Radiation source intention identification method and system based on knowledge and data hybrid driving

Abstract

The invention discloses a radiation source intention identification method and system based on knowledge and data hybrid driving. The method comprises the following steps: at a data driving end, performing feature extraction on data based on a space-time self-attention model to obtain a data feature vector f; at a knowledge driving end, constructing a radiation source intention recognition knowledge graph, and converting the knowledge graph into a knowledge embedding vector ke by adopting a knowledge representation module; integrating features of a knowledge end and a data end by adopting a knowledge alignment module, and converting a knowledge embedding vector ke into a knowledge feature vector fk specific to an intention sample; splicing the data feature vector f and the knowledge feature vector fk to obtain an enhanced feature vector fe; and sending the enhanced feature vector fe into a classifier to complete intention recognition. The radiation source intention recognition performance of the model under small samples can be effectively improved.

Description

Technical Field

[0001] The present invention relates to battlefield intention recognition technology, and particularly to a method and system for radiation source intention recognition based on a hybrid drive of knowledge and data. Background Art

[0002] Battlefield intention recognition (CIR) is an important area in military intelligence analysis and battlefield situation awareness, which involves predicting and analyzing the action intentions of enemy or friendly units. This technology is crucial for making strategic decisions, optimizing resource allocation, and improving battlefield response speed.

[0003] In new battlefield environments such as transient, highly adversarial, and dense targets, conventional intention recognition methods based on rules, template matching, etc. face severe challenges. More and more scholars are exploring deep learning-based methods to solve the CIR task. For example, the Convolutional Neural Network-Recurrent Neural Network (CNN-RNN) framework, using a parallel structure or a sequential structure, automatically extracts the temporal features of data from the time dimension. The intention recognition method based on the Spatio-Temporal Self-Attention Model (STSA) can achieve the joint extraction of features in the time dimension and the space dimension, further improving the intention recognition performance. However, in military application scenarios, the datasets for the above data-driven models are extremely scarce, and it is difficult to train the models. At the same time, due to the lack of interpretability of the models, it affects the decision-making confidence of commanders. Knowledge graphs can store expert knowledge in the form of triples and can efficiently represent the knowledge through embedding for in-depth applications. Combining the respective advantages of knowledge and data-driven, that is, a hybrid drive of knowledge and data, has become a feasible approach for radiation source intention recognition methods under small samples. Summary of the Invention

[0004] Object of the Invention: Aiming at the challenges faced by existing deep learning methods, the present invention proposes a method and system for radiation source intention recognition based on a hybrid drive of knowledge and data, which can effectively improve the recognition performance under small samples.

[0005] Technical Solution: To achieve the above object, the present invention adopts the following technical solutions:

[0006] In a first aspect, a method for radiation source intention recognition based on a hybrid drive of knowledge and data includes the following steps:

[0007] Feature extraction is performed on the data based on a spatio-temporal self-attention model to obtain a data feature vector f. The spatio-temporal self-attention model includes a spatial attention module, a temporal attention module, and a classification layer. The basic units of the spatial attention module and the temporal attention module are the multi-head self-attention mechanism;

[0008] Construct a domain knowledge graph for radiation source intention recognition, vectorize the semantic information of each node in the knowledge graph using a text embedding model, and assign weights to each edge in the knowledge graph. Use the vectorized semantic nodes as the initial input feature vector H 0 , and output a knowledge embedding vector ke through a multi-layer graph convolutional network;

[0009] Use the data feature vector f as the query vector, and use the knowledge embedding vector ke as the key vector and value vector to perform a knowledge graph-based self-attention mechanism to obtain an intention-specific knowledge feature vector f k , and concatenate the data feature vector f and the knowledge feature vector f k to obtain an enhanced feature vector f e ;

[0010] Use the enhanced feature vector f e to obtain the network output through a classifier composed of a fully connected layer and SoftMax. The output is the probability of each target intention, and select the largest one as the result of radiation source intention recognition.

