Interference Device Detection Method Based on Knowledge Graph

Abstract

This invention discloses a knowledge graph-based method for detecting interference devices, belonging to the field of interference perception technology. Addressing the problem of heterogeneous, massive, and redundant big data regarding interference information in complex electromagnetic environments, this invention constructs a knowledge graph by building triples based on the relationships between various interference parameters. On this basis, a deep neural network is used to mine the interference knowledge graph information to identify interference devices. The invention first configures the interference knowledge graph, then obtains the embedding vector representation of the interference information to be identified, as well as the interference device vector representation of each interference device in the interference knowledge graph, based on the embedding vector representation of the interference knowledge graph. The similarity between the two is obtained based on a recognition model, and the interference device perception result of the interference information to be identified is obtained based on the interference device corresponding to the maximum similarity. This invention can significantly improve the accuracy of interference device perception.

Description

Technical Field

[0001] This invention belongs to the field of interference sensing technology, specifically relating to a method for sensing interference devices based on knowledge graphs. Background Technology

[0002] The purpose of interference sensing is to obtain situational information about the target, providing a basis for subsequent interference suppression efforts. Most interference sensing focuses on signal detection to identify the presence and type of interference, belonging to signal-level sensing, with limited research on sensing interference devices. Traditional interference source sensing methods primarily differentiate between different interference sources by extracting subtle features (i.e., fingerprint information). The quality of signal feature extraction significantly impacts the complexity and recognition rate of subsequent classifier design, and this approach does not fully utilize spatial information.

[0003] Knowledge graphs are a recently emerging type of large-scale knowledge base that can better characterize the relational attributes of entities. They can also analyze and predict based on behavioral patterns of entities, providing rich characteristic information for related tasks and bringing new technical means to interference sensing in complex electromagnetic environments. While knowledge graphs have been widely applied in various fields, their application in communication interference sensing has not yet been reported, and currently, there is no large-scale knowledge graph for analyzing and processing communication interference information. Summary of the Invention

[0004] This invention provides a knowledge graph-based method for detecting interference devices, which can be used to improve the accuracy of detecting interference devices.

[0005] The technical solution adopted in this invention is a knowledge graph-based method for detecting interference devices, which includes the following steps:

[0006] Step 1, Configure the interference knowledge graph:

[0007] Interference information is collected, and location and signal features are extracted from the historical behavior information of different interference devices to obtain interference information for each interference device. The interference information includes multiple parameters. The interference information is represented in the form of triples (h, r, t) to form an interference knowledge graph. The head entity h is the name of the interference information, the tail entity t is the model of the interference device or the parameter value of different parameters of the interference information, and the relation r represents the relationship between the head entity h and the tail entity t or the attribute of the entity.

[0008] Step 2: The semantic information of the device information of the interfering device in the interference knowledge graph is mapped into vectors through the word vector generation model to obtain the vector representation of the interfering device and input into the recognition model;

[0009] Step 3: Obtain the embedding vector representations of entities containing interfering information in the interfering knowledge graph through the knowledge graph translation model, and store them in the entity embedding vector table;

[0010] Step 4: Collect the interference information to be identified, and query the embedding vector representation of each parameter value in the entity embedding vector table; concatenate the embedding vectors of all parameter values ​​to obtain the embedding vector representation of the interference information and input it into the recognition model.

[0011] Step 5: The recognition model performs feature enhancement on the embedded vector representation of the interference information to obtain the interference feature vector;

[0012] The identification model performs vector transformation on the vector representation of interfering devices to obtain device feature vectors;

[0013] Furthermore, the device feature vector and the interference feature vector have the same vector dimension;

[0014] The identification model calculates the similarity between the interference feature vector and the device feature vector of each interference device, and obtains the matching result of the interference information to be identified based on the interference device corresponding to the maximum similarity.

[0015] Furthermore, the interference information includes location characteristics (including the geographical location of the interference device and its distance from the target) and signal characteristics (including the type, power, and frequency of the interference signal).

[0016] Furthermore, the device information includes the device model and device attribute information.

[0017] Furthermore, in step 2, the word vector generation model used is the Word2vec model.

