Radar purpose reasoning method based on knowledge graph

Abstract

The invention discloses a radar purpose reasoning method based on a knowledge graph, which comprises the following steps of: firstly, detecting original radar signal parameter data, constructing a radar purpose knowledge graph, accurately expressing semantic association between radar signal parameters and purposes by utilizing the structural characteristics of the knowledge graph, extracting structured triple knowledge in the knowledge graph, and calculating the purpose of the radar purpose. Then, a multi-head attention mechanism is introduced to capture a global dependency relationship among radar signal multi-dimensional parameters, and the sensitivity of the model to fine parameter differences is enhanced through a convolution module, so that the modeling capability for complex combination parameters is effectively improved, and higher reasoning accuracy is obtained; and finally, matching is carried out through prediction vector and candidate tail entity embedding, and reasoning of the radar purpose is realized. The method provided by the invention solves the problem that the complex relationship between the radar signal parameter and the purpose is difficult to model effectively in the existing method, and has the advantages of accuracy, high efficiency and suitability for a reconnaissance party to receive numerous radar signal parameters and complex combination.

Description

Technical Field

[0001] This invention belongs to the field of radar electronic countermeasures, specifically involving a radar application reasoning method based on knowledge graphs. Background Technology

[0002] Radar purpose, or the radar's behavioral pattern, is typically achieved through a combination of multiple operating modes arranged in a specific time sequence. Compared to operating modes alone, this combination method more comprehensively characterizes the radar's tactical functionalities. Radar purpose inference refers to the process by which electronic reconnaissance units deduce the radar's specific purpose based on the parameters of the received radar signals. This method plays a crucial role in key tasks such as electronic warfare, situational awareness, and target identification.

[0003] Faced with the increasingly complex electromagnetic environment, the ever-advancing radar technology, and the rich and varied radar signal parameters, reconnaissance units receive numerous and complex combinations of radar signal parameters. Existing methods struggle to effectively model the complex relationships between radar signal parameters and their intended uses, making modeling and reasoning about radar uses increasingly difficult for electronic reconnaissance equipment. Therefore, existing algorithms are not suitable for radar use reasoning in the current environment. Knowledge Graph (KG), as a technology system integrating structured semantic representation and relational reasoning capabilities, offers a new solution for radar use reasoning. By organizing radar entities, signal features, and their semantic associations into a graph structure, it not only systematically expresses the attributes of radar uses and the relationships between them but also achieves unified modeling of multi-source heterogeneous radar parameter information. By constructing a graph structure of entities and relationships, Knowledge Graph endows the model with powerful reasoning capabilities, enabling it to integrate known knowledge and provide a more structured and semantically rich knowledge foundation for radar use reasoning.

[0004] Solving the problem of radar application inference mainly relies on signal structure modeling and parameter rule matching. The paper "Ma, Y., Li, Y., Zhu, M., et al.: 'Radar working mode recognition based on hierarchical feature representation and clustering'. Proc. IET Int. Radar Conf., Sep. 2020, pp. 1629-1633" proposes an intelligent algorithm based on hierarchical feature representation and unsupervised clustering, which can alleviate the dependence on manual feature extraction to some extent and improve adaptability to diverse signals, effectively realizing the identification of multifunctional radar working modes. However, the generalization ability and robustness to noise in complex electromagnetic environments still need further improvement.

[0005] With the development of machine learning and knowledge graph technology, knowledge graph reasoning built using deep networks has gradually become popular. The paper "PAN Yudai, LIU Jun, ZHAO Tianzhe, et al. Context-Aware Commonsense Knowledge Graph Reasoning with Path-Guided Explanations. IEEE Transactions on Knowledge and Data Engineering, 2024, 36(8): 3725-3738" constructs a CoRPe model through path-guided context, combining semantic information with graph structure information, and effectively improves the representation ability and interpretability of commonsense knowledge graph reasoning by utilizing the collaborative optimization of pre-trained language models and graph convolution. However, its path generation is highly dependent, computationally complex, and resource-intensive in the training and reasoning process. Nowadays, the electromagnetic environment is becoming increasingly complex, and the adaptability and intelligence of modern radar systems are constantly improving. The radar signal parameters received by the reconnaissance side are numerous and complex in combination. Current knowledge graph embedding models are difficult to handle the "many-to-one" relationship between entities. Therefore, radar application reasoning based on knowledge graphs is of great significance. Summary of the Invention

[0006] To address the aforementioned technical problems, this invention proposes a knowledge graph-based radar application reasoning method. Utilizing the structured characteristics of knowledge graphs, it accurately expresses the semantic relationships between radar signal parameters and applications. A multi-head attention mechanism is introduced to capture the global dependencies between multi-dimensional radar signal parameters, and a convolutional module enhances the model's sensitivity to subtle parameter differences, thereby effectively improving the modeling capability for complex combinations of parameters and achieving accurate reasoning of radar applications.