[0011] Furthermore, construct a domain knowledge graph for radiation source intention recognition, including:

[0012] Model the ontology layer, divide the nodes into awareness layer nodes, behavior layer nodes, and sub-behavior layer nodes. The awareness layer corresponds to the output space of the intention recognition model. The behavior layer nodes include enemy behaviors and our behaviors; the sub-behavior layer nodes represent enemy behaviors directly observed through battlefield observation means;

[0013] Use neo4j as a knowledge graph storage tool to extract the knowledge of combat domain experts and complete the construction of the knowledge graph.

[0014] Furthermore, the basic units of the spatial attention module and the temporal attention module are both the multi-head self-attention mechanism. Each self-attention head is expressed as: where Q, K, and V represent the query vector, key vector, and value vector respectively, and d k represents the dimension of the query vector and the key vector;

[0015] The multi-head self-attention mechanism linearly maps the original Q, K, and V vectors h times to obtain the inputs for each attention head. Each attention head calculates the output results in parallel, and all the results are concatenated together and passed through a linear layer to obtain the final output result. The specific expression is:

[0016] MultiHead(Q,K,V)=Concat(head1,head2,…head h )W O

[0017]

[0018] Among them, is the input mapping matrix; represents the output linear layer weight of the multi-head attention mechanism; Concat(·) represents the concatenation operation;

[0019] Based on the spatio-temporal self-attention model, feature extraction is performed on the data to obtain the data feature vector f, including: inputting the transposed time series data X into the spatial attention module to obtain the enhanced time series feature data f space from the spatial dimension. Then, the transposed time series feature data f space T is input into the time attention module to extract time features. Finally, the data feature vector f is obtained through the classification layer.

[0020] Furthermore, the t-th layer of the multi-layer graph convolutional network GCN is represented as a non-linear function: H t =f(H t-1 , A), where represents the nodes of layer t; represents the nodes of layer t-1; f(·,·) represents the graph convolution operation, which is expressed as: Among them represents the normalized adjacency matrix; represents the learned weight; σ(·) represents the non-linear activation function;

[0021] Through the initial input feature vector H 0 , after the graph convolution operation, the knowledge embedding vector is finally obtained, where D represents the output dimension of the last GCN layer, and n represents the number of nodes in the knowledge graph.

[0022] Furthermore, the data feature vector f is used as the query vector, and the knowledge embedding vector ke is used as the key vector and value vector to perform the self-attention mechanism based on the knowledge graph, obtaining the intention-specific knowledge feature vector f k , including:

[0023] The data feature vector f passes through a fully connected layer to obtain the query vector Q, and the knowledge embedding vector ke passes through a fully connected layer to obtain the key vector K and the value vector V, where The query vector Q and the key vector K are subjected to dot product attention to obtain the attention degree α of the data feature vector f for each node on the graph, where A scaling factor is introduced to alleviate the problems of vanishing and exploding gradients; the value V is weighted and summed using the attention scores α of each node to obtain sample-specific knowledge features The specific expression is:

[0024]

[0025] The outputs from multiple attention heads are concatenated and passed through a linear layer to obtain sample-specific knowledge features:

[0026]

[0027] where Concat(·) represents the concatenation operation, represents the final linear learning weight layer, represents the output of the h-th attention head.

[0028] Furthermore, the classifier output is expressed as:

[0029] S p = softmax(FC(f e ))

[0030] Select the maximum value argmax(s p ) as the result of the radiation source intention recognition.

[0031] Furthermore, the spatio-temporal self-attention model adopts the following loss function:

[0032]

[0033] where, N represents the total number of training samples in the training set; M is the total number of intention label categories in the training set; p(x n，m ) is the probability that the n-th sample in the training set corresponds to the true label category m, and q(x n，m ) is the probability that the n-th sample of the network model output value corresponds to the m-th category; β = (N - 1) / N, represents the number of samples y n , represents the negation of the valid number of samples of category y n ; K1 represents the cost factor for misjudging a high-threat intention as a low-threat intention, and K2 represents the cost factor for misjudging a low-threat intention as a high-threat intention, and K1 > K2.

