Active interference identification method based on cross-modal learning

Abstract

The invention discloses an active interference identification method based on cross-modal learning. The method comprises the following specific steps: constructing an active interference identification reference data set; defining a knowledge modeling form of external knowledge in the active interference ontology model, and converting active interference knowledge modeling into semantic representation; based on the interference data feature VAE model and the interference semantic representation VAE model, constructing an active interference data and knowledge hybrid driven cross-modal learning model; and constructing a generalized zero sample classification model to realize active interference identification. According to the invention, through data and knowledge hybrid driven cross-modal learning, the problems of few samples and zero samples faced by active interference identification are solved.

Description

Technical Field

[0001] The present invention relates to the technical field of active interference identification, and in particular to an active interference identification method based on cross-modal learning. Background Art

[0002] Active interference identification refers to the analysis of electromagnetic signals received by radar through specific technical means to determine whether they are active interference signals, and further identify the type and source of interference. Active interference identification is an important prerequisite for radar to resist active interference in complex electromagnetic environments. Due to the long acquisition cycle and high cost of active interference identification tests, as well as the rapid development of interference technology, interference styles are endless and difficult to enumerate, and it is very difficult to obtain a large number of interference signals of various types.

[0003] Existing technologies have improved the recognition performance of active interference under small samples to a certain extent by constructing metric learning or ensemble learning frameworks, but they all follow the general idea of ​​data-driven, have limitations in the algorithm framework, and cannot deal with the zero-sample problem in practical applications. Summary of the invention

[0004] The purpose of the present invention is to provide an active interference identification method based on cross-modal learning, which solves the few-sample and zero-sample problems faced by active interference identification through cross-modal learning driven by a mixture of data and knowledge.

[0005] In order to achieve the above object, the present invention provides an active interference identification method based on cross-modal learning, and the specific steps include:

[0006] Step S1, constructing an active interference identification benchmark data set;

[0007] Step S2, clarifying the knowledge modeling form of external knowledge in the active interference ontology model, and converting the active interference knowledge modeling into semantic representation;

[0008] Step S3, constructing a cross-modal learning model driven by active interference data and knowledge based on the interference data feature VAE model and the interference semantic representation VAE model;

[0009] Step S4, constructing a generalized zero-shot classification model to realize active interference identification.

[0010] Optionally, the knowledge modeling form of the external knowledge in the active interference ontology model in step S2 is a knowledge graph.

[0011] Optionally, the step S2 further includes:

[0012] Step S2.1, construct an active interference ontology model, use knowledge modeling to construct the logical relationship between the required information in the active interference ontology model, and complete the knowledge graph model layer construction;

[0013] Step S2.2, establishing a knowledge graph pattern layer definition model based on the active interference identification benchmark data set in step S1, and obtaining and storing the attribute values ​​of the attributes possessed by each instance through data simulation;

[0014] Step S2.3, the active interference ontology model and the attributes of each instance are converted into triples to form a complete active interference knowledge graph;

[0015] In step S2.4, the active interference knowledge graph is embedded into the model, converted into a vector space, and the semantic representation of various active interference signal types is obtained.

[0016] Optionally, the active interference ontology model includes: the generation mechanism, action mechanism, multi-domain characteristics, multi-domain and multi-dimensional features, radar technology systems that can interfere, and interference countermeasures or suppression measures of different styles of active interference; wherein, the multi-domain and multi-dimensional features are obtained by extracting the features of the active interference identification benchmark data set in step S1; the multi-domain characteristics are summarized and condensed based on the multi-domain and multi-dimensional features; the generation mechanism, action mechanism, radar technology systems that can interfere, and interference countermeasures or suppression measures of different styles of active interference are obtained through external knowledge.

[0017] Optionally, the knowledge graph embedding model is a TranE model.