[0007] The technical solution adopted in this invention is: a radar application reasoning method based on knowledge graphs, the specific steps of which are as follows:

[0008] S1. Detect raw radar signal parameter data and construct a knowledge graph of radar applications;

[0009] The radar signal parameters include: carrier frequency (CF), pulse width (PW), and multiple frequency (PRI). The construction of the radar application knowledge graph includes: entity and relationship definition and visualization.

[0010] First, based on the ability of signal parameters to represent the application of radar, three parameters—carrier frequency, pulse width, and multiple cycles—are selected as modeling attributes. In the knowledge graph, radar signal samples are defined as head entities, and specific applications and their corresponding parameter values ​​are defined as tail entities. Furthermore, statistical features of carrier frequency, pulse width, and multiple cycles are extracted for modeling.

[0011] The statistical features include: median, maximum, and minimum. The formal expression of the knowledge graph G is as follows:

[0012] G=(V,R,T) (1)

[0013] Where V represents the entity set, R represents the relation set, and T represents the triple set.

[0014] S2. Extract structured triplet knowledge from the knowledge graph in step S1;

[0015] Extract structured triples from the knowledge graph in step S1, including head entity, relation, and tail entity. The extracted knowledge graph triples will then be in the form of (h, r, t).

[0016] In this context, the head entity h represents a radar signal sample, the relationship r represents a parameter type or application relationship, and the tail entity t represents the corresponding parameter value or application category.

[0017] S3. Map the triplet knowledge obtained in step S2 into a vector representation with fixed dimensions;

[0018] First, entities and relations are uniquely numbered using a dictionary mapping method, resulting in a set of entity and relation numbers. Then, the head entity, relation, and tail entity in the triples are converted into corresponding index sequences. Next, an embedding layer is used to map the obtained indices into low-dimensional dense vector representations, achieving a continuous representation of triple knowledge. Finally, the original discrete symbolic triples are uniformly mapped to a fixed-dimensional vector space.

[0019] S4. Construct a radar application reasoning network;

[0020] The radar application inference network includes: a multi-head attention mechanism module, a convolution module, and a scoring module.

[0021] The multi-head attention mechanism module includes: an input embedding layer, a stacked Transformer encoder layer, and an output layer.

[0022] Each Transformer encoder layer consists of: a multi-head self-attention layer, a feedforward fully connected layer, a residual connection, and a layer normalization layer.

[0023] The convolutional module includes: one input reshaping layer, one two-dimensional batch normalization layer, one two-dimensional convolutional layer, one convolutional output normalization layer, one ReLU activation function, one Dropout layer, one flattening layer, two fully connected layers, and one layer normalization layer.

[0024] S5. Use the vector obtained in step S3 as the input to the radar purpose inference network constructed in step S4, and train the radar purpose inference network.

[0025] The head entity and relation vector obtained in step S3 are input into the radar application inference network for forward propagation to obtain the prediction score vector and calculate the loss function value; then the model parameters are updated using the backpropagation algorithm based on stochastic gradient descent; by repeatedly performing the forward and backpropagation process, the loss function gradually decreases and tends to stabilize until the training converges, and the trained radar application inference network is obtained.

[0026] Among them, for the score vector s i Apply the Softmax function to obtain the predicted distribution p of the tail entity. ik The expression is as follows:

[0027]

[0028] Among them, s ik =x i ·e k Let x represent the i-th sample. i With the k-th entity e k The matching score, where |U| represents the total number of entities.

[0029] The radar application inference network uses a sample average loss function L, expressed as follows:

[0030]

[0031] Where N represents the number of samples, t i L represents the number of correct labels for the i-th sample. i Let represent the cross-entropy loss of the i-th sample.

[0032] S6. Based on the radar knowledge graph application inference network obtained in step S5, the predicted score vector obtained in step S5 is matched with the candidate tail entity embedding to realize the inference of radar application.

[0033] Furthermore, step S4 is specifically as follows:

[0034] S41, Multi-head Attention Mechanism Module;

[0035] Given an input sequence First, the input is mapped to query Q, key K, and value V using linear mapping. The resulting linear mapping relation vector expression is as follows:

[0036]

[0037] in, Let v = 2 represent the concatenation of the two embedding vectors: the head entity and the relation. Let d represent the embedding dimension. Q i K i V i These represent the Q, V, and K linear mapping vectors of the i-th attention head, respectively. Let d represent the parameter matrix of the i-th attention head, h represent the number of attention heads, and d = h × d k The expression for each attention head is calculated as follows:

[0038]

[0039] Where, d k The embedding dimension of each attention head is represented by the number of heads. i Let represent the i-th attention head, and T represent the transpose operation.

[0040] The outputs of multiple attention heads are concatenated and then linearly mapped to obtain the overall output of the multi-head attention mechanism. The expression is as follows:

[0041] MultiHead(X)=Concat(head1,…,head h )·W O (6)

[0042] Among them, W O This indicates the output mapping matrix.

