Radar active composite interference identification method based on efficient multi-modal cache model

Abstract

The invention discloses a radar active composite interference identification method based on an efficient multi-modal cache model, and belongs to the field of radar active interference identification, and the method comprises the steps: obtaining simulation radar active interference time domain data, splitting the simulation radar active interference time domain data into a training set and a test set, and obtaining a simulation radar active interference time domain data set; dividing the training set into a training support set and a training query set; constructing an efficient multi-modal cache model based on a visual cache module, a text cache module and a residual connection module; training the efficient multi-modal cache model based on the training support set and the training query set, and testing the trained model through a test set to obtain an interference identification model; and radar active composite interference is identified based on the interference identification model. Limited interference samples can be efficiently utilized and deeply excavated, so that the accuracy of radar active composite interference identification is improved.

Description

Technical Field

[0001] The invention belongs to the field of radar active interference identification, and in particular relates to a radar active composite interference identification method based on an efficient multi-modal cache model. Background Art

[0002] Anti-ship missiles are missiles launched from naval ships, shores or aircraft to attack surface warships. In modern warfare, with the rapid development of ship jamming equipment, the survival and operation of anti-ship missiles are facing increasingly severe threats. Therefore, it is particularly urgent to improve the anti-interference capability of the seeker radar of anti-ship missiles. As the primary link of radar anti-interference technology, interference type identification has received extensive attention and research in recent years.

[0003] In order to accurately distinguish the types of radar interference, a variety of traditional interference identification methods have emerged. These methods can be roughly divided into two categories: likelihood-based methods and multi-domain feature extraction methods. Likelihood-based methods focus on in-depth modeling of echo signals and use statistical methods such as generalized likelihood ratio tests to distinguish and identify different types of interference. However, such methods are highly dependent on prior knowledge and usually require manual setting of complex threshold parameters, which to a certain extent limits their wide application in complex and changeable battlefield environments. Feature extraction-based methods emphasize extracting features from multiple fields (such as time domain, frequency domain, spatial domain, etc.) from radar signals, and combining machine learning algorithms or other advanced classifiers to achieve automatic identification of interference signals. Although this method has higher flexibility and adaptability in theory, manual feature extraction is not only time-consuming and labor-intensive, increasing the recognition cost, but also the quality of the extracted features directly determines the accuracy and robustness of the recognition results. Therefore, how to automatically and efficiently extract high-quality features has become a key challenge to improve the performance of radar interference recognition. Summary of the invention

[0004] In order to solve the above technical problems, the present invention provides a radar active composite interference identification method based on an efficient multi-modal cache model, comprising:

[0005] Acquire simulated radar active interference time domain data, split the simulated radar active interference time domain data into a training set and a test set, and divide the training set into a training support set and a training query set;

[0006] Construct an efficient multimodal cache model based on visual cache module, text cache module and residual connection module;

[0007] The efficient multimodal cache model is trained based on the training support set and the training query set, and the trained model is tested by a test set to obtain an interference recognition model;

[0008] Radar active composite interference is identified based on the interference identification model.

[0009] Preferably, the radar active jamming time domain data includes: noise product jamming, intelligent noise jamming, spectrum dispersion jamming, dense false target jamming, intermittent sampling forwarding jamming and frequency modulation jamming.

[0010] Preferably, the operation process of the efficient multimodal cache model includes:

[0011] Based on the visual cache module, the time-frequency image features and label unique hot encoding of the training support set interference are extracted to construct key-value pairs; wherein the visual cache module includes a visual encoder, a visual feature optimizer, a unique hot encoder and a visual cache, the visual feature optimizer is composed of two fully connected layers and a dropout layer, the output dimension of the first fully connected layer is 512, and the output dimension of the second fully connected layer is 36; the visual encoder is composed of the visual encoding part based on the pre-trained CLIP;

[0012] Extracting text features constructed based on label prior information based on the text cache module; wherein the text cache module includes a text encoder, a text feature optimizer and a text cache; the text encoder is composed of a text encoding part based on the pre-trained CLIP, the text feature optimizer is composed of a fully connected layer and a dropout layer, and the output dimension of the fully connected layer is 36;

[0013] Extracting interference time-frequency features of a training query set based on the visual encoder and the visual feature optimizer, retrieving interference time-frequency features of the training query set and visual affinity of a training support set based on the visual cache to obtain visual prediction features;

[0014] Obtain text prediction features based on the text cache retrieval training query set interference time-frequency features and training support set text affinity;

[0015] The visual prediction feature and the text prediction feature are fused based on a residual module to obtain an interference recognition result.