[0018] Optionally, step S3 further includes:

[0019] Step S3.1, based on the active interference identification benchmark dataset of step S1 and the semantic representation of step S2, construct the input of cross-modal learning driven by active interference data and knowledge hybrid, which can be expressed as:

[0020] {(s i , c i , l i )|=1,2,…M}

[0021] In the formula, s i and l i represents the data sample and active interference category label of the i-th example in the active interference identification benchmark dataset constructed in step S1, c i represents the first i The semantic representation of the class, M represents the number of instances based on step S2.2;

[0022] Step S3.2, based on the interference data feature VAE model and the interference semantic representation VAE model, a cross-modal learning model driven by active interference data and knowledge is established;

[0023] Step S3.3, reconstruct the loss of the interference data feature VAE model and the interference semantic representation VAE model. The total loss l of the cross-modal learning model driven by the active interference data and knowledge mixture is composed of the multimodal reconstruction loss Cross-modal feature reconstruction loss Hidden space multimodal feature distribution matching loss and multi-class distribution separation loss The composition can be expressed as:

[0024] l=l VAE +α·l CMFR +β·l JSDM +γ·l MCDD

[0025] Where α, β, and γ represent the weight factors of cross-modal feature reconstruction loss, latent space multimodal feature distribution matching loss, and multi-category distribution separation loss, respectively.

[0026] Optionally, in step S3.3, the multimodal reconstruction loss l VAE VAE model loss for interference data features and interference semantic representation VAE model loss The superposition of can be expressed as:

[0027]

[0028] In the formula, x represents the original feature of the interference data, represents the interference data feature encoder, Interference data feature decoder; (j) represents the interference data features in the latent space, z (a) represents the semantic representation in the latent space; c represents the semantic representation of the interference type, Represents interference semantic representation encoder, Interference semantic representation decoder; D KL (·) indicates the calculation of KL divergence, and η is the weight factor of KL divergence.

[0029] Optionally, the step S1 further includes:

[0030] Step S1.1, generating a target signal, various single interference signals, and a noise signal through simulation; and performing additive compounding on the generated various single interference signals to generate a variety of composite interference signals; superimposing the target signal, the single interference signal or the composite interference signal, and the noise signal with different weights to obtain received signals with different signal-to-noise ratios;

[0031] Step S1.2, performing digital down-conversion on the received signal to obtain a baseband signal, and obtaining frequency domain filter parameters through constant false alarm detection and interference signal parameter estimation;

[0032] Step S1.3, preprocessing and standardizing the interference samples of the baseband signal to obtain interference identification sample data;

[0033] Step S1.4, divide the interference identification sample data in step S1.3 into sample labels of single interference signal and composite interference signal types based on the values ​​of the binary vector labels, and obtain an active interference identification benchmark data set.

[0034] Optionally, in step S1.4, a binary vector label is constructed in the form of a one-dimensional array, where each element in the one-dimensional array represents a single interference signal type, and when the single interference signal type exists in the interference identification sample data, the element in the one-dimensional array takes 1; when the single interference signal type does not exist in the interference identification sample data, the element in the one-dimensional array takes 0.

[0035] Optionally, in step S4, the generalized zero-shot classification model includes: a training phase and a testing phase; and according to the training requirements based on the generalized zero-shot classification model, the active interference identification benchmark data set in step S1 is divided into an unknown class and an unknown class;

[0036] In the training phase, the semantic representation c corresponding to the interference signals of known and unknown classes of the active interference recognition benchmark dataset in step S1 is fed into the interference semantic representation encoder to obtain the latent space semantic representation z (a) , then perform machine learning under the supervision of the binary vector label in step S1.4 to train the generalized zero-shot classifier F GZSL ;

[0037] In the test phase, the interference data samples s of known or unknown classes corresponding to the training phase are sent to the interference feature extraction network, and the interference data features x are output, which are then sent to the interference data feature encoder to obtain the latent space interference features z. (j) , using the trained generalized zero-shot classifier F GZSL Interference feature z in latent space (j) Classify and identify active interference.

[0038] Compared with the prior art, the technical solution of the present invention has at least the following beneficial effects:

[0039] The present invention converts external knowledge into semantic representation, guides the cross-modal learning model machine driven by a mixture of active interference data and knowledge to learn the essential characteristics of active interference data, improves the generalization performance of the cross-modal learning model driven by a mixture of active interference data and knowledge, and makes the present invention applicable to situations with less sample data; uses external knowledge to build a bridge between the target domain (i.e., unknown class) and the source domain (i.e., known class) to solve the problem of zero samples.

[0040] The present invention embeds external knowledge into vector space to obtain semantic representation, achieving good results in active interference identification under few-sample and zero-sample conditions. The present invention can also realize knowledge modeling of knowledge graphs in external knowledge. Since knowledge graphs contain external knowledge of types such as text and attributes, they have stronger semantic richness. Vector-based knowledge graph embedding is used for semantic representation, and thus combined with data-driven deep learning models, more semantic representations can be obtained, greatly improving the performance of cross-modal learning models driven by active interference data and knowledge. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 It is a flow chart of the active interference identification method based on cross-modal learning of the present invention.