[0043] The multi-head attention mechanism module uses a stacked Transformer encoder as its structural basis, with a total of l Transformer encoder layers. Each Transformer encoder layer includes a multi-head attention layer and a feedforward fully connected layer, and the information flow is stably propagated through a residual connection and a normalization layer. The calculation process of the l-th Transformer encoder is expressed as follows:

[0044] Z (l) =LayerNorm(X (l-1) +MultiHead(X (l-1) (7)

[0045] X (l) =LayerNorm(Z) (l) +FFN(Z (l) (8)

[0046] Among them, Z (l) X represents the output of the l-th multi-head attention layer. (l)This represents the output of the l-th Transformer encoder, and FFN(·) represents the feedforward network, which consists of two fully connected layers, as shown in the following expression:

[0047] FFN(Z)=ReLU(ZW1+b1)W2+b2 (9)

[0048] Where Z represents the output of the multi-head attention layer, W1 and W2 represent the weight matrices of the linear transformation, b1 and b2 represent the bias terms of the linear transformation, and finally, the outputs X of each Transformer encoder layer are stacked. (l) As a fusion representation of head entities and relations.

[0049] S42, Convolutional module;

[0050] The fused representation X is obtained in step S41. (l) Subsequently, a spatial awareness module is introduced to enhance the ability to represent local features. First, the fused representation is flattened into a vector and then reshaped into a two-dimensional feature map P, as shown in the following expression:

[0051] P = reshape(vec(X) (l) (10)

[0052] The input is then fed into a two-dimensional convolutional neural network, where local convolutional kernels extract the latent spatial structure patterns between entities and relationships, as shown in the following expression:

[0053] F=Dropout(ReLU(BN(Conv2D(P)))) (11)

[0054] After the obtained convolutional feature map F is flattened into a vector f, it is sequentially passed through two fully connected layers to compress the representation dimension and improve the expressive power. The process is expressed as follows:

[0055] x i =LayerNorm(ReLU(fW3+b3)W4+b4) (12)

[0056] Finally, what was obtained This represents the final fused representation vector, which incorporates the deep semantics and local interaction information of the head entities and relationships. This vector is used to match and score all candidate tail entities.

[0057] S43, Scoring Module;

[0058] Based on step S42, after completing the semantic fusion of the head entity and relation embeddings, convolutional and fully connected layer modules are further introduced to mine spatial features. The resulting fused representation vector is input into the scoring module to evaluate its matching degree with all candidate tail entities. Let the entity set be U, and its embedding matrix be denoted as E. By calculating the dot product of the fused representation and the entity embedding, the score vector s of the i-th sample for all entities is obtained accordingly. i The expression is as follows:

[0059]

[0060] Among them, s ik =x i ·e k Let x represent the i-th sample. i With the k-th entity e k The matching score, where |U| represents the total number of entities.

[0061] The beneficial effects of this invention are as follows: The method of this invention first receives raw radar signal parameter data, constructs a radar usage knowledge graph, and then utilizes the structured characteristics of the knowledge graph to accurately express the semantic relationship between radar signal parameters and usage. It extracts structured triple knowledge from the knowledge graph, then introduces a multi-head attention mechanism to capture the global dependencies between multi-dimensional radar signal parameters. Furthermore, it enhances the model's sensitivity to subtle parameter differences through convolutional modules, thereby effectively improving the modeling ability for complex parameter combinations and achieving higher inference accuracy. Finally, it matches the predicted vector with candidate tail entity embeddings to achieve inference about the radar usage. This invention solves the problem of existing methods' difficulty in effectively modeling the complex relationship between radar signal parameters and usage, and has the advantages of accuracy, efficiency, and applicability to battlefield reconnaissance receiving radar signal parameters with numerous and complex combinations. Attached Figure Description

[0062] Figure 1 This is a flowchart of a radar application reasoning method based on knowledge graphs according to the present invention.

[0063] Figure 2 This is a schematic diagram of the knowledge graph of a single radar signal sample based on the Neo4j graph database in an embodiment of the present invention.

[0064] Figure 3 This is a schematic diagram of the radar application inference network in an embodiment of the present invention.

[0065] Figure 4 This is a schematic diagram of a knowledge graph illustrating some radar applications in an embodiment of the present invention.

[0066] Figure 5 This is a schematic diagram of the radar application reasoning results under a simulated short sequence (dataset A) in an embodiment of the present invention.

[0067] Figure 6 This is a schematic diagram of the radar application reasoning results under a simulated long sequence (dataset B) in an embodiment of the present invention.

[0068] Figure 7 This is a schematic diagram of experimental results under different signal-to-noise ratios in an embodiment of the present invention.

[0069] Figure 8 This is a schematic diagram of experimental results under different embedding dimensions in the embodiments of the present invention.

