Dynamic electronic target sequence intention recognition method and system based on sequence diagram structure

Abstract

The invention discloses a dynamic electronic target sequence intention recognition method and system based on a time sequence diagram structure, and the method comprises the steps: obtaining multi-moment target data for a target object which dynamically changes in a real-time scene; performing sequence division on the multi-moment target data by adopting an FP-Growth algorithm to obtain an electronic target sequence set; modeling and coding each electronic target sequence in the electronic target sequence set by adopting a heterogeneous time sequence diagram model to obtain a heterogeneous time sequence diagram sequence; according to the heterogeneous time sequence diagram sequence, time sequence dynamic hidden relation prediction and time sequence dynamic feature prediction are carried out respectively, and a hidden relation and prediction features are obtained; the hidden relation comprises hidden nodes and hidden edges; and according to the hidden relationship and the prediction features, constructing graph data at the next moment, and performing intention recognition through a graph attention convolutional network. The problem that the time-space incidence relation of the target behavior changing along with time is difficult to capture in the prior art is solved, and the prediction and intention recognition precision is improved.

Description

Technical Field

[0001] The present invention relates to the field of timing analysis, and in particular to a method and system for dynamic electronic target sequence intention recognition based on a timing diagram structure. Background Art

[0002] As a key breakthrough in information confrontation situational awareness, electronic target sequence analysis focuses on behavioral deconstruction and intention mining of cluster targets carrying electronic facilities such as radars and communication equipment. The behavioral pattern evolution, dynamic adjustment of cooperative relationships, and time-varying electromagnetic characteristics of such target groups are key information carriers for mining the opponent's tactical intentions, and their analysis accuracy directly affects the timeliness and accuracy of command decisions.

[0003] In electronic information confrontation scenarios, electronic target sequences have significant dynamic evolution characteristics: the number of target objects increases and decreases with the mission phase, the collaborative relationship changes dynamically with tactical adjustments, and individual properties (such as electromagnetic parameters) are continuously updated over time.

[0004] However, existing analysis methods are mostly based on static graphs or complex network models, which can only characterize static associations at fixed moments and are difficult to adapt to the analysis needs in dynamic evolution scenarios, resulting in insufficient ability to capture temporal behavior patterns, changes in collaborative relationships, and attribute evolution laws, leading to low prediction and intent recognition accuracy. Specific defects include: (1) Insufficient dynamic modeling capabilities: Traditional methods mostly rely on static complex networks or discrete action sequence modeling, which makes it difficult to capture dynamic characteristics such as the increase and decrease in the number of target objects and the evolution of collaborative relationships. For example, in a complex electromagnetic environment, multi-dimensional data such as the dynamic addition and deletion of platform targets, the heterogeneous parameter distribution of electronic equipment, and the real-time evolution of collaborative relationships are intertwined to form a highly dynamic spatiotemporal association network. Due to the solidified topological structure, traditional static models are difficult to characterize temporal dependencies and dynamic evolution laws, resulting in the analysis timeliness and interpretability being difficult to meet the needs of modern battlefields. (2) Split of spatiotemporal features: Existing technologies often handle spatial associations (such as collaboration between targets) and temporal evolution (such as changes in behavior patterns) in isolation, lacking joint modeling of spatiotemporal coupling relationships. Although graph neural networks (GNNs) can effectively extract spatial features of graph structures, their ability to model long-term dependencies is limited; although the Transformer architecture is good at time series analysis, it is difficult to directly handle the spatiotemporal heterogeneity of dynamic graph structures. However, in existing technologies, graph neural networks focus more on static graph reasoning, while Transformers focus on time series modeling. There is a technical gap between the two in terms of dynamic spatiotemporal joint representation. How to build a unified dynamic graph model to achieve spatiotemporal feature fusion and long-term dependency modeling.

[0005] In view of this, this application is hereby filed. Summary of the invention

[0006] The technical problem to be solved by the present invention is that most existing electronic target sequence analysis methods are based on static graphs or complex network models, which can only depict static associations at fixed moments and are difficult to adapt to the analysis requirements in dynamic evolution scenarios, resulting in insufficient ability to capture temporal behavior patterns, changes in collaboration relationships, and attribute evolution laws, and leading to low prediction and intention recognition accuracy.