[0018] Furthermore, in step 5, the identification model uses the inner product method to calculate the similarity between the interference feature vector and the device feature vector.

[0019] Furthermore, the identification model is based on a branch including interference information, a branch including interference devices, and a similarity calculation module;

[0020] The interference information branch consists of a convolutional network layer, a multi-head self-attention network layer, a transformation layer, and at least one fully connected layer, which is used to output the interference feature vector.

[0021] The interference device branch includes a conversion layer and at least one fully connected layer, which are used to output the device feature vector.

[0022] The transformation layer is used to convert multidimensional vectors into one-dimensional vectors.

[0023] The similarity calculation module uses the inner product method to calculate the similarity between the interference feature vector and the device feature vector.

[0024] Furthermore, the convolutional network layer includes a stacked structure of multiple convolutional blocks, wherein each convolutional block sequentially includes at least one convolutional layer, an activation function layer, and a pooling layer.

[0025] Furthermore, the loss function used during training of the recognition model is:

[0026]

[0027] Where, N R Indicates the number of interference information samples. The expression represents the information similarity of different interference information samples calculated by the similarity calculation module. Here, the subscript n represents the number of the interference information sample, the superscript pos represents the positive sample, the superscript neg represents the negative sample, the subscript i represents the number of the negative sample, I represents the number of negative samples, and σ(·) represents the Sigmoid activation function.

[0028] Furthermore, in step 3, the translation model for the knowledge graph adopts the TransR model;

[0029] The TransR model is trained synchronously with the recognition model. The input of the TransR model is the triples in the interference knowledge graph, and all triples belonging to the same interference information are treated as a single interference information sample.

[0030] The embedding vector representations of the parameter values ​​of different parameters of the interference information output by the TransR model are concatenated through a concatenation layer to obtain the embedding vector representation of the current interference information sample, which is then input into the interference information branch of the recognition model.

[0031] Simultaneously, each interference device vector in the interference knowledge graph is used to represent the interference device branch input into the recognition model;

[0032] When training the TransR model and the recognition model simultaneously, the loss function used by the TransR model is:

[0033]

[0034] Among them, L BPR Let N represent the Bayesian personalized ranking loss function. s This represents the number of triplet samples (i.e., the number of triplets input to the TransR model in the current training round). This represents the score of the positive sample. denoted as negative sample score, subscript k indicates triplet sample number, and σ(·) represents the Sigmoid activation function;

[0035] The function for calculating the sample score is: h r=hM r , t r =tM r M r The embedded vector mapping matrix represents the TransR model, denoted by [·]. + Let [x] represent a positive function, where [x] is positive when the independent variable x > 0. + =x; when the independent variable x≤0, [x] + =0.

[0036] The technical solution provided by this invention brings at least the following beneficial effects:

[0037] This invention can be used for real-time perception of jamming devices, effectively improving the perception performance. Furthermore, the jamming device perception model employed is not limited to knowledge representation learning from static knowledge graphs. For real-time perceived data, this model can also randomly embed entities not found in the knowledge base, allowing other entities in the knowledge base to play a role in the jamming information parameters. This avoids the high retraining cost and inability to test in real-time associated with knowledge representation learning from static knowledge graphs. Simultaneously, the CNN network and multi-head self-attention network in the embedding enhancement module of this invention have excellent feature extraction capabilities, accelerating network convergence and improving the recognition performance of jamming devices. Attached Figure Description

[0038] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0039] Figure 1 This is a schematic diagram of the interference knowledge graph displayed in the Neo4j graph database in an embodiment of the present invention;

[0040] Figure 2 This is a schematic diagram of the network structure of the interference device sensing model used in this embodiment of the invention;

[0041] Figure 3 This is a schematic diagram of the TransR model used in this embodiment of the invention;

[0042] Figure 4 This is a schematic diagram of the structure of the convolutional network layer used in an embodiment of the present invention. Detailed Implementation

[0043] To make the objectives, technical solutions, and advantages of the present invention clearer, the embodiments of the present invention will be described in further detail below with reference to the accompanying drawings.

[0044] To address the heterogeneous, massive, and redundant big data problem of interference information in complex electromagnetic environments, this invention constructs triples based on the relationships between various interference parameters, thereby building an Interference Knowledge Graph (IKG). On this basis, deep neural networks are used to mine the interference knowledge graph information to identify interference devices.