[0034] In a second aspect, a radiation source intention recognition system driven by a hybrid of knowledge and data includes the following steps:

[0035] A data-driven module for extracting features from data based on a spatio-temporal self-attention model to obtain a data feature vector f. The spatio-temporal self-attention model includes a spatial attention module, a temporal attention module, and a classification layer, where the basic unit of the spatial attention module and the temporal attention module is the multi-head self-attention mechanism;

[0036] A knowledge representation module for constructing a domain knowledge graph for radiation source intention recognition, vectorizing the semantic information of each node in the knowledge graph using a text embedding model, and assigning weights to each edge in the knowledge graph, and using the vectorized semantic nodes as the initial input feature vector H 0 , and outputting a knowledge embedding vector ke through a multi-layer graph convolutional network;

[0037] A knowledge alignment module for using the data feature vector f as a query vector, and using the knowledge embedding vector ke as a key vector and a value vector to perform a knowledge graph-based self-attention mechanism to obtain an intention-specific knowledge feature vector f k , and concatenating the data feature vector f and the knowledge feature vector f k to obtain an enhanced feature vector f e ;

[0038] A classification and recognition module for obtaining a network output through a classifier composed of a fully connected layer and SoftMax for the enhanced feature vector f e , where the output is the probability of each target intention, and the maximum one is selected as the result of radiation source intention recognition.

[0039] The present invention also provides a computer device, including: one or more processors; a memory; and one or more programs, where the one or more programs are stored in the memory and are configured to be executed by the one or more processors, and when the program is executed by the processor, the steps of the above-mentioned radiation source intention recognition method driven by a hybrid of knowledge and data are implemented.

[0040] The present invention also provides a computer storage medium, on which a computer program is stored, and when the computer program is executed by a processor, the steps of the above-mentioned radiation source intention recognition method driven by a hybrid of knowledge and data are implemented.

[0041] Beneficial effects: Aiming at the dilemmas of existing deep learning methods, the present invention proposes a method and system for identifying radiation source intentions driven by a hybrid of data and knowledge. By using a spatio-temporal self-attention model to extract data features, a knowledge representation module to obtain knowledge features, and a knowledge alignment module to integrate the feature vectors at the data and knowledge ends, the performance limitations caused by data scarcity can be alleviated, and the interpretability of the model can be greatly improved. The simulation results show that the method of the present invention is the best-performing model on both the original data set and the noisy data. The present invention provides a new solution for identifying radiation source intentions with small samples. Description of the Drawings

[0042] Figure 1 It is a flowchart of the method for identifying radiation source intentions driven by a hybrid of data and knowledge in an embodiment of the present invention;

[0043] Figure 2 It is an ontology model of the battlefield intention recognition knowledge graph in an embodiment of the present invention;

[0044] Figure 3 It is an ROC curve graph of five models of the simulation results in an embodiment of the present invention;

[0045] Figure 4 It is the F-1 score of the STSA model and the KEN model for tail intentions under different noises in an embodiment of the present invention;

[0046] Figure 5 It is a heat map drawn from the attention scores of the knowledge alignment module in an embodiment of the present invention. Detailed Embodiments

[0047] The technical solution of the present invention will be further described below with reference to the drawings.

[0048] The present invention proposes a method for identifying radiation source intentions driven by a hybrid of data and knowledge, which is applicable to identifying radiation source intentions with few samples. Identifying radiation source intentions is an important part of battlefield intention recognition. In the implementation of the present invention, intention recognition is mainly carried out for the sea combat scenario, but it should be understood that this is only for illustrative purposes and not a limitation to the present invention. The method of the present invention is generally applicable to the recognition of other combat scenarios, such as air combat scenarios and land combat scenarios. The overall solution is as Figure 1 shown, and specifically includes the following steps:

[0049] Step 1: Obtaining the feature vector at the data end. Based on the STSA model, the data is subjected to feature extraction to obtain the data feature vector f.

[0050] In the present invention, a spatio-temporal self-attention (STSA) model is adopted as the basic feature extractor. The core of the STSA model is to capture spatio-temporal features in the identification of radiation source intentions through a spatio-temporal self-attention mechanism. Among them, the self-attention module is the key part of the STSA model.