[0007] The purpose of the present invention is to provide a dynamic electronic target sequence intention recognition method and system based on a temporal graph structure, innovatively proposing a temporal graph dynamic modeling framework. Through heterogeneous node representation, temporal slice encoding, and spatio-temporal joint reasoning, it breaks through the inherent limitations of static models and provides high-precision and strongly interpretable technical support for tasks such as feature prediction and intention recognition of electronic target sequences. The present invention solves the problem that it is difficult for existing technologies to capture the spatio-temporal correlation relationship of target behaviors changing over time, solves the problem that restricts the analysis accuracy of the dynamic evolution characteristics of electronic target sequences in the spatio-temporal dimension, and improves the prediction and intention recognition accuracy.

[0008] The present invention is realized through the following technical solutions:

[0009] In the first aspect, the present invention provides a dynamic electronic target sequence intention recognition method based on a temporal graph structure, and the method includes:

[0010] For the target objects that dynamically change in real-time scenarios, obtain target data at multiple moments; use the FP-Growth algorithm to perform sequence partitioning on the target data at multiple moments to obtain a set of electronic target sequences;

[0011] Use a heterogeneous temporal graph model to model and encode each electronic target sequence in the set of electronic target sequences to obtain a heterogeneous temporal graph sequence; the heterogeneous temporal graph sequence is a graph structure for structurally representing each electronic target sequence in the set of electronic target sequences;

[0012] According to the heterogeneous temporal graph sequence, perform temporal dynamic hidden relationship prediction and temporal dynamic feature prediction respectively to obtain hidden relationships and predicted features; the hidden relationships include hidden nodes and hidden edges;

[0013] According to the hidden relationships and predicted features, construct the graph data for the next moment; based on the graph data for the next moment, perform intention recognition through a graph attention convolutional network.

[0014] Further, using the FP-Growth algorithm to perform sequence partitioning on the target data at multiple moments to obtain a set of electronic target sequences includes:

[0015] According to the target data at multiple moments, traverse the target data at all moments, extract the frequently co-occurring electronic targets and use them as frequent items, and all the frequently occurring targets form a frequent item set;

[0016] Based on the electronic targets, a compressed prefix tree is generated based on frequent itemsets to quickly locate highly correlated electronic target sequences;

[0017] Based on the highly correlated electronic target sequences, through pruning and backtracking strategies, target formations with strong spatio-temporal coordination characteristics are divided to form a set of electronic target sequences.

[0018] Furthermore, using a heterogeneous temporal graph model, each electronic target sequence in the set of electronic target sequences is modeled and encoded to obtain heterogeneous temporal graph sequences, including:

[0019] According to a fixed time window, each electronic target sequence in the set of electronic target sequences is sliced, and each slice is independently modeled as a graph structure;

[0020] A mask matrix is introduced to unify the feature matrices of each slice to the maximum node dimension to form a temporal graph sequence after mask alignment;

[0021] According to the Data class, the node feature matrix and the edge feature matrix of the temporal graph sequence after mask alignment are concatenated to realize spatio-temporal joint encoding of the temporal graph sequence and obtain a heterogeneous temporal graph sequence.

[0022] Furthermore, the nodes of the graph structure represent platform targets or electronic devices, the node attributes are dynamic characteristics such as electromagnetic parameters of the platform targets or electronic devices, the edges represent the cooperation relationships between targets (such as communication, association), and the edge attributes indicate that the types of edges are dynamically updated with slices.

[0023] Furthermore, using a heterogeneous temporal graph model, each electronic target sequence in the set of electronic target sequences is modeled and encoded to obtain heterogeneous temporal graph sequences, and also includes:

[0024] In the process of independently modeling each slice as a graph structure, non-numerical features are mapped to numerical values through a dynamically extended mapping table.

[0025] Furthermore, according to the heterogeneous temporal graph sequences, temporal dynamic hidden relationship prediction is performed, including:

[0026] According to the heterogeneous temporal graph sequences, a node completion algorithm based on a graph convolutional network is used to predict hidden nodes to obtain hidden nodes;

[0027] According to the heterogeneous temporal graph sequences, a TransGCNE graph structure prediction model is used to predict edges to obtain hidden edges;

[0028] Among them, the TransGCNE graph structure prediction model includes an encoder and a decoder. The encoder includes a Transformer module and a GCN module (Graph Convolutional Network module). The weight H of the GCN module is used as the input of the Transformer module. The Transformer module calculates and extracts temporal features and outputs hidden states. Then, the hidden states output by the Transformer module are used to update the weight of the GCN module. And the heterogeneous temporal graph sequence is input into the GCN module with updated weights for calculation and output.