[0042] Figure 2 A schematic diagram of a cross-modal learning model driven by a hybrid of active interference data and knowledge in the active interference identification method based on cross-modal learning of the present invention.

[0043] Figure 3 It is a schematic diagram of a generalized zero-shot classification model of the active interference identification method based on cross-modal learning of the present invention. DETAILED DESCRIPTION

[0044] The technical solution of the present invention will be described clearly and completely below in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0045] In the description of the present invention, it should be noted that the terms "upper", "lower", "left", "right", "vertical", "horizontal", "inner", "outer", etc., indicating the orientation or positional relationship, are based on the orientation or positional relationship shown in the drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as limiting the present invention. In addition, the terms "first", "second", and "third" are used for descriptive purposes only, and cannot be understood as indicating or implying relative importance.

[0046] In the description of the present invention, it should be noted that, unless otherwise clearly specified and limited, the terms "installed", "connected", and "connected" should be understood in a broad sense, for example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection; it can be a direct connection, or it can be indirectly connected through an intermediate medium, or it can be the internal communication of two components. For ordinary technicians in this field, the specific meanings of the above terms in the present invention can be understood according to specific circumstances.

[0047] like Figure 1 As shown, the present invention provides an active interference identification method based on cross-modal learning, and the specific steps include:

[0048] Step S1, constructing an active interference identification benchmark data set.

[0049] Step S1.1, generating receiving signals with different signal-to-noise ratios.

[0050] Specifically, simulation software is used to simulate and generate target signals, various single interference signals, and noise signals based on radar transmission signal models, target models, single interference signal models, and noise signal models; and various types of generated single interference signals are additively compounded to generate multiple composite interference signals; the various types of single interference signals and multiple composite interference signals together constitute the interference signal. At the same time, the target signal, interference signal, and noise signal are superimposed with different weights to obtain received signals with different signal-to-noise ratios. In a preferred embodiment, the radar transmission signal in the transmission signal model adopts a linear frequency modulation signal, the target model is modeled as a point target, and the noise signal model is modeled as Gaussian white noise, so as to simplify the simulation process.

[0051] Step S1.2, performing signal preprocessing and parameter estimation on the received signal to obtain the baseband signal and filter parameters.

[0052] Specifically, the received signal is digitally down-converted to obtain a baseband signal, and the frequency domain filter parameters are obtained through constant false alarm detection and interference signal parameter estimation.

[0053] Step S1.3, preprocessing and standardizing the interference samples of the baseband signal to obtain interference identification sample data.

[0054] Specifically, the baseband signal is frequency-domain filtered based on the frequency-domain filter parameters in step S1.2 to suppress noise outside the interference frequency band; the filtered baseband signal is subjected to short-time Fourier transform and Gaussian filtering in turn to suppress noise within the interference frequency band; finally, standardized interference identification sample data is obtained through maximum and minimum value normalization, data truncation and bilinear interpolation.

[0055] Step S1.4, annotate and divide the interference identification sample data to form an active interference identification benchmark data set.

[0056] Specifically, the interference identification sample data in step S1.3 is divided into sample annotations of a single interference signal type and a composite interference signal type using binary vector labels to obtain an active interference identification benchmark data set.

[0057] Construct a binary vector label in the form of a one-dimensional array to mark the interference identification sample data. Each element in the one-dimensional array represents a single interference signal type. When the single interference signal type exists in the interference identification sample data, the element in the one-dimensional array takes 1; when the single interference signal type does not exist in the interference identification sample data, the element in the one-dimensional array takes 0. For example: the binary vector label of the sample data is [0,1,0,0,0,0], indicating that the sample data is a single interference signal type; the binary vector label of the sample data is [0,1,0,0,1,0], indicating that the sample data is a composite interference signal type after two single interference signals are combined. At the same time, the sample data of all single interference signals and composite interference signals are marked with binary vector labels.

[0058] Step S2, clarifying the form of knowledge modeling in the active interference ontology model, and converting the active interference knowledge modeling into semantic representation.

[0059] Step S2.1, construct the active interference ontology model and complete the knowledge graph model layer construction.