[0051] The STSA model consists of a space-attention module, a time-attention module, and a fully connected layer (with softmax as the activation function, serving as the classification layer).

[0052] The basic units of both the space-attention module and the time-attention module are the multi-head self-attention mechanism. Each self-attention head can be expressed as:

[0053]

[0054] where Q, K, and V represent the query vector, key vector, and value vector respectively, and d k represents the dimensions of the query vector and the key vector.

[0055] The multi-head self-attention mechanism linearly maps the original Q, K, and V vectors h times to obtain the inputs of each attention head. Each attention head calculates the output results in parallel, and all the results are concatenated together and passed through a linear layer to obtain the final output result. The specific expression is:

[0056] MultiHead(Q,K,V)=Concat(head1,head2,…head h )W O (2)

[0057]

[0058] where, is the input mapping matrix; represents the weight of the output linear layer of the multi-head attention mechanism; d input is the dimension of the input vector of each attention head, and d output is the dimension after concatenating the output vectors of multiple attention heads.

[0059] In the embodiments of the present invention, the specific hyperparameters of the STSA model are shown in Table 1:

[0060] Table 1 Hyperparameters of the SATA model

[0061]

[0062] where, ns is the number of layers of multi - head attention in the spatial attention module, n t is the number of layers of multi - head attention in the temporal attention module, h is the number of attention heads of the multi - head self - attention mechanism in the spatio - temporal attention module, d k is the dimension of the key vector for each attention head, d v is the dimension of the value vector for each attention head, β is the penalty degree of the re - weighted term in the loss function for the head - class samples, dropout rate is the proportion of neurons randomly inactivated during training, batch size is the number of training samples used in each gradient - descent update, epoch is the number of times the model completes a full traversal of the entire training dataset, optimizer is the optimization tool adopted. In the present invention, the Adam optimizer is used as the optimization algorithm, and learning rate is the step size for the model to update the weights according to the direction of gradient descent in each iteration.

[0063] The original data comes from a combat simulation system. The original time - series data is partitioned using a sliding window, and then the partitioned data is submitted to domain experts for annotation. The processed dataset consists of 12,000 samples. Each sample is a two - dimensional time series containing ten simulated time steps, and each time step includes 15 - dimensional feature parameters. Each sample can be represented as The transposed time - series data X is input into the spatial attention module to obtain enhanced temporal - sequence feature data f from the spatial dimension space , then, the transposed temporal - sequence feature data f space T is input into the temporal attention module to extract richer temporal features. Finally, the data feature vector f is obtained through a classification layer (a fully - connected layer with the activation function softmax).

[0064] Step 2: Obtaining the knowledge - end feature vector. Construct a domain knowledge graph G = <N, ε> for radiation source intention recognition, and use the knowledge representation module to convert the knowledge graph into a knowledge embedding vector ke.

[0065] First, model the ontology layer. Refer to Figure 2 , and divide the nodes into awareness - layer nodes, behavior - layer nodes, and sub - behavior - layer nodes. Among them: the awareness layer corresponds to the output space of the intention recognition model; the behavior - layer nodes include enemy behaviors and our - side behaviors; the sub - behavior - layer nodes represent enemy behaviors that can be directly observed through battlefield observation means. Then, use neo4j as the knowledge - graph storage tool to extract the knowledge of sea - surface combat domain experts, thereby completing the construction of the knowledge graph.

[0066] The knowledge representation module first vectorizes the rich semantic information of each node using a text embedding model, and then assigns empirical weights to each edge (relationship) in the knowledge graph. The vectorized semantic nodes are used as the initial input feature vector H 0 , and the knowledge embedding vector ke is output through a multi-layer graph convolutional network (GCN). The t-th layer of the GCN is represented as a non-linear function:

[0067] H t = f(H t-1 , A) (4)

[0068] where represents the nodes of layer t; represents the nodes of layer t-1; f(·,·) represents the graph convolution operation, represents the adjacency matrix, which represents the association information between n nodes in the graph.

