremental learning on a pre-constructed direct push small sample incremental learning radar composite interference identification model; uses an unbalanced test set to perform direct push learning on the incrementally trained direct push small sample incremental learning radar composite interference identification model to obtain a target direct push small sample incremental learning radar composite interference identification model; inputs the radar interference time-frequency image to be tested into the target direct push small sample incremental learning radar composite interference identification model to perform radar composite interference identification. The invention realizes the integrated rapid and accurate identification of known and unknown composite interferences.

Description

Technical Field

[0001] The present invention relates to the technical field of radar interference identification and processing, and in particular to a radar composite interference identification method, system, medium and equipment based on incremental learning. Background Art

[0002] As an all-weather information sensing device on the modern battlefield, radar has undoubtedly become a core component for battlefield situational awareness, reconnaissance and surveillance, weapon guidance, and target identification, playing a crucial role in the military. However, the continuous development of radar jamming technology has led to the emergence of new jamming patterns, which seriously impact the detection, identification, and tracking capabilities of radar systems. Furthermore, the accurate identification of various jammers is a prerequisite for implementing effective anti-jamming measures. To address this issue, jamming identification technology has been extensively researched and has become a hot topic in the field of radar electronic countermeasures.

[0003] While deep learning-based radar jammer recognition technology has made rapid progress, its ability to continuously learn and update from valid samples remains limited, restricting its applicability in real-world scenarios. For example, due to the limited number of samples for new categories, radar jammer recognition models can easily experience catastrophic forgetting of old jammers when learning new ones, thereby reducing the algorithm's overall recognition capability. Summary of the Invention

[0004] In order to solve the technical problems existing in the background technology, the present invention proposes a radar composite interference identification method, system, medium and device based on incremental learning.

[0005] This invention proposes a radar composite interference identification method based on incremental learning, comprising:

[0006] Acquire multiple radar interference time-frequency map data sets with interference type labels; wherein the interference type labels include single interference type labels and composite interference type labels;

[0007] Based on multiple radar interference time-frequency map data sets with interference type labels, a basic training data set, an incremental training data set, and an unbalanced test set are constructed; wherein the basic training data set only includes radar interference time-frequency maps with a single interference type label, the incremental training data set only includes radar interference time-frequency maps with composite interference type labels, and the unbalanced test set includes radar interference time-frequency maps with a single interference type label and radar interference time-frequency maps with composite interference type labels;

[0008] The pre-built direct-induction small-sample incremental learning radar composite interference recognition model is trained with the basic training dataset and incremental training dataset respectively, and the incrementally trained direct-induction small-sample incremental learning radar composite interference recognition model is obtained.

[0009] By using an imbalanced test set to perform push-based learning on the incrementally trained push-based small-sample incremental learning radar composite interference identification model, a target push-based small-sample incremental learning radar composite interference identification model is obtained.

[0010] The time-frequency image of the radar interference to be tested is input into the target direct-pushing small-sample incremental learning radar composite interference recognition model to obtain the predicted recognition result of the time-frequency image of the radar interference to be tested.

[0011] Preferably, the push-type small-sample incremental learning radar composite interference identification model includes a basic interference identification network model based on meta-learning, an interference identification network model based on incremental learning, and an auxiliary interference identification network model based on push-type learning; wherein, the basic interference identification network model based on meta-learning includes a feature extractor based on a CNN network structure, a fully connected layer, and a dynamically growing EM memory.

[0012] Preferably, the pre-built transductive few-sample incremental learning radar composite interference identification model is trained on base classes and incrementally learned using the basic training dataset and the incremental training dataset, respectively, specifically including:

[0013] In the basic training phase, the basic interference recognition network model based on meta-learning is trained and learned using the basic training dataset to obtain the basic interference recognition network model based on meta-learning after basic training.

[0014] In the incremental training phase, the feature extraction network of the basic interference recognition network model based on meta-learning after basic training is used as the feature extraction network of the interference recognition network model based on incremental learning; the interference recognition network model based on incremental learning is incrementally trained using the incremental training dataset to obtain the interference recognition network model based on incremental learning after incremental training.

[0015] Preferably, the direct-learning small-sample incremental learning radar composite interference identification model is trained using a basic training dataset to obtain a basic interference identification network model based on meta-learning after basic training, specifically including:

[0016] Based on the meta-learning model, the feature extractor and fully connected layer are trained on the basic training dataset using cosine similarity to obtain the basic training meta-learning-based interference recognition network model and multiple quasi-orthogonal prototype vectors. The multiple quasi-orthogonal prototype vectors and their corresponding interference type labels are stored in dynamic memory.

[0017] Preferably, the formula for calculating cosine similarity is:

[0018] l i =cos(tanh(g) θ2 (fθ1 (x))),(tanh(P i )));

[0019] In the formula, tanh() is the hyperbolic tangent function, cos() is the cosine similarity, and P i Let x be a quasi-orthogonal prototype vector, and let x be an input sample.

[0020] Preferably, the incremental learning-based interference recognition network model is incrementally trained using the incremental training dataset to obtain the incrementally trained interference recognition network model, specifically including:

[0021] The incremental training dataset is divided into multiple incremental session datasets according to the incremental learning mode, and incremental learning is performed on multiple incremental session datasets in sequence.

[0022] The first incremental learning process is as follows:

[0023] The first incremental session dataset is input into the basic trained incremental learning-based interference recognition network model, and the fully connected layers in the basic trained incremental learning-based interference recognition network model are retrained using the incremental training loss.