[0070] Figure 9 This is a schematic diagram illustrating the experimental results of different numbers of multi-head attention heads in an embodiment of the present invention. Detailed Implementation

[0071] The method of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0072] like Figure 1 The flowchart of a radar application reasoning method based on knowledge graphs according to the present invention is shown below. The specific steps are as follows:

[0073] S1. Detect raw radar signal parameter data and construct a knowledge graph of radar applications;

[0074] The radar signal parameters include: carrier frequency (CF), pulse width (PW), and multiple cycles (PRI). The construction of the radar application knowledge graph includes: entity and relationship definition and visualization.

[0075] First, based on the ability of signal parameters to represent the application of radar, three parameters—carrier frequency, pulse width, and multiple cycles—are selected as modeling attributes. In the knowledge graph, radar signal samples are defined as head entities, and specific applications and their corresponding parameter values ​​are defined as tail entities. Furthermore, statistical features of carrier frequency, pulse width, and multiple cycles are extracted for modeling.

[0076] The statistical features include: median, maximum, and minimum values. Since these parameters typically exhibit a time-varying sequence, their median, maximum, and minimum values ​​need to be extracted for modeling. The formal expression of the knowledge graph G is as follows:

[0077] G=(V,R,T) (1)

[0078] Where V represents the entity set, R represents the relation set, and T represents the triple set. Figure 2 This is a schematic diagram of the knowledge graph of a single radar signal sample based on the Neo4j graph database in this embodiment.

[0079] S2. Extract structured triplet knowledge from the knowledge graph in step S1;

[0080] Extract structured triples from the knowledge graph in step S1, including head entity, relation, and tail entity. The extracted knowledge graph triples will then be in the form of (h, r, t).

[0081] In this table, the head entity h represents a radar signal sample, the relation r represents a parameter type or application relationship, and the tail entity t represents the corresponding parameter value or application category. Table 1 shows some of the triples extracted from the radar application knowledge graph in this embodiment.

[0082] Table 1

[0083] Head Entity relation Tail entity Signal Sample 1 use Radar Application 1 Signal Sample 1 Repeated modulation type Group change Signal Sample 1 Carrier frequency modulation type fixed Signal Sample 1 Carrier median 9700 Signal Sample 1 Maximum value of multiple cycles 400

[0084] S3. Map the triplet knowledge obtained in step S2 into a vector representation with fixed dimensions;

[0085] First, entities and relations are uniquely numbered using a dictionary mapping method, resulting in a set of entity and relation numbers. Then, the head entity, relation, and tail entity in the triples are converted into corresponding index sequences. Next, an embedding layer is used to map the obtained indices into low-dimensional dense vector representations, achieving a continuous representation of triple knowledge. Finally, the original discrete symbolic triples are uniformly mapped to a fixed-dimensional vector space, providing input features for the subsequent training and inference of the radar application reasoning network.

[0086] S4. Construct a radar application reasoning network;

[0087] like Figure 3 As shown, the radar application inference network includes: a multi-head attention mechanism module, a convolution module, and a scoring module. In the figure, h1, h2, t1, t2, r1, and r2 represent examples of triple knowledge, namely, head entity 1-relation 1-tail entity 1, and head entity 2-relation 2-tail entity 2.

[0088] The multi-head attention mechanism module includes: an input embedding layer, a stacked Transformer encoder layer, and an output layer.

[0089] Each Transformer encoder layer consists of: a multi-head self-attention layer, a feedforward fully connected layer, a residual connection, and a layer normalization layer.

[0090] The convolutional module includes: one input reshaping layer, one two-dimensional batch normalization layer, one two-dimensional convolutional layer, one convolutional output normalization layer, one ReLU activation function, one Dropout layer, one flattening layer, two fully connected layers, and one layer normalization layer.

[0091] First, a radar application knowledge graph is constructed based on radar signal parameters, and structured triplet knowledge is extracted from it. Second, the input head entities and relations are mapped to fixed-dimensional vector representations through an embedding layer. Then, a multi-head attention mechanism is used to stack the entity and relation vectors and input them into a Transformer encoder to fully explore their high-order semantic dependencies. Subsequently, this joint representation is reshaped into a two-dimensional tensor and input into a convolutional neural network to extract its local interaction features. Further nonlinear transformation and feature fusion are performed through a fully connected network, ultimately outputting a matching score with the tail entity embedding.

[0092] S5. Use the vector obtained in step S3 as the input to the radar purpose inference network constructed in step S4, and train the radar purpose inference network.

[0093] The head entity and relation vector obtained in step S3 are input into the radar application inference network for forward propagation to obtain the prediction score vector and calculate the loss function value; then the model parameters are updated using the backpropagation algorithm based on stochastic gradient descent; by repeatedly performing the forward and backpropagation process, the loss function gradually decreases and tends to stabilize until the training converges, and the trained radar application inference network is obtained.

[0094] Among them, for the score vector s i Apply the Softmax function to obtain the predicted distribution p of the tail entity. ik The expression is as follows:

[0095]

[0096] Among them, s ik =x i ·e k Let x represent the i-th sample. i With the k-th entity e k The matching score, where |U| represents the total number of entities.