[0069] In the embodiment of the present invention, the knowledge representation module consists of two layers of GCNs, with output dimensions of 512 and 150. First, the knowledge graph is preprocessed. For text embedding, the open-source BGE1 model provided by BAAI is used to convert the description of each node in the knowledge graph into a 1024-dimensional embedding representation, which is used as the initial input feature H0 of the first layer of GCN. In the edge reweighting step, weights in the range (0, 1) are assigned to each relationship in the knowledge graph. Then, through the graph convolution operation, it can be expressed as:

[0070]

[0071] where represents the normalized adjacency matrix; represents the learned weight; σ(·) represents the non-linear activation function, set to ReLU.

[0072] Finally, the knowledge embedding vector is obtained, where D represents the input dimension of the last GCN layer, and n represents the number of nodes in the knowledge graph.

[0073] Step 3: Integrate the data and knowledge-side feature vectors. The knowledge alignment module is used to convert the knowledge embedding vector ke into a knowledge feature vector f specific to the intent sample k . And the data feature vector f and the knowledge feature vector f k are concatenated to obtain the enhanced feature vector f e .

[0074] The core of the knowledge alignment module is the multi-head self-attention mechanism, which uses the basic feature f at the data end as the query vector, and the knowledge representation ke as the key vector and value vector to perform the attention mechanism based on the knowledge graph.

[0075] Specifically, the data feature vector f passes through a fully connected layer to obtain the query vector Q, and the knowledge embedding vector ke passes through a fully connected layer to obtain the key vector K and value vector V. These two fully connected layers are separate linear layers without the softmax activation function, where n represents the number of nodes in the knowledge graph. The query vector Q and the key vector K go through dot-product attention, that is, perform matrix multiplication on the query vector Q and the key vector K to obtain the attention degree α of the data feature vector f for each node on the graph, where Furthermore, a scaling factor is introduced to normalize the result of the matrix multiplication, preventing the input values of the softmax function from being too large, thus alleviating the problems of gradient vanishing and explosion to a certain extent; finally, the value vector V is weighted and summed using the attention scores α of each node to obtain the knowledge features specific to the sample The specific expression is:

[0076]

[0077] α = [α1, α2,..., α n (7)

[0078]

[0079] The specific hyperparameters are shown in Table 2:

[0080] Table 2 Hyperparameters of the self-attention mechanism of the knowledge graph

[0081]

[0082] Among them, h is the number of attention heads of the multi-head self-attention mechanism, d k is the dimension of the key vector of each attention head, d v is the dimension of the value vector of each attention head, d moudle is the dimension of the hidden layer of the model. The number of neurons in the fully connected layer at the data end is 128, the number of neurons in the fully connected layer at the knowledge end is 64, the batch size is the number of training samples used in each gradient descent update, the epoch is the number of times the model completes a full traversal of the entire training dataset, the optimizer is the optimization tool used. In the present invention, the Adam optimizer is used as the optimization algorithm, and the learning rate is the step size for the model to update the weights according to the direction of gradient descent in each iteration.

[0083] Finally, the outputs of the multiple heads are concatenated and passed through a linear layer to obtain the final sample-specific knowledge features. Specific expression:

[0084]

[0085] Among them, Concat(·) represents the concatenation operation. represents the final linear learning weight layer. represents the output of the h-th attention head. In the final fully connected layer, the number of neurons on the data side is set to 128, and the number of neurons on the knowledge side is set to 64. Finally, the knowledge feature f k and the basic feature f are concatenated to obtain the enhanced knowledge vector f e = [f; f k .

[0086] Step 4: The classifier performs intent recognition. The enhanced feature vector f e is sent into the classifier. The enhanced feature vector f e passes through the classifier composed of a fully connected layer and SoftMax to obtain the network output, and the output is the probability of each target intent. Specifically:

[0087] S p = softmax(FC(f e ))

[0088] Select the maximum value argmax(s p ) as the result of the radiation source intent recognition.