[0016] Preferably, the process of extracting the time-frequency image features of the training support set interference and the label unique-hot encoding to construct key-value pairs based on the visual cache module includes:

[0017] The labels of the training support set are converted into one-hot encodings based on the one-hot encoder:

[0018] The time domain signal of the training support set is converted into a time-frequency graph based on short-time Fourier transform;

[0019] Extracting time-frequency features from the time-frequency graph based on the visual encoder and the visual feature optimizer;

[0020] The time-frequency feature is used as a value of a visual cache, and the one-hot encoding is stored in the visual cache as a key of the visual cache.

[0021] Preferably, the process of extracting text features constructed based on tag prior information based on the text cache module includes:

[0022] Convert the training support set labels into simple text prompts, convert the simple text prompts into text vectors and input them into a text encoder to obtain text features;

[0023] Optimizing the text feature based on the text feature optimizer to obtain an optimized text vector;

[0024] The text vectors related to each category in the optimized text vectors are combined into a text feature matrix and stored in the text cache.

[0025] Preferably, the expression for obtaining the visual prediction feature is:

[0026]

[0027] in, is the visual prediction feature, β represents the hyperparameter for adjusting the sharpness of similarity, Indicates the calculation of the sample features f of the training query set query The cosine similarity between the features of any sample in the training support set, L train is the one-hot encoding matrix.

[0028] Preferably, the expression for obtaining the text prediction feature is:

[0029]

[0030] in, is the text prediction feature, is the text feature matrix.

[0031] Preferably, the process of training the efficient multimodal cache model based on the training support set and the training query set further includes:

[0032] The cross entropy loss is calculated using the true labels of the training query set samples, and the visual encoder and text encoder parameters in the model are frozen in combination with an efficient transfer learning strategy. The visual feature optimizer and text feature optimizer parameters in the model are iteratively optimized using a back gradient propagation algorithm until the loss value no longer decreases. The iterative optimization is stopped and the interference recognition model is output.

[0033] Preferably, the expression for obtaining the interference identification result is:

[0034]

[0035] Among them, f test represents the visual time-frequency features obtained after the test set interference is processed by the visual encoder and the visual feature optimizer, S final Interference identification results.

[0036] Compared with the prior art, the present invention has the following advantages and technical effects:

[0037] The present invention uses an optimized efficient multimodal cache model to extract the interference features of the training set and stores them in the visual cache and text cache respectively, inputs the test set interference time-frequency diagram into the visual encoder and visual feature optimizer in the visual cache module to extract the interference time-frequency features of the training query set, and uses the visual cache and text cache to obtain prediction results respectively, and finally obtains the interference recognition result in the residual connection module. The present invention can efficiently utilize and deeply mine limited interference samples, thereby improving the accuracy of radar active composite interference recognition. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] The drawings constituting a part of the present application are used to provide a further understanding of the present application. The illustrative embodiments and descriptions of the present application are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:

[0039] Figure 1 It is a flow chart of a radar active composite interference identification method based on an efficient multi-modal cache model according to an embodiment of the present invention;

[0040] Figure 2 Schematic diagram of an efficient multimodal cache model according to an embodiment of the present invention. DETAILED DESCRIPTION

[0041] It should be noted that, in the absence of conflict, the embodiments and features in the embodiments of the present application can be combined with each other. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0042] It should be noted that the steps shown in the flowcharts of the accompanying drawings can be executed in a computer system such as a set of computer executable instructions, and that, although a logical order is shown in the flowcharts, in some cases, the steps shown or described can be executed in an order different from that shown here.