[0045] Interference knowledge graph is a domain knowledge graph that integrates multi-source heterogeneous data acquired by sensing nodes to extract the complex relationships between the location features and transmitted signal features of interference sources, thereby enabling situational characterization and reasoning for different interference sources.

[0046] In this embodiment of the invention, the interference knowledge graph describes instance knowledge in the form of triples, namely (head entity, relation, tail entity), denoted as (h, r, t). Here, the head entity h is an entity in the knowledge graph; the tail entity t represents an entity or its attribute value; and the relation r represents the relationship between the head and tail entities or an entity's attribute. A complete triple is called a fact, denoted as F = (h, r, t). From the perspective of the graph model, all entities and attribute values ​​in the interference knowledge graph correspond to nodes in the graph, and the relations between entities or the attributes of entities correspond to edges in the graph, forming a network-like graph data model that organizes scattered interference information data together through structured knowledge.

[0047] Assuming multiple jamming devices exist in space, location and signal features are extracted from the historical behavior information of each device. Location features include the geographical location of the jamming source and its distance to the target; signal features include the type, power, and frequency of the jamming signal. The jamming information is represented as triples, forming an jamming knowledge graph. The storage format of triples in the knowledge graph is shown in Table 1.

[0048] Table 1. Storage format of knowledge graph triples

[0049] Entity 1 (Head Entity) relation Entity 2 (Tail Entity) Interference information name belong jamming device model Interference Information Name power Power value Interference information name frequency Frequency value …… …… ……

[0050] Visualizing the interference knowledge graph within a graph database (such as Neo4j) facilitates querying specific interference devices, thereby retrieving historical situational information for individual devices. The Neo4j graph database provides an example of this visualization. Figure 1 As shown.

[0051] In this embodiment of the invention, the interference device perception model based on the knowledge graph consists of a Knowledge Representation Learning (KRL) module and a recognition model. The KRL module includes a word vector generation model and a translation model based on the knowledge graph. The former is used to obtain the vector representation of the interference device, and the latter is used to obtain the word vector representation of the interference information. In this embodiment, the word vector generation model is preferably the Word2vec model, and the knowledge graph translation model is preferably the TransR model. During training, the triple information in the interference knowledge graph is first converted into embedded representation vectors through the KRL module. Then, the powerful feature extraction capability of the neural network is used to enhance the features of the embedded representation vectors. Finally, the interference device is identified by similarity scoring.

[0052] As one possible implementation, in this embodiment of the invention, the network structure of the interference device sensing model is as follows: Figure 2 As shown, specifically:

[0053] (1) KRL module.

[0054] The KRL module concatenates the interference device model and its corresponding attribute information, and then uses the Word2vec model to map the semantic information into vector form. The interference information and its parameters are stored in the interference knowledge graph as triples. The TransR model is used to embed the structural information contained in the interference information into a low-dimensional vector space, which facilitates subsequent feature extraction and feature enhancement.

[0055] In this embodiment, the TransR model models entities and relations in different vector spaces, namely, an entity space and a relation space, and translates the head and tail entities in the corresponding relation spaces. The TransR model principle diagram is shown below. Figure 3 As shown.

[0056] For a triplet entity pair (h, r, t) in a distracting knowledge graph, each head entity h and tail entity t are linked through a relation matrix M. r The projection matrix (i.e., the projection matrix) is projected onto the relation space to obtain h. r and t r By continuously adjusting the vectors h, r, t and the projection matrix M r , making h r +r and t r Getting closer and closer, the formula is as follows:

[0057] h r =hM r ,t r =tM r (1)

[0058]

[0059]

[0060] Where L represents the loss function of the TransR model, γ is a hyperparameter of the loss function used to characterize the margin distance, S represents the set of all positive triples, S' represents the set of all negative triples, [·] + Let [x] represent a positive function, where [x] is positive when the independent variable x > 0. + =x; when the independent variable x≤0, [x] + =0.