[0089] Based on Steps 1 to 4, the knowledge enhanced network model (knowledge enhanced network, KEN) constructed by the present invention is trained. The original data of the dataset comes from a typical combat simulation system. The original data is partitioned using a sliding window, and then the partitioned data is submitted to domain experts for annotation. The processed dataset consists of 12,000 samples, where each sample is two-dimensional time data containing ten simulated time steps. Each time step includes 15-dimensional feature parameters. The dataset is divided into a training set and a test set in a ratio of 8:2. The training set consists of 9,600 samples, with the search and detection intent accounting for 47.083%, the target locking intent accounting for 41.875%, the interception intent and the blocking and suppression intent accounting for 5.417%, and the out-of-band deception intent accounting for 0.208%. The test set consists of 2,400 samples, with 480 samples in each intent category. In addition, noise is introduced into the original test set at different signal-to-noise ratios (SNRs) to simulate the complex electromagnetic spectrum environment in a real battlefield scenario, which is called the noise dataset.

[0090] For the original dataset, accuracy, precision, recall, and F1-score were used as evaluation metrics, and the main results are shown in Table 3. As can be seen from Table 3, the overall performance of the KEN model is the best, with an improvement of approximately 1.4% compared to STSA and an average improvement of approximately 5.3% compared to other benchmark models.

[0091] Table 3 Recognition Results of Each Model on the Original Dataset

[0092]

[0093] For the noisy dataset, the average precision was calculated as a measure of robustness. The results obtained are shown in Table 4. As can be seen from Table 4, at different signal-to-noise ratios, the accuracy of the KEN model is the highest, demonstrating the superiority of the KEN model.

[0094] Table 4 Recognition Results of Each Model on the Noisy Dataset

[0095]

[0096] To further evaluate the classification performance of the models, the ROC curve of each model was plotted on the original dataset, and the AUC value was calculated. As Figure 3 shown, the KEN model obtained the highest AUC value of 0.9981, indicating the superior classification performance of the KEN model.

[0097] To further verify the effective recognition performance of the models under small samples, the F1-scores of the tail class intentions were compared. Specifically, the performance of the KEN model and the STSA model for outboard deception intentions (with only 20 training samples) under different noises was compared. As Figure 4 shown, the KEN model achieved a performance improvement of approximately 4% compared to STSA under various noise conditions, and as the noise increased, the performance improvement of KEN for the tail class intentions became more obvious, indicating that the KEN model has more effective recognition performance under small samples.

[0098] To further verify the interpretability of the models, the attention scores of the knowledge alignment module were plotted as a heatmap, which can visually observe the attention degree of the models to different knowledge. As Figure 5 shown, the visualization of the KEN for a certain outboard deception intention shows different attention degrees to different knowledge.

[0099] Based on the same technical concept as the method embodiment, the present invention also provides a radiation source intention recognition system based on a hybrid drive of knowledge and data, including:

[0100] A data-driven module, which is used to extract features from data based on a spatio-temporal self-attention model to obtain a data feature vector f. The spatio-temporal self-attention model includes a spatial attention module, a temporal attention module, and a classification layer. The basic units of the spatial attention module and the temporal attention module are the multi-head self-attention mechanism;

[0101] A knowledge representation module, which is used to construct a domain knowledge graph for radiation source intention recognition, vectorize the semantic information of each node in the knowledge graph using a text embedding model, assign weights to each edge in the knowledge graph, and use the vectorized semantic nodes as the initial input feature vector H 0 , and output a knowledge embedding vector ke through a multi-layer graph convolutional network;

[0102] A knowledge alignment module, which is used to use the data feature vector f as a query vector, and use the knowledge embedding vector ke as a key vector and a value vector to perform a knowledge graph-based self-attention mechanism to obtain an intention-specific knowledge feature vector f k , and concatenate the data feature vector f and the knowledge feature vector f k to obtain an enhanced feature vector f e ;

[0103] A classification and recognition module, which is used to obtain a network output by passing the enhanced feature vector f e through a classifier composed of a fully connected layer and a softmax function. The output is the probability of each target intention, and the largest one is selected as the result of radiation source intention recognition.

[0104] The present invention also provides a computer device, including: one or more processors; a memory; and one or more programs, where the one or more programs are stored in the memory and are configured to be executed by the one or more processors. When the program is executed by the processor, the steps of the above-mentioned radiation source intention recognition method based on knowledge and data hybrid driving are implemented.