[0060] The active interference ontology model includes: the generation mechanism, action mechanism, multi-domain characteristics, multi-domain multi-dimensional features of different styles of active interference, radar technology systems that can interfere, interference countermeasures or suppression methods, etc. Among them, the active interference ontology model is stored in owl format, and the multi-domain multi-dimensional features are obtained by extracting the features of the active interference identification benchmark data set in step S1.4 using the Protégé ontology modeling tool, and the multi-domain characteristics are obtained by summarizing and condensing the multi-domain multi-dimensional features; other information in the active interference ontology model is obtained through external knowledge.

[0061] The external knowledge is composed of four categories: text, attributes, knowledge graph, ontology and rules, and its semantic richness ranges from weak to strong. Compared with external knowledge of types such as text and attributes, knowledge graph has stronger semantic richness and can contain external knowledge of types such as text and attributes; compared with rule-based external knowledge, knowledge graph can represent knowledge in the underlying graph structure model or vector-based knowledge graph embedding, thereby combining with data-driven deep learning models to improve the performance of deep learning models.

[0062] In a preferred embodiment, the knowledge modeling form of the knowledge graph is selected, and the knowledge modeling is used to construct the logical relationship between the required information (i.e., the knowledge graph) obtained from external knowledge in the active interference ontology model to complete the knowledge graph model layer construction.

[0063] Step S2.2, establish a knowledge graph pattern layer definition model, obtain and store the attribute values ​​of each instance.

[0064] According to the external knowledge analysis, the key parameters of the active interference type are obtained, and according to the key parameters of the active interference type and the typical values ​​of the key parameters, several instances are established based on the active interference identification benchmark data set in step S1.4, and then the pattern layer definition model of the knowledge graph is established. Among them, each instance has key parameter attributes and multi-domain multi-dimensional feature attributes; through data simulation, multi-domain multi-dimensional data feature extraction and multi-domain feature analysis are performed to obtain the key parameter attribute values ​​and multi-domain multi-dimensional feature attribute values ​​of each instance, and the structured attributes of the instance are stored in a format such as csv to expand the data of the knowledge graph pattern layer.

[0065] Step S2.3, complete the construction of the active interference knowledge graph, and store and manage the active interference external knowledge in a graphical form.

[0066] Import the active interference ontology model stored in owl format constructed in step S2.1 and the structured attributes stored in csv and other formats extracted in step S2.2 into the graph database management system Neo4j, so that the active interference ontology model and structured attributes are converted into triples in the form of "class-object attribute-class", "class-data attribute-instance", and "entity-attribute-attribute value" to form a complete active interference knowledge graph. Among them, "entity" is an instance in the knowledge graph; "class" represents the set of entities with the same attributes in the knowledge graph, including the binary vector label in step S1.4; "object attribute" represents the association between entities in the knowledge graph; "data attribute" represents the characteristics of entities in the knowledge graph, and "attribute" and "attribute value" can be obtained according to step S2.2.

[0067] In step S2.4, external knowledge is embedded into the relevant transformation model to obtain the semantic representation of various types of active interference.

[0068] When the knowledge modeling form is text, it is converted into vector space through text embedding model. When the knowledge modeling form is attribute, the attribute set is first selected based on the principles of distinguishability and shareability, and then the non-numeric attributes are converted into numerical values ​​through definition, and finally it is converted into vector space through binarization or normalization. When the knowledge modeling form is knowledge graph, it is converted into vector space through knowledge graph embedding model.

[0069] The active interference knowledge graph in step S2.3 is converted into a vector space. In a preferred embodiment, the knowledge graph embedding model is a TranE model, which can convert the knowledge graph into a vector space to obtain semantic representations of various types of active interference signals.

[0070] Step S3, constructing a cross-modal learning model driven by a mixture of active interference data and knowledge.

[0071] Step S3.1, construct the input of cross-modal learning driven by the hybrid of active interference data and knowledge.

[0072] Based on the sample data and category labels of each example in the active interference identification benchmark dataset in step S1.4 and the semantic representation obtained in step S2.4, the input of cross-modal learning driven by active interference data and knowledge hybrid represented by multimodal data is constructed, which can be expressed as:

[0073] {(s i , c i , l i )|i=1,2,…M} (Formula 1)

[0074] In the formula, s i and l i They represent the sample data and active interference category label of the i-th example in the active interference identification benchmark dataset constructed in step S1.4, and c i represents the first i The semantic representation of the class, M represents the number of instances based on step S2.2.

[0075] Step S3.2, construct a cross-modal learning model driven by a mixture of active interference data and knowledge.