[0029] In the above technical solution, the TransGCNE graph structure prediction model combines the Transformer module and the GCN module, and uses the Transformer module to update the parameters of the GCN module to better extract temporal features.

[0030] Furthermore, according to the heterogeneous temporal graph sequence, temporal dynamic feature prediction is performed, including:

[0031] According to the heterogeneous temporal graph sequence, the TransCGAT feature parameter prediction model is used to perform temporal dynamic feature prediction to obtain predicted features;

[0032] Among them, the TransCGAT feature parameter prediction model includes a GAT processing module, a CNN temporal encoding module, and a Transformer encoder connected in sequence. The heterogeneous temporal graph sequence is used as the input of the GAT processing module. The GAT processing module aggregates the node features to obtain a new feature matrix. Then, the new feature matrix output by the GAT processing module is input into the CNN temporal encoding module for shape transformation, and finally, it is output after being encoded by the Transformer encoder.

[0033] In a second aspect, the present invention further provides a dynamic electronic target sequence intention recognition system based on a temporal graph structure. The system includes:

[0034] An acquisition unit for acquiring multi-moment target data for a target object that dynamically changes in a real-time scenario;

[0035] A sequence division unit for using the FP-Growth algorithm to divide the multi-moment target data to obtain an electronic target sequence set;

[0036] A temporal graph modeling and encoding unit for using a heterogeneous temporal graph model to model and encode each electronic target sequence in the electronic target sequence set to obtain a heterogeneous temporal graph sequence; the heterogeneous temporal graph sequence is a graph structure for structurally representing each electronic target sequence in the electronic target sequence set.

[0037] A prediction unit, which respectively performs temporal dynamic hidden relationship prediction and temporal dynamic feature prediction according to the heterogeneous temporal graph sequence to obtain a hidden relationship and predicted features; the hidden relationship includes hidden nodes and hidden edges.

[0038] An intention recognition unit, which is used to construct graph data for the next moment according to the hidden relationship and predicted features; and perform intention recognition through a graph attention convolutional network based on the graph data for the next moment.

[0039] Furthermore, the temporal graph modeling and encoding unit includes:

[0040] A slicing subunit, which slices each electronic target sequence in the electronic target sequence set according to a fixed time window, and independently models each slice as a graph structure; the nodes of the graph structure represent platform targets or electronic devices, the node attributes are dynamic features such as the electromagnetic parameters of the platform targets or electronic devices, the edges represent the cooperation relationships between targets (such as communication, association), and the edge attributes indicate that the edge types are dynamically updated with the slices.

[0041] A mask alignment subunit, which is used to introduce a mask matrix to unify the feature matrices of each slice to the maximum node dimension to form a masked-aligned temporal graph sequence.

[0042] A spatio-temporal joint encoding subunit, which is used to splice the node feature matrix and the edge feature matrix of the masked-aligned temporal graph sequence according to the Data class to implement spatio-temporal joint encoding of the temporal graph sequence and obtain a heterogeneous temporal graph sequence.

[0043] In a third aspect, the present invention further provides a computer-readable storage medium, which stores a computer program, and when the computer program is executed by a processor, it implements the above-mentioned method for identifying the intention of a dynamic electronic target sequence based on a temporal graph structure.

[0044] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0045] 1. The method and system for identifying the intention of a dynamic electronic target sequence based on a temporal graph structure of the present invention innovatively proposes a temporal graph dynamic modeling framework. Through heterogeneous node representation, temporal slice encoding, and spatio-temporal joint reasoning, it breaks through the inherent limitations of static models and provides high-precision and strongly interpretable technical support for tasks such as feature prediction and intention recognition of electronic target sequences. The present invention solves the problem that it is difficult for the prior art to capture the spatio-temporal correlation relationship of target behaviors changing over time, solves the problem that leads to limited accuracy in analyzing the dynamic evolution characteristics of electronic target sequences in the spatio-temporal dimension, and improves the prediction and intention recognition accuracy.