[0043] Embodiment 1

[0044] like Figure 1-2 As shown, this embodiment provides a radar active composite interference identification method based on an efficient multi-modal cache model, including:

[0045] Simulate radar active jammer time domain data and split the samples into training set and test set. In each epoch, further split the training set into training support set and training query set.

[0046] Build an efficient multimodal cache model;

[0047] The visual cache module is used to extract the time-frequency image features and label unique-hot encoding of the training support set interference to construct key-value pairs and store them in the visual cache;

[0048] The text cache module is used to extract text features based on label prior information and store them in the text cache;

[0049] The visual encoder and visual feature optimizer in the visual cache module are used to extract the interference time-frequency features of the training query set. At the same time, the visual cache is used to retrieve the interference time-frequency features of the training query set and the visual affinity of the training support set to obtain the visual prediction features;

[0050] The text cache is used to retrieve the training query set interference time-frequency features and the training support set text affinity to obtain text prediction features;

[0051] The residual connection module is used to fuse the prediction results of the two modalities;

[0052] The cross entropy loss is calculated using the true labels of the training query set samples. The parameters of the visual encoder and text encoder in the model are frozen in combination with an efficient transfer learning strategy, and only the parameters of the visual feature optimizer and text feature optimizer in the model are optimized.

[0053] The optimized efficient multimodal cache model is used to extract the interference features of the training set and store them in the visual cache and text cache respectively. The interference time-frequency diagram of the test set is input into the visual encoder and visual feature optimizer in the visual cache module to extract the interference time-frequency features of the training query set, and the visual cache and text cache are used to obtain the prediction features respectively. Finally, the interference recognition result is obtained in the residual connection module.

[0054] This embodiment can fully mine and utilize the information in limited interference samples by constructing an efficient multimodal cache model. Specifically, the model designs a visual cache module and a text cache module. The visual cache module is used to extract the time-frequency image features and unique hot encoding labels of interference to form key-value pairs. By adopting an efficient key-value affinity calculation mechanism, the model effectively utilizes and deeply mines the prior information of interference, thereby obtaining visual prediction features with discriminativeness. At the same time, the text description constructed based on the interference prior information is input into the text cache module to extract text features and obtain text prediction features. In the designed residual connection module, the prediction features of the two modalities are fused to make full use of the complementarity between different modalities to improve the sensitivity and accuracy of the model in identifying similar interference types. In the model optimization process, an efficient transfer learning strategy is combined to optimize only some parameters of the model to improve its optimization efficiency.

[0055] In this embodiment, radar active jamming time domain data is simulated, and the samples are split into a training set and a test set. Each epoch further splits the training set into a training support set and a training query set, specifically including:

[0056] MATLAB is used to simulate and generate time domain data of six common single types and six efficient combined types of radar active jamming. The single jamming types include noise product jamming (NPJ), smart noise jamming (SNJ), smeared spectrum jamming (SMSP), dense false target jamming (DFTJ), intermittent sampling repeater jamming (ISRJ) and modulated intermittent sampling repeater jamming (MISRJ). In addition, the six combined jamming types are composed of the above single jamming types, namely, the combination of DFTJ and ISRJ, the combination of DFTJ and MISRJ, the combination of MISRJ and ISRJ, the combination of MISRJ and SMSP, the combination of SMSP and DFTJ, and the combination of SMSP and ISRJ. Each jamming type contains 500 simulated samples, 3 of which are randomly selected to form the training set, and the rest are included in the test set. In each training cycle (epoch), the samples in the training set are randomly assigned again, some of which serve as training support sets and the rest as training query sets.