[0076] The cross-modal learning model driven by a hybrid of active interference data and knowledge is established based on two VAE models (Variational Autoencoder), which are an interference data feature VAE model and an interference semantic representation VAE model.

[0077] like Figure 2As shown, the interference data feature VAE model and the interference semantic representation VAE model both include an encoder and a decoder, wherein the encoder maps the input sample data to a latent space, and the decoder realizes the reconstruction of the data in the latent space and restores the data in the latent space to samples. The active interference recognition benchmark data set features of step S2.1 and the vector space corresponding to the semantic representation of step S2.4 are mapped to the latent space, and the common latent space of multimodal distribution alignment is learned by optimizing the joint loss of multimodal reconstruction, cross-modal alignment, and multi-category distribution separation, to obtain the interference data feature encoder and the interference semantic representation encoder. Then, according to the joint loss, the interference data feature decoder and the interference semantic representation decoder are obtained.

[0078] The active interference data in step S3.1 is mixed with the sample data s in the input of the cross-modal learning driven by knowledge mixing (i.e., formula 1) i Input interference feature extraction network, output interference data feature x i ; and the interference data feature x i , No.1 i Semantic representation of class c i Input the interference data feature VAE model and the interference semantic representation VAE model respectively, and finally output the reconstructed interference feature and reconstruct semantic representation

[0079] Step S3.3, reconstruct the loss of the interference data feature VAE model and the interference semantic representation VAE model.

[0080] The total loss l of the cross-modal learning model driven by the active interference data and knowledge is composed of the multimodal reconstruction loss Cross-modal feature reconstruction loss Hidden space multimodal feature distribution matching loss and multi-class distribution separation loss It can be expressed as Equation 2. Among them, the multimodal reconstruction loss is composed of the VAE model loss of each modality (i.e., the interference data feature VAE model and the interference semantic representation VAE model), which is used to ensure the feature expression ability of each modality in the latent space.

[0081] l=l VAE +α·l CMFR +β·l JSDM +γ·l MCDD (Formula 2)

[0082] Where α, β, and γ represent the weight factors of cross-modal feature reconstruction loss, latent space multimodal feature distribution matching loss, and multi-category distribution separation loss, respectively.

[0083] The multimodal reconstruction loss l VAEVAE model loss for interference data features and interference semantic representation VAE model loss The superposition of each of which is composed of reconstruction loss and regularization term (approximately distributed in the KL divergence of the actual posterior distribution), can be expressed as:

[0084]

[0085] In the formula, x represents the original feature of the interference data, represents the interference data feature encoder, Interference data feature decoder; (j) represents the interference data features in the latent space, z (a) represents the semantic representation in the latent space; c represents the semantic representation of the interference type, Represents interference semantic representation encoder, Interference semantic representation decoder; D KL (·) indicates the calculation of KL divergence, and η is the weight factor of KL divergence.

[0086] The cross-modal feature reconstruction loss l CMFR It means that the interference data features and semantic representation are respectively sent to the interference data feature encoder and the interference semantic representation encoder, and then the interference features in the latent space are cross-input to the decoder of another modality, that is, the interference data features in the latent space are input to the interference semantic representation decoder, and the semantic representation in the latent space is input to the interference data feature decoder to achieve the reconstruction error of cross-modal reconstruction. The cross-modal reconstruction error is defined as the L1 distance between the cross-modal reconstructed data and the input data. The cross-modal feature reconstruction loss l CMFR It is expressed as:

[0087]

[0088] Where N represents the number of training samples in the training set divided from the active interference identification benchmark dataset in step S1.4 according to project requirements, and x i and c i They respectively represent the interference data features and semantic representations of the same type of interference.

[0089] The latent space multimodal feature distribution matching loss l JSDM It represents the distance between the mean and variance of the interfering data feature distribution and the interfering semantic representation distribution, which is used to align the feature distributions of different modes in the latent space and can be expressed as:

[0090]

[0091] Where N represents the number of training samples in the training set divided from the active interference identification benchmark dataset in step S1.4 according to project requirements. and They represent the mean and variance of the latent space interference data features, and denote the mean and variance of the latent space interference semantic representation, respectively, ||·||2 denotes the 2-norm of the matrix, ||·|| F Represents the F-norm of the matrix.