[0097] To measure overall performance, the sample average loss function L is used, and its expression is as follows:

[0098]

[0099] Where N represents the number of samples, t i L represents the number of correct labels for the i-th sample. i Represents the cross-entropy loss of the i-th sample.

[0100] S6. Based on the radar knowledge graph application inference network obtained in step S5, the predicted score vector obtained in step S5 is matched with the candidate tail entity embedding to realize the inference of radar application.

[0101] In this embodiment, step S4 is specifically as follows:

[0102] S41, Multi-head Attention Mechanism Module;

[0103] Given an input sequence First, the input is mapped to query Q, key K, and value V using linear mapping. The resulting linear mapping vector expression is as follows:

[0104]

[0105] in, Let v = 2 represent the concatenation of the two embedding vectors: the head entity and the relation. Let d represent the embedding dimension. Q i K i V i These represent the Q, V, and K linear mapping vectors of the i-th attention head, respectively. Let d represent the parameter matrix of the i-th attention head, h represent the number of attention heads, and d = h × d k The expression for each attention head is calculated as follows:

[0106]

[0107] Where, d k The embedding dimension of each attention head is represented by the number of heads. i Let represent the i-th attention head, and T represent the transpose operation.

[0108] The outputs of multiple attention heads are concatenated and then linearly mapped to obtain the overall output of the multi-head attention mechanism. The expression is as follows:

[0109] MultiHead(X)=Concat(head1,...,head h )·W O (6)

[0110] Among them, W O The output mapping matrix is ​​represented by the multi-head attention mechanism, which enables the model to capture semantic associations of different dimensions in parallel in multiple subspaces, thereby improving representation and generalization capabilities.

[0111] To further enhance the depth and nonlinear modeling capabilities of knowledge representation learning, the multi-head attention mechanism module employs a stacked Transformer encoder as its structural foundation, consisting of l layers. Each Transformer encoder layer includes a multi-head attention layer and a feedforward fully connected layer, with stable information flow propagation achieved through residual connections and a normalization layer. The computational process of the l-th Transformer encoder is expressed as follows:

[0112] Z (l) =LayerNorm(X (l-1) +MultiHead(X (l-1) (7)

[0113] X (l) =LayerNorm(Z) (l) +FFN(Z (l) (8)

[0114] Among them, Z (l) X represents the output of the l-th multi-head attention layer. (l) This represents the output of the l-th Transformer encoder, and FFN(·) represents the feedforward network, which consists of two fully connected layers, as shown in the following expression:

[0115] FFN(Z)=ReLU(ZW1+b1)W2+b2 (9)

[0116] Where Z represents the output of the multi-head attention layer, W1 and W2 represent the weight matrices of the linear transformation, b1 and b2 represent the bias terms of the linear transformation, and finally, the outputs X of each Transformer encoder layer are stacked. (l) As a fusion representation of head entities and relationships, the multi-head attention mechanism module effectively models global semantic dependencies while capturing local interactions, providing a more discriminative feature foundation for subsequent convolutional operations.

[0117] S42, Convolutional module;

[0118] The fused representation X is obtained in step S41. (l) Subsequently, a spatial awareness module is introduced to enhance the ability to represent local features. First, the fused representation is flattened into a vector and then reshaped into a two-dimensional feature map P, as shown in the following expression:

[0119] P = reshape(vec(X) (l) (10)

[0120] The input is then fed into a two-dimensional convolutional neural network, where local convolutional kernels extract the latent spatial structure patterns between entities and relationships, as shown in the following expression:

[0121] F=Dropout(ReLU(BN(Conv2D(P)))) (11)

[0122] After the obtained convolutional feature map F is flattened into a vector f, it is sequentially passed through two fully connected layers to compress the representation dimension and improve the expressive power. The process is expressed as follows:

[0123] x i =LayerNorm(ReLU(fW3+b3)W4+b4) (12)

[0124] Finally, what was obtained This represents the final fused representation vector, which incorporates the deep semantics and local interaction information of the head entities and relationships. This vector is used to match and score all candidate tail entities.

[0125] S43, Scoring Module;

[0126] Based on step S42, after completing the semantic fusion of the head entity and relation embeddings, convolutional and fully connected layer modules are further introduced to mine spatial features. The resulting fused representation vector is input into the scoring module to evaluate its matching degree with all candidate tail entities. Let the entity set be U, and its embedding matrix be denoted as E. By calculating the dot product of the fused representation and the entity embedding, the score vector s of the i-th sample for all entities is obtained accordingly. i The expression is as follows:

[0127]

[0128] Among them, s ik =x i ·e k Let x represent the i-th sample. i With the k-th entity e k The matching score, |U|, represents the total number of entities. This step can be viewed as a large-scale vector similarity retrieval in the tail entity space of the knowledge graph, with the goal of matching the correct tail entity t. i Achieve the highest score.