[0057] In this embodiment, an efficient multimodal cache model is constructed, referring to Figure 2, specifically including: the efficient multimodal cache model mainly includes a visual cache module, a text cache module, and a residual connection module; the visual cache module mainly includes a visual encoder, a visual feature optimizer, a one-hot encoder, and a visual cache. The visual encoder is taken from the visual encoding part of the pre-trained CLIP, and CLIP is pre-trained on 200 million image-text pairs. The output of the visual encoder is connected to the input of the visual feature optimizer. The visual feature optimizer consists of two fully connected layers and a dropout layer, where the output dimension of the first fully connected layer is 512 and the output dimension of the second fully connected layer is 36. The one-hot encoder encodes according to the labels of the training support set samples. The output of the visual feature optimizer and the one-hot encoder are connected to the input of the visual cache. The interference time-frequency features and label one-hot encoding output by the visual feature optimizer are stored as key-value pairs in the visual cache for visual modality query training query set sample categories. The text cache module mainly includes three parts: a text encoder, a text feature optimizer, and a text cache. The text encoder is taken from the text encoding part of the pre-trained CLIP. CLIP is pre-trained on 200 million image-text pairs. The output of the text encoder is connected to the input of the text feature optimizer. The text feature optimizer consists of a fully connected layer and a dropout layer, where the output dimension of the fully connected layer is 36. The output of the text feature optimizer is connected to the input of the text cache. The text features output by the text feature optimizer are stored in the text cache for text modality query training query set sample categories. The residual connection module consists of a feature fusion and a softmax. The input of the feature fusion is connected to the visual cache and the output of the text cache. The output of the feature fusion is subjected to a softmax operation to obtain the interference recognition result.

[0058] In a further embodiment, the visual cache module is used to extract the time-frequency image features of the training support set interference and the label unique hot encoding to construct a key-value pair stored in the visual cache, specifically including:

[0059] In the visual cache module, a one-hot encoder is used to convert the training support set labels into one-hot encodings:

[0060] L train =OneHot(L N ). (1)

[0061] Among them, OneHot(L N ) represents a one-hot encoding operation. L N is the true label of the samples in the training support set.

[0062] The short-time Fourier transform is used to transform the time domain signal of the training support set into a time-frequency graph, and the visual encoder and visual feature optimizer are used to extract the time-frequency features:

[0063] F train =VFEVFO(S K ). (2)

[0064] Among them, VFEVFO(S K ) represents the joint operation of the visual encoder and the visual feature optimizer. S K The two-dimensional time-frequency original images of the samples in the training support set.

[0065] Based on the cache model paradigm, the interference time-frequency feature F train The values ​​and labels of the visual cache are one-hot encoded L train The keys that are considered as vision cache are stored in the vision cache.

[0066] In a further embodiment, the text features constructed based on the tag prior information are extracted and stored in the text cache by using the text cache module, specifically including:

[0067] In the text cache module, the training support set labels are converted into simple text in the form of "Aphotoofa{class}". For example, for MISRJ interference, the text prompt constructed in this embodiment is Aphoto ofa{MISRJ}. Subsequently, the text prompt is converted into a text vector through tokenize and input into the frozen CLIP pre-trained text encoder for text feature extraction. Next, the text feature representation is further optimized using the designed text feature adaptation head:

[0068] f t =TETFO(T N ) (3)

[0069] Among them, TETFO(T N ) represents the joint operation of the text encoder and the text feature adaptation head. N is the text vector corresponding to the text prompt.

[0070] The text vectors related to each category in the text vector are combined into a text feature matrix Stored in the text cache.

[0071] In a further embodiment, the visual encoder and visual feature optimizer in the visual cache module are used to extract the interference time-frequency features of the training query set. At the same time, the visual cache is used to retrieve the interference time-frequency features of the training query set and the visual affinity of the training support set to obtain the visual prediction features, specifically including:

[0072] The short-time Fourier transform is used to obtain the time-frequency graph of the training query set samples, and the visual encoder in the visual cache module is used to preliminarily extract features, and the visual feature optimizer is used to further extract features f query. Using the interference visual time-frequency features f query Perform efficient key-value affinity calculation in the visual cache to obtain the visual prediction features of the sample:

[0073]