[0024] After learning is complete, the quasi-orthogonal prototype vectors of the new class and the old class are recalibrated using a dual-polarization strategy to obtain a brand-new quasi-orthogonal prototype vector after the first incremental session, which is then stored in dynamic memory.

[0025] By repeating the above steps until the last incremental session dataset, we obtain the trained final quasi-orthogonal prototype vectors and the interference recognition network model based on incremental learning.

[0026] Preferably, the formula for calculating the incremental training loss is:

[0027]

[0028] In the formula, A is the average activation, β is the update rate, and K is the value of K. * θ1 is the bipolarization prototype vector, and θ2 is the parameter of the fully connected layer.

[0029] Preferably, the calculation formula for the dual-polarization strategy is as follows:

[0030] K * =sign(P);

[0031] In the formula, K * Let P be the bipolar prototype vector, and let P be the quasi-orthogonal prototype vector.

[0032] Preferably, the incrementally trained interference recognition network model is subjected to transductive learning using an imbalanced test set to obtain a target transductive small-sample incremental learning radar composite interference recognition model, specifically including:

[0033] The auxiliary interference identification network model based on transductive learning is transductively learned on an imbalanced test set. The parameters of the fully connected layer are updated based on the transductive training loss. The quasi-orthogonal prototype vectors are further reinforced until the loss of the auxiliary interference identification network model based on transductive learning converges, thus obtaining the target transductive small sample incremental learning radar composite interference identification model.

[0034] Preferably, the formula for calculating the direct-push training loss is:

[0035]

[0036] Where CE is the standard cross entropy loss, I KL For mutual information based on KL divergence, I α For mutual information based on α divergence, X Q is the interference sample in the input unbalanced test set, Y Q The output is the predicted label, λ is the parameter, and p ik is the posterior distribution of the label, p k is the marginal distribution of labels.

[0037] Preferably, the single interference type label includes range deception interference, velocity deception interference, intermittent sampling direct interference, intermittent sampling repetitive interference, intermittent sampling cyclic interference, and dense false target interference;

[0038] The composite jamming type labels include range deception and intermittent sampling repeated composite jamming, range deception and intermittent sampling direct composite jamming, velocity deception and intermittent sampling direct composite jamming, dense false targets and intermittent sampling direct composite jamming, and range deception and velocity deception composite jamming.

[0039] This invention also proposes a radar composite interference identification system based on incremental learning, comprising:

[0040] The acquisition module is used to acquire multiple radar interference time-frequency map datasets labeled with interference types;

[0041] The training set construction module is used to construct a basic training dataset, an incremental training dataset, and an imbalanced test set based on multiple radar jamming time-frequency map datasets labeled with jamming types.

[0042] The processing module pre-builds a direct-induction small-sample incremental learning radar composite interference recognition model;

[0043] It is also used to perform base class training and incremental learning on the direct-push small-sample incremental learning radar composite interference identification model using the basic training dataset and the incremental training dataset, respectively, to obtain the incrementally trained direct-push small-sample incremental learning radar composite interference identification model.

[0044] It is also used to perform push-through learning on the incrementally trained push-through small sample incremental learning radar composite interference identification model using an unbalanced test set, so as to obtain the target push-through small sample incremental learning radar composite interference identification model.

[0045] The identification module is used to input the time-frequency image of the radar interference to be tested into the target direct-pull small sample incremental learning radar composite interference identification model to obtain the prediction and identification results of the time-frequency image of the radar interference to be tested.

[0046] The present invention also proposes an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the radar composite interference identification method based on incremental learning as described in any of the above.

[0047] The present invention also proposes a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the radar composite interference identification method based on incremental learning as described in any of the above.

[0048] This invention proposes a radar composite interference identification method, system, medium, and device based on incremental learning, constructing a transductive few-shot incremental learning radar composite interference identification model (hereinafter referred to as the CTFSCIL model) based on an incremental learning framework. First, the transductive few-shot incremental learning radar composite interference identification model is trained on base classes and incrementally learned using a basic training dataset and an incremental training dataset, respectively, resulting in an incrementally trained transductive few-shot incremental learning radar composite interference identification model. This model enables rapid and accurate identification of new types of interference in special scenarios while simultaneously ensuring accurate identification of known interference. Finally, addressing the issue of incomplete model classifier training during the incremental learning stage, the transductive learning method is combined to efficiently utilize interference samples from the imbalanced test set. During the transductive learning stage, the imbalanced test set is used to transductively learn the identification model, fully utilizing all effective information from the interference samples in the imbalanced test set, ultimately achieving integrated, rapid, and accurate identification of known and unknown composite interference. Attached Figure Description

[0049] Figure 1 This is a flowchart illustrating a radar composite interference identification method based on incremental learning in one embodiment of the present invention.

[0050] Figure 2This is a schematic diagram of the framework of a radar composite interference identification method based on incremental learning in one embodiment of the present invention.

[0051] Figure 3 This is a schematic diagram of the basic interference identification network based on meta-learning in one embodiment of the present invention.

[0052] Figure 4 This is a schematic diagram of the feature extractor in one embodiment of the present invention.

[0053] Figure 5 It is a sample randomly selected from each interference category label in a -5dB scenario in an embodiment of the present invention.

[0054] Figure 6 The graph shows the recognition accuracy curves of existing methods and the method described in this invention in 1-way 5-shot incremental mode under the -5dB scenario.

[0055] Figure 7 This is a visualization of the T-SNE classification features of the method described in this invention during the incremental learning process in a -5dB scenario.