[0129] This embodiment further verified the simulation results. First, based on simulated radar signal parameters, a radar application reasoning knowledge graph was constructed. Specifically, two datasets were constructed through simulation experiments, simulating short-sequence and long-sequence radar signals, represented by dataset A (simulated short sequence) and dataset B (simulated long sequence), respectively. This embodiment simulated ten typical radar applications, generating 1000 signal samples for each application. Dataset A contains four types of relationships: radar carrier frequency value, pulse width value, repeatability value, and radar application, with a total of 10096 entities and 40000 triples. Dataset B contains 13 types of relationships: radar carrier frequency modulation type, maximum carrier frequency value, minimum carrier frequency value, median carrier frequency value, pulse width modulation type, maximum pulse width value, minimum pulse width value, median pulse width value, repeatability modulation type, maximum repeatability value, minimum repeatability value, median repeatability value, and radar application, with a total of 10210 entities and 130000 triples. The Neo4j graph database was used as the underlying support platform for structured storage and visualization. Knowledge graph of partial radar applications based on dataset B, such as Figure 4 As shown.

[0130] To verify the feasibility and superiority of the method of the present invention, this embodiment conducts a comparative experiment on the trained radar application inference network, specifically as follows:

[0131] (1) Comparative Experiment 1;

[0132] Experiments were conducted on datasets A and B, which were randomly divided into training, validation, and test sets, representing 60%, 10%, and 30% of the datasets, respectively. This embodiment compares the method of this invention with some commonly used classification networks, including traditional graph embedding models: TransE and TransH, and common classification algorithms: SVM and ResNet. Mean Reciprocal Rank (MRR) and Hits@n were used as evaluation metrics. MRR measures the average rank of the correct answer in the prediction results, while Hits@n represents the proportion of correct entities whose rank in the prediction results does not exceed n. Both reflect the model's accuracy and retrieval ability in knowledge reasoning tasks, and their calculation expressions are as follows:

[0133]

[0134] Where m represents the test set samples, and l(·) represents the indicator function, which has a value of 1 when the condition is met and 0 otherwise. rank i This represents the rank of the correct answer in the i-th test sample.

[0135] The results of the comparative experiment 1 in this embodiment are shown in Table 2, which shows the radar application reasoning performance index results under different models.

[0136] Table 2

[0137]

[0138] As shown in Table 2, compared to dataset A, all models in dataset B show improved performance across all metrics. This indicates that the inference accuracy of the model significantly improves with the increase in the number of relations and the size of triples in the radar application knowledge graph. The embodiment of this invention achieves the best performance on all three mainstream evaluation metrics—MRR, Hits@1, and Hits@3—in both datasets. On dataset A, the Hits@1 metric of the embodiment of this invention is improved by approximately 30% to 40% compared to traditional embedding models, while the improvement is approximately 20% on dataset B. Furthermore, ResNet's performance on dataset A is close to that of the embodiment of this invention, but its inference performance is slightly insufficient in dataset B due to the increased complexity of entity relations. This indicates that although ResNet has strong capabilities in feature extraction, it still has certain limitations in handling radar application inference tasks with rich structural information. The performance improvement can be attributed to the effective fusion of multi-head attention mechanisms and local convolutional features, which significantly enhances the model's ability to model global semantic dependencies and local structural patterns. Figure 5 and Figure 6 The radar usage inference results of the method of the present invention in datasets A and B are shown respectively.

[0139] (2) Comparative Experiment 2;

[0140] In Comparative Experiment 2, the comparison method is that the method of this invention does not employ the multi-head attention mechanism, but only uses simple embedding vector concatenation as the input to the subsequent convolutional neural network. The results of the comparative experiments without the multi-head attention mechanism are shown in Table 3.

[0141] Table 3

[0142]

[0143] As shown in Table 3, after removing the multi-head attention mechanism module, the comparative methods showed varying degrees of decline in all three evaluation metrics on both datasets, especially in Hits@1, where the declines were 16% and 7%, respectively, exhibiting the most significant performance degradation. In contrast, the changes in the Hits@3 metrics were relatively smaller. Despite the overall performance decline, the methods still significantly outperformed the traditional embedding methods TransE and TransH on MRR and Hits@1. These experimental results demonstrate that introducing an attention mechanism can effectively enhance the representation fusion capability between entities and relations, thereby exhibiting superior performance in terms of prediction accuracy and generalization ability.

[0144] (3) Comparative Experiment 3;

[0145] This comparative experiment will set different signal-to-noise ratio (SNR) conditions to simulate the application of the method of this invention in real-world scenarios. In knowledge graph embedding tasks, the signal-to-noise ratio (SNR) is defined as the ratio between the number of true triples and the number of noisy triples, expressed as follows:

[0146]

[0147] Figure 7 Experimental results under different signal-to-noise ratio (SNR) conditions are presented. The results show that the method of this invention outperforms traditional graph embedding methods and classification networks at all SNR levels, and maintains high inference accuracy even under a low SNR of -5dB.