[0092] The multi-class distribution separation loss l MCDD Represents the distribution separation between various types of interference semantic distributions, which is used to enhance the separation of different categories in the latent space and can be expressed as:

[0093]

[0094] H=(N·PW) / N (Formula 10)

[0095] In the formula, is a matrix composed of semantic representations of various interferences in the latent space; N represents the number of training samples in the training set divided from the active interference identification benchmark data set in step S1.4 according to project requirements, is the identity matrix; is a matrix whose elements are all real 1; is the identity matrix; ‖·‖ F is the F-norm of the matrix.

[0096] Step S4, constructing a generalized zero-shot classification model to realize active interference identification.

[0097] like Figure 3 As shown, the generalized zero-shot classification model includes: a training phase and a testing phase.

[0098] Based on the training requirements of the generalized zero-shot classification model, the active interference identification benchmark dataset is divided into the target domain (i.e., unknown class) and the source domain (i.e., known class).

[0099] The total loss of the cross-modal learning model driven by the active interference data and knowledge in step S3.3 is used to train the interference data feature encoder and the interference semantic representation encoder. In the training phase of the generalized zero-shot classification model, the semantic representation c corresponding to the interference signals of the known and unknown classes of the active interference recognition benchmark dataset in step S1.4 is sent to the interference semantic representation encoder as shown in Formula 1 to obtain the latent space semantic representation z (a) , then perform machine learning under the supervision of the binary vector label in step S1.4 to train the generalized zero-shot classifier F GZSL .

[0100] In the test phase of the generalized zero-shot classification model, the interference data samples s of the known or unknown classes corresponding to the training phase, as shown in Formula 1, are sent to the interference feature extraction network to output the interference data feature x, which is then sent to the interference data feature encoder to obtain the latent space interference feature z (j) , using the trained generalized zero-shot classifier F GZSL Interference feature z in latent space (j) Classify and identify active interference.

[0101] In summary, the present invention addresses the few-sample and zero-sample problems faced by active interference identification in practice by constructing a cross-modal learning framework driven by a mixture of active interference data and knowledge. At the same time, the present invention can convert knowledge graphs with high semantic richness into semantic representations, and combine them with data-driven deep learning models to improve the performance of cross-modal learning models driven by a mixture of active interference data and knowledge.

[0102] Although the content of the present invention has been described in detail through the above preferred embodiments, it should be appreciated that the above description should not be considered as a limitation of the present invention. After reading the above content, it will be apparent to those skilled in the art that various modifications and substitutions of the present invention will occur. Therefore, the protection scope of the present invention should be limited by the appended claims.

Claims

1. A method for identifying active interference based on cross-modal learning, characterized in that: The specific steps include: Step S1, constructing an active interference identification benchmark data set; Step S2, clarifying the knowledge modeling form of external knowledge in the active interference ontology model, and converting the active interference knowledge modeling into semantic representation; Step S3, constructing a cross-modal learning model driven by active interference data and knowledge based on the interference data feature VAE model and the interference semantic representation VAE model; Step S4, constructing a generalized zero-shot classification model to realize active interference identification.

2. The active interference identification method based on cross-modal learning according to claim 1, characterized in that: The knowledge modeling form of the external knowledge in the active interference ontology model in step S2 is a knowledge graph.

3. The active interference identification method based on cross-modal learning according to claim 2 is characterized in that: The step S2 further comprises: Step S2.1, construct an active interference ontology model, use knowledge modeling to construct the logical relationship between the required information in the active interference ontology model, and complete the knowledge graph model layer construction; Step S2.2, establishing a knowledge graph pattern layer definition model based on the active interference identification benchmark data set in step S1, and obtaining and storing the attribute values ​​of the attributes possessed by each instance through data simulation; Step S2.3, the active interference ontology model and the attributes of each instance are converted into triples to form a complete active interference knowledge graph; In step S2.4, the active interference knowledge graph is embedded into the model, converted into a vector space, and the semantic representation of various active interference signal types is obtained.

4. The active interference identification method based on cross-modal learning according to claim 3 is characterized in that: The active interference ontology model includes: the generation mechanism, action mechanism, multi-domain characteristics, multi-domain multi-dimensional features, radar technology systems that can interfere, and interference countermeasures or suppression methods of different styles of active interference; wherein, the multi-domain multi-dimensional features are obtained by extracting the features of the active interference identification benchmark data set in step S1; the multi-domain characteristics are obtained by summarizing and condensing the multi-domain multi-dimensional features; the generation mechanism, action mechanism, radar technology systems that can interfere, and interference countermeasures or suppression methods of different styles of active interference are obtained through external knowledge.