[0105] The present invention also provides a computer storage medium, on which a computer program is stored. When the computer program is executed by a processor, the steps of the above-mentioned radiation source intention recognition method based on knowledge and data hybrid driving are implemented.

[0106] Those skilled in the art should understand that the embodiments of the present invention can be provided as a method, a device, a computer device, or a computer program product. Therefore, the present invention can adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present invention can adopt the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0107] The present invention is described with reference to the flowchart of a method according to an embodiment of the present invention. It should be understood that each process in the flowchart and the combination of processes in the flowchart can be implemented by computer program instructions. These computer program instructions can be provided to the processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing devices to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing devices generate a device for implementing the functions specified in one process Figure 1 or multiple processes. These computer program instructions can also be stored in a computer-readable memory capable of guiding the computer or other programmable data processing devices to work in a specific manner, so that the instructions stored in the computer-readable memory generate a manufactured article including an instruction device, and the instruction device implements the functions specified in one process Figure 1 or multiple processes. These computer program instructions can also be loaded onto the computer or other programmable data processing devices, so that a series of operation steps are executed on the computer or other programmable devices to generate a computer-implemented process, and thus the instructions executed on the computer or other programmable devices provide steps for implementing the functions specified in one process Figure 1 or multiple processes.

Claims

1. A method for identifying radiation source intentions driven by a hybrid of knowledge and data, characterized in that Including the following steps: Performing feature extraction on data based on a spatio-temporal self-attention model to obtain a data feature vector f. The spatio-temporal self-attention model includes a spatial attention module, a temporal attention module, and a classification layer. The basic units of the spatial attention module and the temporal attention module are multi-head self-attention mechanisms; Construct a domain knowledge graph for radiation source intention recognition, vectorize the semantic information of each node in the knowledge graph using a text embedding model, assign weights to each edge in the knowledge graph, and use the vectorized semantic nodes as the initial input feature vector H 0 , and output the knowledge embedding vector ke through a multi-layer graph convolutional network; Use the data feature vector f as the query vector, and use the knowledge embedding vector ke as the key vector and value vector to perform a knowledge graph-based self-attention mechanism to obtain an intent-specific knowledge feature vector f k , and concatenate the data feature vector f and the knowledge feature vector f k to obtain an enhanced feature vector f e ; The enhanced feature vector f e The network output is obtained through a classifier composed of a fully connected layer and a softmax function. The output is the probability of each target intention, and the largest one is selected as the result of the radiation source intention recognition.

2. The method according to claim 1, wherein The basic units of both the spatial attention module and the temporal attention module are multi-head self-attention mechanisms. Each self-attention head is expressed as: Among them, Q, K, and V represent the query vector, key vector, and value vector respectively, and d k represents the dimensions of the query vector and the key vector; The multi-head self-attention mechanism linearly maps the original Q, K, V vectors h times to obtain the inputs of each attention head. Each attention head calculates the output results in parallel, and all the results are concatenated and passed through a linear layer to obtain the final output result. The specific expression is: MultiHead(Q,K, V) = Concat(head1, head2, … head h )W O Among them, is the input mapping matrix; represents the output linear layer weights of the multi-head attention mechanism; Concat(·) represents the concatenation operation; Extract features from the data based on the spatio-temporal self-attention model to obtain the data feature vector f, including: input the transposed time series data X into the spatial attention module to obtain the enhanced time series feature data f from the spatial dimension space , then, input the transposed time series feature data f space T into the time attention module to extract time features, and finally, obtain the data feature vector f through the classification layer.

3. The method according to claim 1, wherein Constructing a domain knowledge graph for radiation source intention recognition, including: Modeling the ontology layer, dividing the nodes into awareness layer nodes, behavior layer nodes, and sub-behavior layer nodes. The awareness layer corresponds to the output space of the intention recognition model. The behavior layer nodes include enemy behaviors and our behaviors; the sub-behavior layer nodes represent enemy behaviors directly observed through battlefield observation means; Using neo4j as a knowledge graph storage tool, extracting the knowledge of combat domain experts to complete the construction of the knowledge graph.