[0148] (4) Comparative Experiment 4;

[0149] This comparative experiment will verify the effectiveness of the parameters of the method proposed in this invention through parameter experiments, specifically testing the impact of two parameters—embedding dimension and the number of heads in the multi-head attention mechanism—on the model's performance. During knowledge reasoning, the dimension of entity and relation vector embeddings has a certain impact on model performance. Low-dimensional vectors may limit the model's expressive power, making it difficult to fully capture the semantic relationships between entities, while high-dimensional vectors, although better able to represent complex structures and semantic information, can also lead to problems such as increased parameter count and overfitting. Figure 8 It can be seen that, under the condition that other parameters remain unchanged, the Hit@1 value of the method of the present invention steadily increases with the increase of the vector embedding dimension in both datasets, reaching the maximum at 256 and achieving suboptimal performance at 512.

[0150] (5) Comparative Experiment 5;

[0151] The number of attention heads in a multi-head attention mechanism also affects the model's inference performance. Appropriately increasing the number of attention heads helps provide more combinations of attention weights, allowing the model to capture richer feature information; however, an excessive number of heads increases the risk of overfitting. Since the embedding dimension must be divisible by the number of attention heads, this comparative experiment fixed the embedding dimension at 240. When the number of attention heads is 8, the model performs optimally on both datasets, as shown in the results. Figure 9 As shown.

[0152] In summary, the method of this invention utilizes the parameter information of the received radar signals to construct a radar purpose reasoning knowledge graph, extracts triplet knowledge, and, based on the head entity and relation embedding sequence of the radar knowledge graph, effectively captures high-order semantic dependencies between entities and relations through a multi-head attention mechanism and a Transformer encoder, generating deep interactive representations. The encoded results are then reshaped into two-dimensional vectors and input into a convolutional neural network to extract local semantic patterns, thereby enhancing sensitivity to complex entity combinations. Furthermore, cross-entropy supervision signals are used to match the predicted vector with candidate tail entity embeddings, achieving accurate reasoning of tail entities. The method of this invention has the advantages of accuracy, efficiency, and applicability to radar signals received by the reconnaissance side with numerous and complex combinations of parameters.

[0153] Those skilled in the art will recognize that the embodiments described herein are intended to help the reader understand the principles of the invention, and should be understood that the scope of protection of the invention is not limited to such specific statements and embodiments. Various modifications and variations can be made to the invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the invention should be included within the scope of the claims of the invention.