[0056] Figure 8 This is a visualization of the classification confusion matrix of the method described in this invention during the incremental learning process in a -5dB scenario. Detailed Implementation

[0057] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0058] Reference Figure 1 The present invention proposes a radar composite interference identification method based on incremental learning, comprising:

[0059] Acquire multiple radar interference time-frequency map data sets with interference type labels; wherein the interference type labels include single interference type labels and composite interference type labels;

[0060] Based on multiple radar interference time-frequency map data sets with interference type labels, a basic training data set, an incremental training data set, and an unbalanced test set are constructed; wherein the basic training data set only includes radar interference time-frequency maps with a single interference type label, the incremental training data set only includes radar interference time-frequency maps with composite interference type labels, and the unbalanced test set includes radar interference time-frequency maps with a single interference type label and radar interference time-frequency maps with composite interference type labels;

[0061] The pre-built direct-induction small-sample incremental learning radar composite interference recognition model is trained with the basic training dataset and incremental training dataset respectively, and the incrementally trained direct-induction small-sample incremental learning radar composite interference recognition model is obtained.

[0062] By using an imbalanced test set to perform push-based learning on the incrementally trained push-based small-sample incremental learning radar composite interference identification model, a target push-based small-sample incremental learning radar composite interference identification model is obtained.

[0063] The time-frequency image of the radar interference to be tested is input into the target direct-pushing small-sample incremental learning radar composite interference recognition model to obtain the predicted recognition result of the time-frequency image of the radar interference to be tested.

[0064] The present invention proposes a radar composite interference recognition method based on incremental learning, and constructs a direct-pushing small-sample incremental learning radar composite interference recognition model based on an incremental learning framework (hereinafter referred to as the CTFSCIL model). First, the direct-pushing small-sample incremental learning radar composite interference recognition model is respectively subjected to base class training and incremental learning using a basic training data set and an incremental training data set, thereby obtaining a direct-pushing small-sample incremental learning radar composite interference recognition model after incremental training, which realizes rapid and accurate recognition of new categories of interference in special scenarios and simultaneously ensures accurate recognition of known interference; finally, in order to address the problem of incomplete training of the model classifier in the incremental learning stage, the direct-pushing learning method is combined to efficiently utilize the interference samples of the unbalanced test set. In the direct-pushing learning stage, the unbalanced test set is used to perform direct-pushing learning on the recognition model using the direct-pushing learning method, making full use of all the effective information of the interference samples in the unbalanced test set, and ultimately achieving integrated rapid and accurate recognition of known and unknown composite interferences.

[0065] Specifically, multiple radar interference time-frequency map datasets labeled with interference type are as follows: The input data is x i and the corresponding true value label y i Among them, label y i Different training sets are mutually exclusive.

[0066] The first radar jamming time-frequency map dataset D 1 The basic session dataset provides a large-scale training sample of single radar jamming. The remaining datasets are incremental session datasets, containing small samples of new categories of radar composite jamming, i.e., N categories, with K samples in each category, referred to as N-way K-shot.

[0067] The overall framework diagram of the radar compound interference identification method based on incremental learning is as follows: Figure 2 shown.

[0068] In this embodiment, the direct-induction small-sample incremental learning radar composite interference recognition model includes a basic interference recognition network model based on meta-learning, an interference recognition network model based on incremental learning, and an auxiliary interference recognition network model based on direct-induction learning.

[0069] Specifically, the CTFSCIL model combines incremental learning, few-shot learning, and transductive learning to form a novel integrated intelligent network model for radar active jamming identification. On the one hand, this model can update the pre-trained radar active jamming identification model using only a very small number of new category composite jamming samples, greatly overcoming the catastrophic forgetting problem and achieving rapid identification of new category composite jamming under small-shot conditions. On the other hand, in the transductive learning stage, the model uses transductive learning to efficiently utilize the effective information of the imbalanced test set, further improving the overall performance of radar active jamming identification.

[0070] The construction of the CTFSCIL model consists of three stages. The first stage is the basic training stage, which constructs a basic interference recognition network model based on meta-learning. Firstly, the basic network model structure of the CTFSCIL model is proposed, which consists of a freezeable feature extractor, a trainable fixed-size fully connected layer, and a rewriteable dynamically growing EM memory. The freeze part is separated from the incremental part by inserting a fully connected layer, which outputs class vectors in a hyperdimensional embedding space with a fixed dimension. All three stages of the CTFSCIL model will adopt this basic network model structure, such as... Figure 3 shown.

[0071] In the basic training phase, the feature extractor is a CNN network model that performs meta-learning through appropriate sharpened attention. It strives to represent different interference categories with quasi-orthogonal vectors to achieve accurate recognition of basic known single-category interference, and also serves as a pre-training model in the incremental training phase.

[0072] The second stage is the incremental training stage, which constructs an interference recognition network model based on incremental learning. Based on the pre-trained model from the first stage, a very small number of new category composite interference samples are input into the network model, forming new category quasi-orthogonal prototype vectors in a dynamically growing EM memory. Then, the quasi-orthogonal prototype vectors stored in the dynamically growing EM memory are bipolarized, and the fully connected layers are retrained iteratively to obtain a classifier model suitable for all current interference types. This achieves the effect of incremental learning, rapidly updating the model while efficiently identifying both new and old radar interference.

[0073] The third stage is the transductive learning stage, which constructs an auxiliary interference recognition network model based on transductive learning. Building upon the model trained in the second stage, interference samples from the imbalanced test set are used as the query set, and the training set from the second stage is used as the support set. The transductive learning model performs transductive learning on the class quasi-orthogonal prototype vectors stored in the dynamically growing EM memory, resulting in a better classifier model. After training, testing is conducted to effectively evaluate the recognition performance of the entire network model. The third stage primarily aims to fully utilize the interference sample information from the imbalanced test set to further improve the interference recognition performance of the CTFSCIL model.