4. The method according to claim 2, wherein The t-th layer of the multi-layer graph convolutional network GCN is represented as a non-linear function: H t = f(H t-1 , A), where represents the nodes of layer t; represents the nodes of layer t-1; f(·, ·) represents the graph convolution operation, which is expressed as: where represents the normalized adjacency matrix; represents the learned weights; σ(·) represents the non-linear activation function; Through the initial input feature vector H 0 , after graph convolution operation, finally obtain the knowledge embedding vector where D represents the output degree of the last GCN layer, and n represents the number of nodes in the knowledge graph.

5. The method according to claim 4, characterized in that, Using the data feature vector f as the query vector, and the knowledge embedding vector ke as the key vector and value vector to perform a knowledge graph-based self-attention mechanism to obtain an intent-specific knowledge feature vector f k , including: The data feature vector f passes through a fully connected layer to obtain the query vector Q, and the knowledge embedding vector ke passes through a fully connected layer to obtain the key vector K and the value vector V, where The query vector Q and the key vector K are subjected to dot product attention to obtain the attention degree α of the data feature vector f for each node on the graph, where Introduce a scaling factor to alleviate the problems of gradient vanishing and explosion; use the attention scores α of each node to perform weighted summation on V to obtain the sample-specific knowledge features The specific expression is as follows: Concatenating the outputs from multiple attention heads and passing through a linear layer to obtain sample-specific knowledge features: where Concat(·) represents the concatenation operation, represents the final linear learning weight layer, represents the output of the h-th attention head.

6. The method according to claim 1, wherein The classifier output is expressed as: S p = softmax(FC(f e )) Select the maximum value argmax(s p ) as the result of the radiation source intention recognition.

7. The method according to claim 1, wherein The spatio-temporal self-attention model adopts the following loss function: Among them, N represents the total number of training samples in the training set; M is the total number of intent label categories in the training set; p(x n，m ) is the probability that the nth sample in the training set corresponds to the true label category m, and q(x n，m ) is the probability that the nth sample of the network model output value corresponds to the mth category; β = (N - 1) / N, represents the number of samples y n . represents the negation of the effective number of samples of category y n . K1 represents the cost factor for misjudging a high-threat intent as a low-threat intent, and K2 represents the cost factor for misjudging a low-threat intent as a high-threat intent, and K1 > K2.

8. A radiation source intention recognition system based on a hybrid drive of knowledge and data, characterized in that Including: A data-driven module for performing feature extraction on data based on a spatio-temporal self-attention model to obtain a data feature vector f. The spatio-temporal self-attention model includes a spatial attention module, a temporal attention module, and a classification layer. The basic units of the spatial attention module and the temporal attention module are multi-head self-attention mechanisms; A knowledge representation module, which is used to construct a domain knowledge graph for emitter intention recognition, vectorize the semantic information of each node in the knowledge graph by using a text embedding model, assign weights to each edge in the knowledge graph, and use the vectorized semantic nodes as the initial input feature vector H 0 , and output a knowledge embedding vector ke through a multi-layer graph convolutional network; A knowledge alignment module, which is used to execute a knowledge graph-based self-attention mechanism by taking the data feature vector f as the query vector and the knowledge embedding vector ke as the key vector and the value vector, and obtain an intention-specific knowledge feature vector f k , and concatenate the data feature vector f and the knowledge feature vector f k to obtain an enhanced feature vector f e ; The classification and recognition module is used to process the enhanced feature vector f e The network output is obtained through a classifier composed of a fully connected layer and SoftMax. The output is the probability of each target intention, and the largest one is selected as the result of the radiation source intention recognition.

9. A computer device, characterized in that, Including: One or more processors; A memory; And one or more programs, where the one or more programs are stored in the memory and are configured to be executed by the one or more processors. When the program is executed by the processor, it implements the steps of the method for radiation source intention recognition based on a hybrid drive of knowledge and data as described in any one of claims 1-7.

10. A computer storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the method for radiation source intention recognition based on a hybrid drive of knowledge and data as described in any one of claims 1-7.

Images