[0046] 2. With dynamic modeling capabilities: Most existing methods are based on static graphs or complex networks, which can only describe the association relationships at fixed moments and cannot capture the changes in the number of target objects, the evolution of collaborative behaviors, and the time-varying characteristics of attributes. In contrast, the present invention realizes the full life-cycle tracking of target behavior patterns through a time-series graph modeling framework, combining time slicing technology with a dynamic graph structure.

[0047] 3. Implementing heterogeneous time-series data processing: The invention proposes a dynamic mask alignment mechanism, which realizes the dimension unification of multi-time-slice feature matrices through zero-value filling and mask identification, while reducing the interference of invalid features.

[0048] 4. Feature parameter prediction: In existing technologies, GNN (Graph Neural Network) and Transformer architectures are mostly applied independently, making it difficult to balance local spatial associations and long-time-series dependencies. The present invention designs a TransCGAT model, integrating a graph attention network (GAT), one-dimensional convolution, and a Transformer encoder.

[0049] 5. Enhancement of hidden relationship speculation and intention recognition: The present invention designs a TransGCNE model to realize a dynamic weight update mechanism of GCN based on Transformer, for predicting hidden edge relationships; a node completion algorithm based on a graph convolutional network for predicting hidden nodes; and combining a graph attention network for intention classification. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] The drawings described herein are used to provide a further understanding of the embodiments of the present invention, form a part of this application, and do not limit the embodiments of the present invention. In the drawings:

[0051] Figure 1 It is a schematic diagram of the heterogeneous time-series graph sequence of the present invention;

[0052] Figure 2 It is a flowchart of the method for identifying the intention of a dynamic electronic target sequence based on a time-series graph structure of the present invention;

[0053] Figure 3 It is a flowchart of the time-series sequence division of the present invention;

[0054] Figure 4 It is the mask matrix representation of the present invention;

[0055] Figure 5 It is a structural diagram of the TransGCNE Layer of the present invention;

[0056] Figure 6 It is a flowchart of the hidden edge prediction algorithm of the present invention;

[0057] Figure 7 It is a network structure diagram of the TransCGAT model of the present invention;

[0058] Figure 8 This is the network structure diagram of the GAT processing module in the TransCGAT model of the present invention;

[0059] Figure 9 This is the network structure diagram of the CNN processing module in the TransCGAT model of the present invention;

[0060] Figure 10 This is the flowchart of intention recognition of the present invention;

[0061] Figure 11 This is the structure block diagram of the dynamic electronic target sequence intention recognition system based on the timing diagram structure of the present invention. Detailed implementation manners

[0062] To make the objectives, technical solutions and advantages of the present invention clearer and more understandable, the present invention will be further described in detail below in conjunction with embodiments and the accompanying drawings. The illustrative embodiments and descriptions of the present invention are only used to explain the present invention and are not intended to limit the present invention.

[0063] In the electronic information confrontation scenario, the electronic target sequence has significant dynamic evolution characteristics: the number of target objects increases or decreases with the task stage, the cooperation relationship changes dynamically with the tactical adjustment, and the individual attributes (such as electromagnetic parameters) are continuously updated over time. However, most of the existing electronic target sequence analysis methods are based on static graphs or complex network models, which can only depict the static associations at a fixed moment and are difficult to adapt to the analysis requirements in the dynamic evolution scenario, resulting in insufficient ability to capture the temporal behavior patterns, the changes in cooperation relationships, and the attribute evolution laws, and leading to low prediction and intention recognition accuracy.

[0064] The present invention aims to overcome the above technical bottlenecks and focuses on two core issues: (1) how to design a data structure that can represent the temporal dynamic characteristics to accurately model the structural changes (node addition and deletion, edge relationship evolution) and attribute evolution of the electronic target sequence over time; (2) how to construct an efficient computing framework to solve problems such as inconsistent node dimensions and long temporal dependence difficulties in the dynamic scenario. For this purpose, the present invention proposes a method for dynamic electronic target sequence intention recognition based on the timing diagram structure, innovatively integrating the following technical elements:

[0065] (1) Adopt a heterogeneous timing diagram model, abstract the platform target and electronic equipment as dynamic nodes, map the electromagnetic parameters of the platform target and electronic equipment to the attributes of the nodes, and map the cooperation relationship and the type of edges to time-varying edges and the attributes of the edges, so as to construct a multi-dimensional spatio-temporal association graph sequence, that is, a heterogeneous timing diagram sequence;

[0066] (2) Introduce the time slicing technology to segmentally model the electronic target sequence, and depict the stage evolution characteristics of the behavior of the electronic target sequence through the timing diagram sequence;

[0067] (3) Use a dynamic masking mechanism to align the dimensional differences of nodes in different time slices, and support the unified encoding of dynamic graph structures by graph neural networks (GNNs).