[0074] Among them, β represents the hyperparameter for adjusting the sharpness of similarity, Indicates the calculation of the sample features f of the training query set query The cosine similarity between the training query sample and any sample feature in the training support set. The similarity value range is limited to 0-1 through the exponential function exp. This affinity describes the closeness between the training query sample and the sample in the training support set in the visual feature space. After calculating the affinity, this embodiment adds the affinity to the one-hot encoding matrix L of the label. train Multiply to obtain the visual prediction features of the training query set samples

[0075] In a further embodiment, the text prediction features are obtained by using the text cache retrieval training query set interference time-frequency features and the training support set text affinity, specifically including:

[0076] Using the interference visual time-frequency features f query Multiply the text feature matrix in the text cache to obtain the text prediction features of the training query set samples

[0077]

[0078] Preferably, the residual connection module is used to fuse the two modal prediction features, specifically including:

[0079] In the residual connection module, the interference prediction features of the query set retrieved in the two modal caches are combined through residual connection, and then the probability of the sample belonging to different interference categories is obtained through the softmax layer:

[0080]

[0081] Among them, α represents the residual ratio, which is used to balance the visual prediction features. and text prediction features

[0082] In a further embodiment, the cross entropy loss is calculated using the true labels of the training query set samples, and the visual encoder and text encoder parameters in the model are frozen in combination with an efficient transfer learning strategy, and only the visual feature optimizer and text feature optimizer parameters in the model are optimized, specifically including:

[0083] The cross entropy loss function is calculated using the true labels and predicted labels of the training query set samples:

[0084]

[0085] Where M represents the number of samples in the training query set, y m Represents the true label of the sample. The model combines an efficient transfer learning strategy to freeze the parameters of the visual encoder and text encoder, and uses the reverse gradient propagation algorithm to iteratively optimize the parameters of the visual feature optimizer and text feature optimizer in the model. The iterative optimization is stopped until the loss value no longer decreases, and the training of the proposed model is completed.

[0086] The optimization process uses the AdamW optimizer, with the learning rate set to 0.001 and the weight decay coefficient set to 0.0001.

[0087] In a further embodiment, the interference features of the training set are extracted using the optimized efficient multimodal cache model and stored in the visual cache and the text cache respectively, the interference time-frequency diagram of the test set is input into the visual encoder and the visual feature optimizer in the visual cache module to extract the interference time-frequency features of the training query set, and the prediction features are obtained using the visual cache and the text cache respectively, and finally the interference recognition result is obtained in the residual connection module, which specifically includes:

[0088] The visual cache module in the optimized efficient multimodal cache model is used to extract the interference video features of all samples in the training set as the key F train Stored in the visual cache, and all sample labels are one-hot encoded to form a one-hot encoding matrix as the value L train Stored in the visual cache. Use the text cache module to extract text features of different categories and store them in the text cache.

[0089] The short-time Fourier transform is used to obtain the interference time-frequency diagram of the test set, and the interference feature f of the test machine is obtained through the visual encoder and visual feature optimizer in the visual cache module. test , and then obtain the interference prediction features of different modes in the visual cache and text cache respectively, and connect them in the residual connection module to obtain the interference recognition result:

[0090]

[0091] Among them, f test represents the visual time-frequency features of the test set after being disturbed by the visual encoder and the visual feature optimizer. The meanings of the other letters are the same as above.

[0092] The effect of this embodiment can be further illustrated by the following simulation, as shown in Table 1 and Table 2.

[0093] 1. Simulation conditions

[0094] Table 1

[0095] CPU AMD Ryzen 7 4800H Memory 32GB Graphics card model NVIDIAGeForceGTX4070Ti Graphics card memory 12GB

[0096] Table 2

[0097] operating system Windows 10 64 bit Compiler Pycharm2021.3.1 CUDA Version 11.2.0

[0098] 2. Simulation method

[0099] (1) Radar interference classification methods based on convolutional neural network (CNN) in the prior art, including deep fusion model (DFCNN), Siamese-CNN (S-CNN), meta-learning (Model-Agnostic Meta-Learning, MAML), JR-TFSAD, CL-TD; (2) The method described in this embodiment, namely, radar active composite interference identification based on efficient multimodal cache model.