Claims

1. A knowledge graph-based radar purpose reasoning method, the specific steps being as follows: S1, intercepting original radar signal parameter data, and constructing a radar purpose knowledge graph; The radar signal parameters include: carrier frequency CF, pulse width PW, and pulse repetition interval PRI; The construction of the radar purpose knowledge graph includes entity and relationship definition and visual display; First, based on the signal parameter representation ability of the radar purpose, select the carrier frequency, pulse width, and pulse repetition interval as the modeling attributes, define the radar signal sample as the head entity in the knowledge graph, and the specific purpose and the corresponding parameter value as the tail entity; and extract the statistical features of the carrier frequency, pulse width, and pulse repetition interval for modeling; The statistical features include: median, maximum, and minimum; the knowledge graph G is formally expressed as follows: G=(V, R, T) (1) Wherein, V represents the entity set, R represents the relationship set, and T represents the triple set; S2, extract the structured triple knowledge in the knowledge graph of step S1; The structured triple knowledge in the knowledge graph of step S1 includes the head entity, the relationship, and the tail entity; the extracted knowledge graph triple is in the form of (h, r, t); Wherein, the head entity h represents the radar signal sample, the relationship r represents the parameter type or purpose relationship, and the tail entity t represents the corresponding parameter value or purpose category; S3, map the triple knowledge obtained in step S2 into a fixed-dimensional vector representation; First, the entity and relationship are uniquely numbered based on the dictionary mapping method to obtain the numbered set of entities and relationships; then the head entity, relationship, and tail entity in the triple are converted into corresponding index sequences; then the obtained index is mapped into a low-dimensional dense vector representation using an embedding layer to realize the continuous representation of the triple knowledge; finally, the original discrete symbolized triple is uniformly mapped into a fixed-dimensional vector space; S4, construct a radar purpose reasoning network; The radar purpose reasoning network includes: 1 multi-head attention mechanism module, 1 convolution module, and 1 scoring module; The multi-head attention mechanism module includes: 1 input embedding layer, 1 stacked Transformer encoder layer, and 1 output layer; Each Transformer encoder layer includes: 1 multi-head self-attention layer, 1 feed-forward fully connected layer, 1 residual connection, and 1 layer normalization. The convolution module includes: 1 input reshaping layer, 1 two-dimensional batch normalization layer, 1 two-dimensional convolution layer, 1 convolution output normalization layer, 1 ReLU activation function, 1 Dropout layer, 1 flattening layer, 2 fully connected layers, and 1 layer normalization layer; S5, use the vector obtained in step S3 as the input of the radar purpose reasoning network constructed in step S4 to train the radar purpose reasoning network. The head entity obtained in step S3 and the relation vector are input into the radar use reasoning network for forward propagation to obtain a predicted score vector and calculate a loss function value; a back propagation algorithm based on a stochastic gradient descent is used to update the model parameters; the forward propagation and the back propagation are repeatedly performed until the loss function gradually decreases and tends to be stable, and the training converges, so that the trained radar use reasoning network is obtained; where the score vector s i Applying the Softmax function, the prediction distribution p ik is obtained for the tail entity, with the expression as follows: where s ik = x i · e k represents the matching score of the ith sample x i and the kth entity e k , and |U| represents the total number of entities. The radar use reasoning network adopts a sample average loss function L, and the expression is as follows: where N represents the number of samples, t i represents the number of correct labels of the i-th sample, L i represents the cross-entropy loss of the i-th sample; S6, based on the trained radar knowledge graph use reasoning network obtained in step S5, the predicted score vector obtained in step S5 is matched with a candidate tail entity embedding to realize reasoning of the radar use. 2.The knowledge graph-based radar use reasoning method according to claim 1, characterized in that, The step S4 is specifically as follows: S41, a multi-head attention mechanism module; Given input sequence The input is first mapped by linear mapping into query Q, key K and value V, respectively, and the linear mapping relationship vector expression is as follows: Q i = XW i Q K i = XW i K V i = XW i V i = 1,..., h (4) wherein, represents the concatenation of the two embedding vectors of the head entity and the relation, d represents the embedding dimension, Q i , K i , V i represent the Q, V, K linear mapping relation vectors of the i-th attention head, respectively, represents the parameter matrix of the i-th attention head, h represents the number of attention heads, and d = h x d k The calculation expression of each attention head is as follows: wherein d k denotes the embedding dimension of each attention head, head i denotes the i-th attention head, denotes the transpose operation; The outputs of the plurality of attention heads are linearly mapped after being concatenated to obtain the overall output of the multi-head attention mechanism The expression is as follows: MultiHead(X) = Concat(head1,...,head h )·W O (6) where W O represents an output mapping matrix; The multi-head attention mechanism module adopts a stacked Transformer encoder as a structural basis, and there are l layers of Transformer encoders; each layer of the Transformer encoder includes a multi-head attention layer and a feed-forward full connection layer, and information flow is stably propagated through residual connection and a normalization layer; the calculation process expression of the lth layer of the Transformer encoder is as follows: Z (l) = LayerNorm(X (l-1) + MultiHead(X (l-1) )) (7) X (l) = LayerNorm(Z (l) + FFN(Z (l) )) (8) wherein Z (l) represents the output of the l-th multi-head attention layer, X (l) represents the output of the l-th Transformer encoder, FFN(·) represents a feed-forward network consisting of two fully connected layers, and the expression is as follows: FFN(Z)=ReLU(ZW1+b1)W2+b2 (9) wherein Z represents the output of the multi-head attention layer, W1 and W2 represent the weight matrices of the linear transformation, b1 and b2 represent the bias terms of the linear transformation, and the outputs X of each layer of the Transformer encoder are finally stacked (l) as a fused representation of the head entity and the relation; S42, a convolution module; At step S41, the fusion representation X is obtained (l) After that, a spatial perception module is further introduced to enhance the local feature expression capability. First, the fusion representation is flattened into a vector and then reshaped into a two-dimensional feature map P, and the expression is as follows: P = reshape(vec(X (l) )) (10) Then, the input is input into a two-dimensional convolutional neural network, and a local convolution kernel is used to extract a potential spatial structure mode between the entity and the relation, and the expression is as follows: F=Dropout(ReLU(BN(Conv2D(P)))) (11) After the obtained convolution feature map F is flattened into a vector f, the vector f is sequentially mapped through two layers of full connection to compress the representation dimension and improve the expression ability, and the process expression is as follows: x i = LayerNorm(ReLU(fW3 + b3)W4 + b4) (12) Finally, the obtained The final fusion representation vector is obtained, which represents the deep semantic and local interaction information of the fused head entity and relation. The vector is used for matching and scoring with all candidate tail entities. S43, a scoring module; Based on step S42, after the semantic fusion of the head entity and the relationship embedding is completed, a convolution and a fully connected layer module are further introduced to mine spatial features, and the obtained fusion representation vector is input into a scoring module to evaluate the matching degree between it and all candidate tail entities; let the entity set be U, and the embedding matrix of U be E. By calculating the dot product of the fusion representation and the entity embedding, the score vector s of the i-th sample with respect to all entities is obtained accordingly i , and the expression is as follows: where s ik = x i · e k represents the matching score of the ith sample x i and the kth entity e k , and |U| represents the total number of entities.

Images