[0061] To maximize the score difference between positive and negative samples, the TransR model uses the Bayesian Personalized Ranking (BPR) loss function, expressed as follows:

[0062]

[0063] Among them, L BPR Let N be the BPR loss function. s The number of triplet samples. The score for the positive sample. σ is the score of the negative sample, where the sample score is calculated using formula (2). σ(·) is the Sigmoid activation function.

[0064] During training, the embedding vector representations of the parameter values ​​of different parameters of the interference information output by the TransR model are concatenated through a concatenation layer to obtain the embedding vector representation of the current interference information sample, which is then input into the recognition model.

[0065] Once training is complete, the embedding vector representations of entities related to interference information in the interference knowledge graph can be obtained based on the output of the TransR model and stored in the entity embedding vector table. This allows the embedding vector representation of the interference information to be identified to be obtained by looking up the table. Specifically, for the interference information to be identified, the embedding vector representation of each parameter value in the interference information is queried in the entity embedding vector table; the embedding vectors of all parameter values ​​are concatenated to obtain the embedding vector representation of the interference information.

[0066] (2) Identification model.

[0067] The identification model comprises two branches: an interference information branch (feature enhancement branch) and an interference device branch (vector transformation branch), as well as a similarity calculation module. The interference information branch enhances the embedded vector representation of the interference information to obtain the interference feature vector. The interference device branch transforms the vector representation of the interference device to obtain the device feature vector; the device feature vector and the interference feature vector have the same vector dimension. The similarity calculation module uses an inner product method to calculate the similarity between the device feature vector and the interference feature vector.

[0068] Assuming each set of interference information in the scenario has K parameters, and the embedding dimension of each entity and relation is D, during training, the embedding vectors of each interference information parameter are concatenated head to tail, and each set of interference information is represented as an embedding vector of dimension K*D.

[0069] To improve the accuracy of recognition, feature enhancement is required on the concatenated embedding vector. The feature enhancement branch uses convolutional neural networks (CNNs) and multi-head self-attention networks.

[0070] CNN networks possess powerful feature extraction capabilities, enabling feature extraction and enhancement from concatenated embedded representation vectors. The CNN network used in this embodiment consists of multiple convolutional layers and pooling layers, employing Leaky ReLU as the activation function. The network structure is as follows: Figure 4 As shown.

[0071] In this embodiment of the invention, the convolutional network layer used is a stacked structure of multiple convolutional blocks. Each convolutional block sequentially includes at least one convolutional layer (Conv2D), an activation function layer (e.g., Leaky ReLU activation function), and a pooling layer. The pooling method of the pooling layer can be either average pooling (AvePool) or max pooling (MaxPool). In this embodiment of the invention, the convolutional network layer is used for feature extraction. The network parameters and structure can be adjusted according to the actual application scenario and requirements. This embodiment of the invention does not impose specific limitations.

[0072] The expression for the Leaky ReLU activation function is:

[0073]

[0074] Where x represents the independent variable of the Leaky ReLU activation function, and α is the scaling factor.

[0075] Self-attention mechanisms enable networks to extract correlations between different features and highlight the importance of certain features. The basic process is as follows: given a query vector q... j Calculate the key vector k i Attention weight α i And based on this weight, the value vector v i Perform weighted average output z j The expression for calculating Self-Attention is:

[0076]

[0077] Here, x is the input to Self-Attention, consisting of the embedding representation vector processed by the KRL module. W K W V W Q b K b V b Q These are the weights and bias parameters that need to be learned to generate the Key vector, Value vector, and Query vector, respectively. K, V, and Q are matrices composed of the Key vector, Value vector, and Query vector obtained through different spatial transformations, and z is the output after Self-Attention weighted averaging.

[0078] Multi-head self-attention mechanisms can extend the model's focus on different perturbation information parameters, enabling the model to learn relevant information in different representation subspaces. The approach involves projecting the input matrix x onto different representation subspaces, i.e., using multiple sets of K, V, Q matrices to describe the input matrix X, and finally concatenating the weighted average result z from multiple different subspaces as the output.

[0079] The Flatten layer flattens the multidimensional embedded vector features of the interference information parameters and the interference device model into a one-dimensional vector, and then extracts the correlation between these features through a fully connected layer (Dense layer) to obtain an output vector of the same dimension.