[0100] 3. Simulation content and simulation results

[0101] The simulation experiment selected six single types of interference, namely noise product interference, intelligent noise interference, spectrum dispersion interference, dense false target interference, intermittent sampling forwarding interference and frequency modulation interference, as well as six efficient composite interferences, including the combination of dense false target interference and intermittent sampling forwarding interference, the combination of dense false target interference and frequency modulation interference, the combination of frequency modulation interference and intermittent sampling forwarding interference, the combination of frequency modulation interference and spectrum dispersion interference, the combination of spectrum dispersion interference and dense false target interference, and the combination of spectrum dispersion interference and intermittent sampling forwarding interference. The time-frequency diagram of the twelve interferences obtained by short-time Fourier transform.

[0102] In simulation experiment 1, DFCNN, S-CNN, MAML, JR-TFSAD, and CL-TD are used to identify and classify radar active deception jamming at a fixed interference-to-noise ratio and sample size, as shown in Table 3. The active jamming recognition method based on small sample learning and multi-structure feature fusion in this embodiment is used to identify and classify radar active deception jamming signals at a fixed interference-to-noise ratio and sample size, as shown in Table 3.

[0103] Table 3

[0104]

[0105]

[0106] In order to verify the superiority of this embodiment under the condition of small samples, when each interference type has 3 training samples, the recognition results of different models for 12 interference types are provided, as shown in Table 1. As can be seen from Table 1, this embodiment is superior to other comparative models in terms of interference recognition precision, accuracy, F1 and kappa. Specifically, the recognition accuracy of this embodiment is as high as 92%, which is significantly better than methods such as DFCNN that rely heavily on the number of training samples. In addition, compared with a few existing recognition algorithms (such as S-CNN, MAML, JR-TFSAD and CL-TD), the recognition accuracy of this embodiment is improved by more than 10%. More specifically, this embodiment achieves the highest recognition accuracy on 9 interference signals and maintains a recognition accuracy of more than 90% on 7 interferences. In particular, when facing DFTJ+ISRJ interference, the recognition accuracy of EMMCM is as high as 99.38%, while the recognition accuracy of the comparison model is less than 60%. In addition, the standard deviation of the EMMCM interference recognition result is much lower than that of the comparison model, indicating that this embodiment has excellent stability. In summary, the results in Table 3 show that the present embodiment outperforms other comparison models in terms of accuracy and stability of interference identification.

[0107] Table 4

[0108]

[0109] Simulation experiment 2, in order to quantitatively evaluate the optimization efficiency of this embodiment, the training set of 2, 3 and 30 labeled interference samples of each type is used to train this embodiment for 100 iteration cycles. At the same time, a few sample interference recognition models such as S-CNN, MAML, JR-TFSAD and CL-TD are used as comparison models. The training time of different models under different sample sizes is shown in Table 4. As can be seen from Table 4, compared with the comparison model, this embodiment has the highest optimization efficiency under different sample sizes. In particular, when there are only 2 training samples for each interference type, this embodiment only needs 15.41 seconds to train 100 iteration cycles, which is much less than the few sample recognition model S-CNN with relatively high optimization efficiency. This shows the effectiveness of this embodiment combined with an efficient transfer learning strategy, by freezing the parameters of the visual encoder and text encoder in the model, and only optimizing the parameters of the visual feature optimizer and the text feature optimizer. This makes the proposed method have the highest optimization efficiency.

[0110] The above are only preferred specific implementations of the present application, but the protection scope of the present application is not limited thereto. Any changes or substitutions that can be easily thought of by a person skilled in the art within the technical scope disclosed in the present application should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

Claims

1. A radar active composite interference identification method based on an efficient multi-modal cache model, characterized in that: include: Acquire simulated radar active interference time domain data, split the simulated radar active interference time domain data into a training set and a test set, and divide the training set into a training support set and a training query set; Construct an efficient multimodal cache model based on visual cache module, text cache module and residual connection module; The efficient multimodal cache model is trained based on the training support set and the training query set, and the trained model is tested by a test set to obtain an interference recognition model; Radar active composite interference is identified based on the interference identification model.