[0074] In this embodiment, the basic interference identification network model based on meta-learning includes a feature extractor based on a CNN network structure, fully connected layers, and a dynamically growing EM memory. The dynamically growing EM memory stores quasi-orthogonal prototype vectors equal to the number of interference categories encountered so far. The feature extractor in this embodiment maps the input interference training samples to a feature space, where θ1 is a learnable parameter of the feature extractor.

[0075] For the feature extractor, a five-layer CNN structure is proposed to map the input image to the feature space, such as... Figure 4 As shown. To form an embedded network model with a hyperdimensional distributed representation, the feature extractor is connected to a fully connected layer, where θ2 is a trainable parameter of the fully connected layer.

[0076] The fully connected layer generates a support vector for each training input. These support vectors are combined to compute a set of d-dimensional quasi-orthogonal prototype vectors, which are stored in a dynamically growing EM memory. In addition to storing the quasi-orthogonal prototype vectors, the dynamically growing EM memory also contains a single-hot value store, and this is collectively referred to as the key-value store. The quasi-orthogonal prototype vectors stored in the dynamically growing EM memory are mapped to different perturbation categories using cosine similarity.

[0077] In a further embodiment, the push-type small-sample incremental learning radar composite interference identification model also includes a GAA memory.

[0078] In this embodiment, the pre-built transductive few-shot incremental learning radar composite interference identification model is trained on base classes and incrementally learned using a basic training dataset and an incremental training dataset, respectively. Specifically, this includes:

[0079] In the basic training phase, the basic interference recognition network model based on meta-learning is trained and learned using the basic training dataset to obtain the basic interference recognition network model based on meta-learning after basic training.

[0080] In the incremental training phase, the feature extraction network of the basic interference recognition network model based on meta-learning after basic training is used as the feature extraction network of the interference recognition network model based on incremental learning. The interference recognition network model based on incremental learning is incrementally trained using the incremental training dataset to obtain the interference recognition network model after incremental training.

[0081] This setup, based on a meta-learning training approach, leverages its ability to quickly adapt to new samples and its strong generalization capabilities to construct a meta-learning-based network model for identifying known radar interference. This results in a highly accurate model for identifying known single interference, which serves as a pre-training model for subsequent incremental learning to identify unknown combined interference. This incremental learning network model is then constructed to enable rapid and accurate identification of new types of combined interference in special scenarios while also ensuring accurate identification of known single interference.

[0082] In a further embodiment, the direct-learning small-sample incremental learning radar composite interference identification model is trained using a basic training dataset to obtain a basic interference identification network model based on meta-learning after basic training, specifically including:

[0083] Based on the meta-learning model, the feature extractor and fully connected layer are trained on the basic training dataset using cosine similarity to obtain the basic training meta-learning-based interference recognition network model and multiple quasi-orthogonal prototype vectors. The quasi-orthogonal prototype vectors and their corresponding interference type labels are stored in dynamic memory.

[0084] In one embodiment, the cosine similarity is calculated as follows:

[0085] l i =cos(tanh(g) θ2 (f θ1 (x))),(tanh(P i )));

[0086] In the formula, tanh() is the hyperbolic tangent function, cos() is the cosine similarity, and P i Let x be a quasi-orthogonal prototype vector, and let x be an input sample.

[0087] Since hyperbolic tangent has proven to be a useful nonlinearity in high-dimensional artificial neural networks, it constrains the norm of both the activation prototype and the embedded output. Furthermore, cosine similarity addresses the norm and bias problems frequently encountered in few-shot incremental learning by focusing on the angle between the activation prototype and the embedded output while ignoring their norms. Overall, this design forms a content-based attention mechanism.

[0088] This network model is trained through meta-learning by solving various task problem sets, which gradually improve the quality of the mapping. This is achieved by progressively meta-learning to assign nearly orthogonal vectors to different interference classes and mapping them away from each other in a hyperdimensional space. For newly encountered interference classes, this space will not lack such vectors. To improve inter-class separation, the basic interference recognition network model is built on a meta-learning setup.

[0089] In a further embodiment, the incremental learning-based interference recognition network model is incrementally trained using an incremental training dataset to obtain an incrementally trained interference recognition network model, specifically including:

[0090] The incremental training dataset is divided into multiple incremental session datasets according to the incremental learning mode, and incremental learning is performed on multiple incremental session datasets in sequence.

[0091] The first incremental learning process is as follows:

[0092] The first incremental session dataset is input into the basic trained incremental learning-based interference recognition network model, and the fully connected layers in the basic trained incremental learning-based interference recognition network model are retrained using the incremental training loss.

[0093] After learning is complete, the quasi-orthogonal prototype vectors of the new class and the old class are recalibrated using a dual-polarization strategy to obtain a brand-new quasi-orthogonal prototype vector after the first incremental session, which is then stored in dynamic memory.

[0094] By repeating the above steps until the last incremental session dataset, we obtain the trained final quasi-orthogonal prototype vectors and the interference recognition network model based on incremental learning.

[0095] Specifically, in the incremental training phase, based on the basic interference recognition network, the incrementally learned interference recognition network in the CTFSCIL model does not need to be trained in a centralized manner for representations of all interference classes. The proposed incrementally learned interference recognition network relies on a hyperdimensional embedding space and uses an incremental learning mode to incrementally create the most separable class vectors. In the incremental mode, the average prototype in the dynamically growing EM memory is bipolarized to obtain better quasi-orthogonality, and then the fully connected layers are retrained to enable them to recognize new interference classes. At the same time, the weights in the feature extractor after meta-learning in the basic training phase are frozen to ensure the model's recognition performance for old interference classes and overcome the catastrophic forgetting problem.