[0080] The similarity calculation module uses the inner product method to calculate the similarity probability score of each set of interference information to each interference device. The expression is:

[0081]

[0082] in, The similarity probability score is represented by the output of the similarity calculation module (i.e., Softmax). The interference feature vector of the i-th interference information is represented by e. i The device feature vector of the j-th interfering device is represented as e. jVector e i With e j The same dimension is defined as D. o d represents the dimension index of the vector, i.e. Representing vector e respectively i e j The d-th dimension.

[0083] As one possible implementation, in this embodiment of the invention, when the interference device perception model trains the network parameters, the preferred loss function is the improved BPR loss function, which can further maximize the positive sample score and minimize all negative sample scores, as expressed in:

[0084]

[0085] Where, N R The number of interference information samples. The score for the positive sample. Let be the score of the i-th negative sample, n be the sample number, I be the number of negative samples, and σ(·) be the Sigmoid activation function.

[0086] The knowledge graph-based interference device sensing method provided in this embodiment of the invention is as follows:

[0087] First, an interference knowledge graph is configured. Then, a word vector generation model is used to map the semantic information of the device information of interfering devices in the interference knowledge graph into vectors, obtaining the vector representation of the interfering devices, which is then input into the trained recognition model. Next, the TransR model is used to obtain the embedding vector representations of entities related to the interference information in the interference knowledge graph and store them in an entity embedding vector table. For the interference information to be identified, the embedding vector representation of each parameter value in the interference information is queried from the entity embedding vector table. The embedding vectors of all parameter values ​​are concatenated to obtain the embedding vector representation of the interference information, which is then input into the recognition model. The interference information branch of the recognition model performs feature enhancement on the embedding vector representation of the interference information to obtain the interference feature vector, which is then input into the similarity calculation module. The interference device branch performs vector transformation on the vector representation of the interference devices to obtain the device feature vector, which is then input into the similarity calculation module. Finally, the similarity calculation module calculates the similarity between the interference feature vector and the device feature vector of each interfering device. Based on the interfering device corresponding to the maximum similarity, the matching result of the interference information to be identified is obtained.

[0088] To further verify the interference device sensing method based on knowledge graphs provided in this embodiment of the invention.

[0089] In the simulation application scenario, the number of interference device types is set to N, with models a, b, and c, the total number of interference information is M, and the number of parameters included in each group of interference information is K, such as interference type, frequency, and interference distance. The corresponding data range is set according to the different characteristics of the interference devices.

[0090] The interference information is divided into training and testing sets, and set in a certain proportion. The loss function of the TransR module and the recognition model can use the Adam optimizer with a learning rate of α. The training process is sequential training.

[0091] For interference information data obtained in real-time, interference devices can be identified based on the saved embedding vectors and recognition models. Test data is randomly generated within a range of N interference device models. If an entity exists in the knowledge base, it is represented by the corresponding embedding vector in the knowledge base; otherwise, it is represented by a random embedding vector using the Word2vec model. This vector is input into the trained interference device perception model for real-time interference device perception. In the set simulation parameters, the recognition accuracy of the test set reaches 99.8% when the number of iterations is 10. The CNN network and multi-head self-attention network of the embedding enhancement module have excellent feature extraction capabilities, which can accelerate the convergence speed of the network and improve the recognition performance of the interference device perception model. This model is not limited to knowledge representation learning from static knowledge graphs. For real-time perceived data, the model randomly embeds entities not in the knowledge base, allowing the remaining entities in the interference information parameters that exist in the knowledge base to play a role. This avoids the high retraining cost and inability to test in real-time of knowledge representation learning from static knowledge graphs, enabling real-time perception of interference devices.

[0092] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

[0093] The above descriptions are merely some embodiments of the present invention. Those skilled in the art can make various modifications and improvements without departing from the inventive concept of the present invention, and these all fall within the scope of protection of the present invention.