2. The method according to claim 1, characterized in that The radar active jamming time domain data includes: noise product jamming, intelligent noise jamming, spectrum dispersion jamming, dense false target jamming, intermittent sampling forwarding jamming and frequency modulation jamming.

3. The method according to claim 1, characterized in that The operation process of the efficient multimodal cache model includes: Based on the visual cache module, the time-frequency image features and label unique hot encoding of the training support set interference are extracted to construct key-value pairs; wherein the visual cache module includes a visual encoder, a visual feature optimizer, a unique hot encoder and a visual cache, the visual feature optimizer is composed of two fully connected layers and a dropout layer, the output dimension of the first fully connected layer is 512, and the output dimension of the second fully connected layer is 36; the visual encoder is composed of the visual encoding part based on the pre-trained CLIP; Extracting text features constructed based on label prior information based on the text cache module; wherein the text cache module includes a text encoder, a text feature optimizer and a text cache; the text encoder is composed of a text encoding part based on the pre-trained CLIP, the text feature optimizer is composed of a fully connected layer and a dropout layer, and the output dimension of the fully connected layer is 36; Extracting interference time-frequency features of a training query set based on the visual encoder and the visual feature optimizer, retrieving interference time-frequency features of the training query set and visual affinity of a training support set based on the visual cache to obtain visual prediction features; Obtain text prediction features based on the text cache retrieval training query set interference time-frequency features and training support set text affinity; The visual prediction feature and the text prediction feature are fused based on a residual module to obtain an interference recognition result.

4. The method according to claim 3, characterized in that The process of extracting the time-frequency image features of the training support set interference and the label unique hot encoding to construct key-value pairs based on the visual cache module includes: The labels of the training support set are converted into one-hot encodings based on the one-hot encoder: The time domain signal of the training support set is converted into a time-frequency graph based on short-time Fourier transform; Extracting time-frequency features from the time-frequency graph based on the visual encoder and the visual feature optimizer; The time-frequency feature is used as a value of a visual cache, and the one-hot encoding is stored in the visual cache as a key of the visual cache.

5. The method according to claim 3, characterized in that: The process of extracting text features constructed based on label prior information based on the text cache module includes: Convert the training support set labels into simple text prompts, convert the simple text prompts into text vectors and input them into a text encoder to obtain text features; Optimizing the text feature based on the text feature optimizer to obtain an optimized text vector; The text vectors related to each category in the optimized text vectors are combined into a text feature matrix and stored in the text cache.

6. The method according to claim 3, characterized in that The expression for obtaining the visual prediction feature is: in, is the visual prediction feature, β represents the hyperparameter for adjusting the sharpness of similarity, Indicates the calculation of the sample features f of the training query set query The cosine similarity between the features of any sample in the training support set, L train is the one-hot encoding matrix.

7. The method according to claim 3, characterized in that The expression for obtaining the text prediction feature is: in, is the text prediction feature, is the text feature matrix.

8. The method according to claim 3, characterized in that The process of training the efficient multimodal cache model based on the training support set and the training query set also includes: The cross entropy loss is calculated using the true labels of the training query set samples, and the visual encoder and text encoder parameters in the model are frozen in combination with an efficient transfer learning strategy. The visual feature optimizer and text feature optimizer parameters in the model are iteratively optimized using a back gradient propagation algorithm until the loss value no longer decreases. The iterative optimization is stopped and the interference recognition model is output.

9. The method according to claim 3, characterized in that: The expression for obtaining the interference identification result is: Among them, f test represents the visual time-frequency features obtained after the test set interference is processed by the visual encoder and the visual feature optimizer, S final Interference identification results.

Images