[0096] As the number of new perturbation categories increases, so does the number of prototypes. Due to insufficient separability between classes, they cannot distinguish between different classes. To address this, the prototypes and fully connected layers were tuned to better guide the activations provided by the frozen feature extractor. A two-stage strategy was followed: first, the prototypes were tuned, and then the fully connected layers were retrained to align the average activations with the newly tuned prototypes.

[0097] First, attempt to create separation between nearby prototype pairs, which in the best case produces near-zero cross-correlation between prototype pairs. A computationally simple and efficient option is to add some kind of noise to the prototypes, such as quantization noise. This is achieved by quantizing the prototypes into bipolar vectors using sign operations, defined as:

[0098] K * =sign(P);

[0099] In the formula, K * Let P be the bipolar prototype vector, and let P be the quasi-orthogonal prototype vector.

[0100] Next, the embedding network must be retrained to align its output with the bipolar prototype. Instead of optimizing every training sample, the available global average activations in memory are aligned with the bipolar prototype. The task of the last fully connected layer of the embedding network is to map local features from the feature extractor to the distributed representation. The incremental phase only updates the parameters θ2 of the fully connected layer, while the parameters θ1 of the feature extractor remain frozen during retraining.

[0101] The retraining fine-tuning of fully connected layers is based on iterative updates using a loss function that aims to maximize the similarity between the average activation and the bipolar prototype. The incremental training loss function is as follows:

[0102]

[0103]

[0104] In the formula, A is the average activation, β is the update rate, and K is the value of K. * θ1 is the bipolarization prototype vector, and θ2 is the parameter of the fully connected layer.

[0105] After T iterations of parameter updates, the final prototype P is determined by passing the global average activation value through a fully connected layer. * The fully connected layers have been fine-tuned on the quasi-orthogonal dual-polarization prototypes, so the final prototypes also tend to be quasi-orthogonal, thus obtaining a classifier model that can simultaneously identify new and old interference.

[0106] In this embodiment, an unbalanced test set is used to perform direct learning on the incrementally trained interference recognition network model based on incremental learning to obtain a target direct learning small sample incremental learning radar composite interference recognition model, which specifically includes:

[0107] The auxiliary interference recognition network model based on transductive learning is transductively learned on an unbalanced test set. The parameters of the fully connected layer are updated based on the transductive training loss. The quasi-orthogonal prototype vector is further reinforced until the loss of the auxiliary interference recognition network model based on transductive learning converges, and the target transductive small-sample incremental learning radar composite interference recognition model is obtained.

[0108] Specifically, since the incremental learning stage uses a small sample space training set, there is insufficient training, resulting in the obtained classifier model not having the optimal performance and failing to achieve a high interference recognition accuracy.

[0109] To further improve the interference recognition performance of the model, an auxiliary interference recognition network based on transductive learning was constructed and placed within the entire CTFSCIL model during the testing phase. This auxiliary interference recognition network, based on transductive learning, was then used to train the interference recognition network by incorporating transductive information maximization. Combined with standard cross-entropy loss based on the imbalanced test set, the mutual information between the features of the imbalanced test set and their predicted labels was maximized, while minimizing the cross-entropy loss on the imbalanced test set.

[0110] In the direct learning phase, after samples from the unbalanced test set are input, direct learning is performed based on the model obtained in the incremental training phase. The parameters θ2 of the fully connected layer are further updated using samples from the unbalanced test set to obtain a quasi-orthogonal dual-polarization prototype with more complete separation. This further strengthens the trained classifier and improves the overall interference recognition performance of the CTFSCIL model.

[0111] To this end, the following two loss functions are designed:

[0112]

[0113] Where CE is the standard cross entropy loss, I KL For mutual information based on KL divergence, I α For mutual information based on α divergence, X Q is the sample in the unbalanced test set, Y Q The output is the predicted label, λ is the parameter, and p ik is the posterior distribution of the label, p k is the marginal distribution of labels.

[0114] The two different loss functions described above are respectively applicable to cases where the samples in the imbalanced test set are evenly distributed and unevenly distributed. Mutual information based on KL divergence is applicable when the interference samples are evenly distributed, while mutual information based on α divergence is applicable when the interference samples are unevenly distributed. In actual radar active jamming identification scenarios, the interference data obtained is inevitably unevenly distributed, with different numbers of samples for each type of jammer. Therefore, based on practical applications, the identification scenario of uneven sample distribution during the test phase was considered.

[0115] In this embodiment, the single interference type labels include Distance Deception Jamming (DDJ), Velocity Deception Jamming (VDJ), Intermittent Sampling Direct Jamming (ISDJ), Intermittent Sampling Repeater Jamming (ISRJ), and Dense False Target Jamming (DFTJ).

[0116] The composite jamming type labels include range spoofing and intermittent sampling repeated composite jamming (DDJ+ISRJ), range spoofing and intermittent sampling direct composite jamming (DDJ+ISDJ), velocity spoofing and intermittent sampling direct composite jamming (VDJ+ISDJ), dense false targets and intermittent sampling direct composite jamming (DFTJ+ISDJ), and range spoofing and velocity spoofing composite jamming (DDJ+VDJ).