Claims

1. A knowledge graph-based method for detecting interference devices, characterized in that, Includes the following steps: Step 1, Configure the interference knowledge graph: Interference information is collected, and location and signal features are extracted from the historical behavior information of different interference devices to obtain interference information for each interference device. The interference information includes multiple parameters. The interference information is represented in the form of triples (h, r, t) to form an interference knowledge graph. The head entity h is the name of the interference information, the tail entity t is the model of the interference device or the parameter value of different parameters of the interference information, and the relation r represents the relationship between the head entity h and the tail entity t or the attribute of the entity. Step 2: The semantic information of the device information of the interfering device in the interference knowledge graph is mapped into vectors through the word vector generation model to obtain the vector representation of the interfering device and input into the recognition model; Step 3: Obtain the embedding vector representations of entities containing interfering information in the interfering knowledge graph through the knowledge graph translation model, and store them in the entity embedding vector table; Step 4: Collect the interference information to be identified, and query the embedding vector representation of each parameter value in the entity embedding vector table; concatenate the embedding vectors of all parameter values ​​to obtain the embedding vector representation of the interference information and input it into the recognition model. Step 5: The recognition model performs feature enhancement on the embedded vector representation of the interference information to obtain the interference feature vector; The identification model performs vector transformation on the vector representation of interfering devices to obtain device feature vectors; Furthermore, the device feature vector and the interference feature vector have the same vector dimension; The identification model calculates the similarity between the interference feature vector and the device feature vector of each interference device, and obtains the matching result of the interference information to be identified based on the interference device corresponding to the maximum similarity. The identification model includes an interference information branch, an interference device branch, and a similarity calculation module. The interference information branch consists of a convolutional network layer, a multi-head self-attention network layer, a transformation layer, and at least one fully connected layer, which is used to output the interference feature vector. The interference device branch includes a conversion layer and at least one fully connected layer, which are used to output the device feature vector. The transformation layer is used to convert multidimensional vectors into one-dimensional vectors. The similarity calculation module uses the inner product method to calculate the similarity between the interference feature vector and the device feature vector.

2. The method as described in claim 1, characterized in that, The interference information includes location features and signal features.

3. The method as described in claim 1, characterized in that, The equipment information includes the equipment model and equipment attribute information.

4. The method as described in claim 1, characterized in that, In step 2, the word vector generation model used is the Word2vec model.

5. The method as described in claim 1, characterized in that, In step 5, the identification model uses the inner product method to calculate the similarity between the interference feature vector and the device feature vector.

6. The method as described in claim 1, characterized in that, The convolutional network layer comprises a stacked structure of multiple convolutional blocks, wherein each convolutional block comprises at least one convolutional layer, an activation function layer, and a pooling layer in sequence.

7. The method as described in claim 1 or 6, characterized in that, The loss function used during training of the recognition model is: Where, N R Indicates the number of interference information samples. The expression represents the information similarity of different interference information samples calculated by the similarity calculation module. Here, the subscript n represents the number of the interference information sample, the superscript pos represents the positive sample, the superscript neg represents the negative sample, the subscript i represents the number of the negative sample, I represents the number of negative samples, and σ(·) represents the Sigmoid activation function.

8. The method as described in claim 1, characterized in that, In step 3, the translation model for the knowledge graph adopts the TransR model; The TransR model and the recognition model are trained sequentially. The input of the TransR model is the triples in the interference knowledge graph, and all triples belonging to the same interference information are taken as a interference information sample. The embedding vector representations of the parameter values ​​of different parameters of the interference information output by the TransR model are concatenated through a concatenation layer to obtain the embedding vector representation of the current interference information sample, which is then input into the interference information branch of the recognition model. Simultaneously, each interference device vector in the interference knowledge graph is used to represent the interference device branch input into the recognition model; When training the TransR model and the recognition model sequentially, the loss function used by the TransR model is: Among them, L BPR Let N represent the Bayesian personalized ranking loss function. s Indicates the number of triplet samples. This represents the score of the positive sample. denoted as negative sample score, subscript k indicates triplet sample number, and σ(·) represents the Sigmoid activation function; The function for calculating the sample score is: h r =hM r , t r =tM r M r The embedded vector mapping matrix represents the TransR model, denoted by [·]. + Let represent a positive function, where [x] is positive when the independent variable x > 0. + =x; when the independent variable x≤0, [x] + =0.

9. The method as described in claim 1, characterized in that, Also includes: The interference knowledge graph is visualized in a graph database.

Images