5. The active interference identification method based on cross-modal learning according to claim 3 is characterized in that: The knowledge graph embedding model is the TranE model.

6. The active interference identification method based on cross-modal learning according to claim 3 is characterized in that: The step S3 further comprises: Step S3.1, based on the active interference identification benchmark dataset of step S1 and the semantic representation of step S2, construct the input of cross-modal learning driven by active interference data and knowledge hybrid, which can be expressed as: {(s i ,c i ,l i )|=1,2,…M} In the formula, s i and l i represents the data sample and active interference category label of the i-th example in the active interference identification benchmark dataset constructed in step S1, c i represents the first i The semantic representation of the class, M represents the number of instances based on step S2.2; Step S3.2, based on the interference data feature VAE model and the interference semantic representation VAE model, a cross-modal learning model driven by active interference data and knowledge is established; Step S3.3, reconstruct the loss of the interference data feature VAE model and the interference semantic representation VAE model. The total loss l of the cross-modal learning model driven by the active interference data and knowledge mixture is composed of the multimodal reconstruction loss Cross-modal feature reconstruction loss Hidden space multimodal feature distribution matching loss and multi-class distribution separation loss The composition can be expressed as: l=l VAE +a·l CMFR +β·l JSDM +γ·l MCDD Where α, β, and γ represent the weight factors of cross-modal feature reconstruction loss, latent space multimodal feature distribution matching loss, and multi-category distribution separation loss, respectively.

7. The active interference identification method based on cross-modal learning according to claim 6, characterized in that: In step S3.3, the multimodal reconstruction loss l VAE VAE model loss is the interference data feature and interference semantic representation VAE model loss The superposition of can be expressed as: In the formula, x represents the original feature of the interference data, represents the interference data feature encoder, Interference data feature decoder; (j) represents the interference data features in the latent space, z (a) represents the semantic representation in the latent space; c represents the semantic representation of the interference type, Represents interference semantic representation encoder, Interference semantic representation decoder; D KL (·) indicates the calculation of KL divergence, and η is the weight factor of KL divergence.

8. The active interference identification method based on cross-modal learning according to claim 1, characterized in that: The step S1 further comprises: Step S1.1, generating a target signal, various single interference signals, and a noise signal through simulation; and performing additive compounding on the generated various single interference signals to generate a variety of composite interference signals; superimposing the target signal, the single interference signal or the composite interference signal, and the noise signal with different weights to obtain received signals with different signal-to-noise ratios; Step S1.2, performing digital down-conversion on the received signal to obtain a baseband signal, and obtaining frequency domain filter parameters through constant false alarm detection and interference signal parameter estimation; Step S1.3, preprocessing and standardizing the interference samples of the baseband signal to obtain interference identification sample data; Step S1.4, divide the interference identification sample data in step S1.3 into sample labels of single interference signal and composite interference signal types based on the values ​​of the binary vector labels, and obtain an active interference identification benchmark data set.

9. The active interference identification method based on cross-modal learning according to claim 8, characterized in that: In step S1.4, a binary vector label is constructed in the form of a one-dimensional array, where each element in the one-dimensional array represents a single interference signal type, and when the single interference signal type exists in the interference identification sample data, the element in the one-dimensional array takes 1; When the single interference signal type does not exist in the interference identification sample data, the element in the one-dimensional array is set to 0.

10. The active interference identification method based on cross-modal learning according to claim 1, characterized in that: In step S4, the generalized zero-shot classification model includes: a training phase and a testing phase; and according to the training requirements based on the generalized zero-shot classification model, the active interference identification benchmark data set in step S1 is divided into an unknown class and an unknown class; In the training phase, the semantic representation c corresponding to the interference signals of known and unknown classes of the active interference recognition benchmark dataset in step S1 is fed into the interference semantic representation encoder to obtain the latent space semantic representation z (a) , then perform machine learning under the supervision of the binary vector label in step S1.4 to train the generalized zero-shot classifier F GZSL ; In the test phase, the interference data samples s of known or unknown classes corresponding to the training phase are sent to the interference feature extraction network, and the interference data features x are output, which are then sent to the interference data feature encoder to obtain the latent space interference features z. (j) , using the trained generalized zero-shot classifier F GZSL Interference feature z in latent space (j) Classify and identify active interference.

Images