[0117] This invention also proposes a radar composite interference identification system based on incremental learning, comprising:

[0118] The acquisition module is used to acquire multiple radar interference time-frequency map datasets labeled with interference types;

[0119] The training set construction module is used to construct a basic training dataset, an incremental training dataset, and an imbalanced test set based on multiple radar jamming time-frequency map datasets labeled with jamming types.

[0120] The processing module pre-builds a direct-induction small-sample incremental learning radar composite interference recognition model;

[0121] It is also used to perform base class training and incremental learning on the direct-push small-sample incremental learning radar composite interference identification model using the basic training dataset and the incremental training dataset, respectively, to obtain the incrementally trained direct-push small-sample incremental learning radar composite interference identification model.

[0122] It is also used to perform push-through learning on the incrementally trained push-through small sample incremental learning radar composite interference identification model using an unbalanced test set, so as to obtain the target push-through small sample incremental learning radar composite interference identification model.

[0123] The identification module is used to input the time-frequency image of the radar interference to be tested into the target direct-pull small sample incremental learning radar composite interference identification model to obtain the prediction and identification results of the time-frequency image of the radar interference to be tested.

[0124] The present invention also proposes an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the radar composite interference identification method based on incremental learning as described in any of the above.

[0125] The present invention also proposes a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the radar composite interference identification method based on incremental learning as described in any of the above.

[0126] The effects of the present invention can be further illustrated by the following simulation.

[0127] 1. Simulation conditions

[0128] The parameters of the test hardware platform are shown in Table 1, and the parameters of the software platform are shown in Table 2.

[0129] Table 1 Hardware Platform Parameters

[0130] CPU Intel i7-12700F Memory 32GB Graphics card model NVIDIA GeForce RTX 3060 Graphics card memory 12GB

[0131] Table 2 Software Platform Parameters

[0132] operating system Windows 10 64-bit Compiler PyCharm 2023.1 CUDA version 12.1

[0133] In this embodiment, a lightweight CNN is used as the feature extractor. During the basic training phase, the model is trained for 100 epochs using a stochastic gradient descent optimizer (SGD) with momentum of 0.9 and weight decay of 0.0005, with a batch size of 15. The dimension of the fully connected layer is 512. The learning rate is initially set to 0.01 and reduced by a factor of 10 at iterations 30 and 60. In incremental mode, the fully connected layer is retrained for 10 iterations with an update rate of 0.01.

[0134] All input images are normalized and cropped to 128×128 pixels. In terms of data augmentation, all training images are randomly rotated by small angles to increase their diversity.

[0135] 2. Simulation Method

[0136] (1) The fine-tuned convolutional neural network (Ft-CNN) method in the prior art;

[0137] (2) Incremental classifiers and representation learning (ICaRL) methods in the prior art;

[0138] (3) The existing joint convolutional neural network (Joint_CNN) method;

[0139] (4) The method described in the present invention is the composite interference identification for radar based on incremental learning (CTFSCIL) method.

[0140] 3. Simulation content and simulation results

[0141] The dataset configuration used in the simulation experiments is shown in Table 3. Both the method of this invention and existing methods use a lightweight CNN as a feature extractor. In the basic training phase, the model is trained for 100 epochs using a stochastic gradient descent optimizer (SGD) with momentum of 0.9 and weight decay of 0.0005, with a batch size of 15. The dimension of the fully connected layer is 512. The learning rate is initially set to 0.01 and reduced by a factor of 10 at iterations 30 and 60. In incremental mode, the fully connected layer is retrained for 10 iterations with an update rate of 0.01.

[0142] All input images are normalized and cropped to 128 × 128 pixels. For data augmentation, all training images are randomly rotated by small angles to increase their diversity.

[0143] Table 3 Incremental Configuration of Data Base

[0144]

[0145] To verify the performance of the proposed CTFSCIL model, a 1-way 5-shot incremental mode was adopted in a -5dB scenario, dividing the entire training process into 1 basic session, 5 incremental sessions, and a total of 6 training sessions.

[0146] This embodiment uses average accuracy to evaluate the performance of the method. Incremental classification accuracy is a common metric in incremental learning, used to judge the performance of the current model in classifying and recognizing all encountered categories. Furthermore, the incremental learning process is visualized using T-SNE and a confusion matrix.

[0147] The recognition accuracy curves of the CTFSCIL model and the existing Ft-CNN, ICaRL, and Joint_CNN radar interference recognition models in the incremental process are shown in the following figure: Figure 6 As shown in the figure. The T-SNE feature visualization diagram after the present invention has undergone 5 incremental learning cycles to reach 10 types of radar interference is shown in the figure. Figure 7As shown in the figure, the confusion matrix visualization diagram of the present invention after 5 incremental learning to 10 types of radar interference is as follows Figure 8 shown.

[0148] like Figure 6 The experiment simulates six incremental stages. Initially, there are five types of single radar interference. Each incremental batch adds one type of composite radar interference, with five samples for each new composite interference type. After five incremental learning iterations, the model can identify all ten types of radar interference. It is evident that, compared to traditional interference identification networks, the CTFSCIL model can quickly identify new interference under small sample conditions. For known basic interference, the accuracy rate is above 98%, and after five incremental iterations, the overall average accuracy stabilizes above 80%, representing a performance improvement of over 5% compared to similar classic incremental learning models. In other words, the CTFSCIL model achieves optimal interference identification accuracy and optimal overall identification performance on the constructed simulated radar interference time-frequency map dataset.

[0149] from Figure 7 and Figure 8 As can be seen, the method of the present invention has significant advantages over various benchmark methods, particularly its ability to quickly and accurately identify unknown composite interference in open dynamic scenarios.

[0150] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A radar composite interference identification method based on incremental learning, characterized in that: include: Acquire multiple radar interference time-frequency diagram data sets with interference type labels; wherein the interference type labels include single interference type labels and composite interference type labels; According to multiple radar interference time-frequency graph data sets with interference type labels, a basic training data set, an incremental training data set and an unbalanced test set are constructed; wherein the basic training data set only includes radar interference time-frequency graphs with a single interference type label, the incremental training data set only includes radar interference time-frequency graphs with a composite interference type label, and the unbalanced test set includes radar interference time-frequency graphs with a single interference type label and radar interference time-frequency graphs with a composite interference type label; The pre-built direct-induction small-sample incremental learning radar composite interference recognition model is trained with the basic training data set and the incremental training data set, respectively, to obtain the direct-induction small-sample incremental learning radar composite interference recognition model after incremental training. The non-balanced test set is used to perform direct learning on the incrementally trained direct-pushing small-sample incremental learning radar composite interference recognition model to obtain the target direct-pushing small-sample incremental learning radar composite interference recognition model. The time-frequency image of the radar interference to be tested is input into the target direct-pushing small sample incremental learning radar composite interference recognition model to obtain the prediction recognition result of the time-frequency image of the radar interference to be tested; Among them, the direct-pushing small sample incremental learning radar composite interference recognition model includes a basic interference recognition network model based on meta-learning, an interference recognition network model based on incremental learning, and an auxiliary interference recognition network model based on direct-pushing learning; The basic network model of the direct-pushing small-sample incremental learning radar composite interference recognition model consists of a freezable feature extractor, a trainable fixed-size fully connected layer, and a rewritable dynamically growing EM memory; the frozen part is separated from the incremental part by inserting a fully connected layer, which outputs a class vector in a hyperdimensional embedding space with a fixed dimension. Among them, the construction of the direct-pushing small sample incremental learning radar composite interference recognition model includes three stages, all of which adopt the basic network model structure; The first stage is the basic training stage, in which a basic interference recognition network model based on meta-learning is constructed. In the basic training stage, the feature extractor is a CNN network model that performs meta-learning through appropriate sharpened attention, and uses quasi-orthogonal vectors to represent different interference categories, which is used to achieve accurate recognition of basic known single-category interference and serves as a pre-training model for the incremental training stage. The second stage is the incremental training stage, which builds an interference recognition network model based on incremental learning. In the second stage, based on the pre-trained model in the first stage, a very small number of new category composite interference samples are input into the network model to form new category quasi-orthogonal prototype vectors in the dynamically growing EM memory. Then, the quasi-orthogonal prototype vectors stored in the dynamically growing EM memory are bipolarized, and the fully connected layer is retrained in an iterative manner to obtain a classifier model suitable for all current interference types. The third stage is the direct learning stage, in which an auxiliary interference recognition network model based on direct learning is constructed; in the third stage, based on the model trained in the second stage, the interference samples of the unbalanced test set are used as the query set, and the training set of the second stage is used as the support set. The direct learning model is used to perform direct learning on the category quasi-orthogonal prototype vectors stored in the dynamically growing EM memory to obtain a better classifier model.

2. The radar composite interference identification method based on incremental learning according to claim 1 is characterized in that: The pre-built direct push small sample incremental learning radar composite interference recognition model is trained with the basic training data set and the incremental training data set, respectively, for base class training and incremental learning, including: In the basic training stage, the basic interference recognition network model based on meta-learning is trained and learned using the basic training data set to obtain the basic interference recognition network model based on meta-learning after basic training; In the incremental training stage, the feature extraction network of the basic interference recognition network model based on meta-learning after basic training is used as the feature extraction network of the interference recognition network model based on incremental learning; the interference recognition network model based on incremental learning is incrementally learned using the incremental training data set to obtain the interference recognition network model based on incremental learning after incremental training.

3. The radar composite interference identification method based on incremental learning according to claim 2 is characterized in that: The basic training data set is used to train the direct-pushing small sample incremental learning radar composite interference recognition model, and the basic interference recognition network model based on meta-learning after basic training is obtained, which specifically includes: Based on the meta-learning mode, the feature extractor and the fully connected layer are trained on the basic training data set by cosine similarity to obtain the basic interference recognition network model based on meta-learning and multiple quasi-orthogonal prototype vectors after basic training, and the multiple quasi-orthogonal prototype vectors and the corresponding interference type labels are stored in the dynamic memory.

4. The radar composite interference identification method based on incremental learning according to claim 3 is characterized in that: The calculation formula for cosine similarity is: l i =cos(tanh(g θ2 (f θ1 (x))),(tanh(P i ))); In the formula, tanh() is the hyperbolic tangent function, cos() is the cosine similarity, and P i is the quasi-orthogonal prototype vector, and x is the input sample.

5. The radar composite interference identification method based on incremental learning according to claim 2 is characterized in that: The incremental training data set is used to perform incremental learning on the interference recognition network model based on incremental learning, and the interference recognition network model based on incremental learning after incremental training is obtained, which specifically includes: Dividing the incremental training data set into multiple incremental session data sets according to the incremental learning mode, and performing incremental learning on the multiple incremental session data sets in sequence; The first incremental learning process is: Input the first incremental session data set into the incremental learning-based interference recognition network model after basic training, and retrain the fully connected layer in the incremental learning-based interference recognition network model after basic training through the incremental training loss; After learning is completed, the quasi-orthogonal prototype vectors of the new class and the old class are recalibrated through a dual-polarization strategy to obtain a completely new quasi-orthogonal prototype vector after the first incremental session, and stored in a dynamic memory; By repeating the above steps until the last incremental session data set, the trained final quasi-orthogonal prototype vector and the interference recognition network model based on incremental learning are obtained.

6. The radar composite interference identification method based on incremental learning according to claim 5 is characterized in that: The calculation formula for incremental training loss is Where A is the average activation, β is the update rate, and K * is the dual-polarization prototype vector, θ2 is the parameter of the fully connected layer; The calculation formula of the dual-polarization strategy is: K * =sign(P); In the formula, K * is the dual-polarization prototype vector, and P is the quasi-orthogonal prototype vector.

7. The radar composite interference identification method based on incremental learning according to claim 1 is characterized in that: The non-balanced test set is used to perform direct learning on the interference recognition network model based on incremental learning after incremental training, and a target direct learning small sample incremental learning radar composite interference recognition model is obtained, which specifically includes: The auxiliary interference identification network model based on transductive learning is transductively learned on the unbalanced test set. The parameters of the training fully connected layer are updated based on the transductive training loss. The quasi-orthogonal prototype vector is further reinforced until the loss of the auxiliary interference identification network model based on transductive learning converges, and the target transductive small sample incremental learning radar composite interference identification model is obtained.

8. The radar composite interference identification method based on incremental learning according to claim 7 is characterized in that: The calculation formula for the transductive training loss is In the formula, CE is the standard cross entropy loss, I KL is the mutual information based on KL divergence, I α is the mutual information based on α divergence, X Q is the interference sample in the input unbalanced test set, Y Q is the predicted label of the output, λ is the parameter, and p ik is the posterior distribution of the label, p k is the marginal distribution of labels.

9. The radar composite interference identification method based on incremental learning according to claim 1 is characterized in that: Single jammer type labels include distance deception jammer, speed deception jammer, intermittent sampling direct jammer, intermittent sampling repeated jammer, and dense false target jammer; The composite jamming type labels include distance deception and intermittent sampling repetitive composite jamming, distance deception and intermittent sampling direct composite jamming, speed deception and intermittent sampling direct composite jamming, dense false targets and intermittent sampling direct composite jamming, and distance deception and speed deception composite jamming.

10. A radar composite interference identification system based on incremental learning, characterized in that: include: An acquisition module, used for acquiring multiple radar interference time-frequency diagram data sets with interference type labels; A training set construction module, used to construct a basic training data set, an incremental training data set and an unbalanced test set according to a plurality of radar interference time-frequency diagram data sets with interference type labels; The processing module pre-builds a direct-pushing small-sample incremental learning radar composite interference recognition model; It is also used to perform base class training and incremental learning on a direct-pushing small sample incremental learning radar composite interference recognition model using a basic training data set and an incremental training data set, respectively, to obtain a direct-pushing small sample incremental learning radar composite interference recognition model after incremental training; It is also used to perform direct learning on the incrementally trained direct-pushing small sample incremental learning radar composite interference recognition model using an unbalanced test set to obtain a target direct-pushing small sample incremental learning radar composite interference recognition model; The recognition module is used to input the time-frequency image of the radar interference to be tested into the target direct push small sample incremental learning radar composite interference recognition model to obtain the prediction recognition result of the time-frequency image of the radar interference to be tested; Among them, the direct-pushing small sample incremental learning radar composite interference recognition model includes a basic interference recognition network model based on meta-learning, an interference recognition network model based on incremental learning, and an auxiliary interference recognition network model based on direct-pushing learning; among them, the basic interference recognition network model based on meta-learning includes a feature extractor based on a CNN network structure, a fully connected layer, and a dynamically growing EM memory; The basic network model of the direct-pushing small-sample incremental learning radar composite interference recognition model consists of a freezable feature extractor, a trainable fixed-size fully connected layer, and a rewritable dynamically growing EM memory; the frozen part is separated from the incremental part by inserting a fully connected layer, which outputs a class vector in a hyperdimensional embedding space with a fixed dimension. Among them, the construction of the direct-pushing small sample incremental learning radar composite interference recognition model includes three stages, all of which adopt the basic network model structure; The first stage is the basic training stage, in which a basic interference recognition network model based on meta-learning is constructed. In the basic training stage, the feature extractor is a CNN network model that performs meta-learning through appropriate sharpened attention, and uses quasi-orthogonal vectors to represent different interference categories, which is used to achieve accurate recognition of basic known single-category interference and serves as a pre-training model for the incremental training stage. The second stage is the incremental training stage, which builds an interference recognition network model based on incremental learning. In the second stage, based on the pre-trained model in the first stage, a very small number of new category composite interference samples are input into the network model to form new category quasi-orthogonal prototype vectors in the dynamically growing EM memory. Then, the quasi-orthogonal prototype vectors stored in the dynamically growing EM memory are bipolarized, and the fully connected layer is retrained in an iterative manner to obtain a classifier model suitable for all current interference types. The third stage is the direct learning stage, in which an auxiliary interference recognition network model based on direct learning is constructed; in the third stage, based on the model trained in the second stage, the interference samples of the unbalanced test set are used as the query set, and the training set of the second stage is used as the support set. The direct learning model is used to perform direct learning on the category quasi-orthogonal prototype vectors stored in the dynamically growing EM memory to obtain a better classifier model.

11. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the computer program, the radar composite interference identification method based on incremental learning as described in any one of claims 1 to 9 is implemented.

12. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the radar composite interference identification method based on incremental learning as described in any one of claims 1 to 9 is implemented.

Images