[0068] (4) Propose a new hidden edge prediction network TransGCNE (i.e., the TransGCNE graph structure prediction model, abbreviated as the ransGCNE model): Combine Transformer and GCN, and use Transformer to update the parameters of GCN to predict hidden edges.

[0069] (5) Propose an electronic target sequence feature parameter prediction network TransCGAT (i.e., the TransCGAT feature parameter prediction model, abbreviated as the TransCGAT model): Combine the graph attention convolutional network (GAT), CNN, Transformer architecture and attention mechanism to achieve fine-grained spatio-temporal feature extraction and long-time series dependence extraction, and realize the prediction of electronic target sequence features.

[0070] Through the above technical breakthroughs, the present invention significantly improves the modeling accuracy and computational efficiency of electronic target sequences in dynamic scenarios, and provides more timely and interpretable analysis capabilities for tasks such as electronic target sequence feature prediction and intention recognition.

[0071] The present invention first segments the electronic target sequence by time through time slicing technology to construct a heterogeneous time series graph model, where the nodes represent platform targets and electronic devices, and the edges represent multi-dimensional collaboration relationships; secondly, it uses a masking technology to align the dimensional differences of nodes in different time slices, and combines graph neural networks (GNNs), one-dimensional convolution (CNN) and Transformer architecture to achieve hierarchical spatial feature extraction and long-time series dependence modeling; and intention recognition driven by the graph attention convolutional network. The specific technical solutions include: (1) Division of electronic target sequences based on the FP-Growth algorithm; (2) Dynamic representation of time series graph encoding and mask alignment; (3) Hidden node and edge prediction of the node completion algorithm based on the graph convolutional network and the TransGCNE model; (4) Time series feature prediction based on the TransCGAT model (graph attention network + convolutional network + Transformer); (5) Intention recognition driven by the graph attention convolutional network.

[0072] The present invention breaks through the limitations of traditional static models and significantly improves the prediction accuracy of the dynamic behavior of electronic target sequences and intention recognition in complex electromagnetic environments. Verified by public datasets, the method of the present invention has strict mathematical interpretability in the dimension of time series feature modeling, and its analysis efficiency and accuracy are better than those of the prior art.

[0073] Embodiment 1

[0074] Such asFigure 2 As shown in the figure, the dynamic electronic target sequence intention recognition method based on the timing diagram structure of the present invention includes:

[0075] Step S1, in a complex electromagnetic environment, for the target objects (number increase or decrease, evolution of cooperation relationship) that dynamically change in real-time scenarios, obtain target data at multiple moments; use the FP-Growth algorithm to divide the target data at multiple moments into sequences to obtain an electronic target sequence set;

[0076] In this embodiment, as Figure 3 shown, step S1 specifically includes:

[0077] Step S11, obtain target data at multiple moments;

[0078] Step S12, frequent pattern mining (that is, statistical frequent item sets): According to the target data at multiple moments, traverse the target data at all moments, and extract frequently co-occurring electronic targets (which can be called frequent items) from them; and all frequently occurring targets form a frequent item set;

[0079] Among them, frequently occurring means traversing the target data at all moments. When the number of times a target appears is greater than a set threshold, it is considered that the target frequently appears and is a frequent item. The set threshold is usually set to 3, 4, 5, etc.

[0080] Step S13, construct an FP tree: Based on the electronic targets, generate a compressed prefix tree (FrequentPattern Tree) based on the frequent item set to quickly locate highly relevant electronic target sequences;

[0081] Step S14, divide the time series sequence: According to the highly relevant electronic target sequences, through pruning and backtracking strategies, divide out target formations with strong spatio-temporal cooperation characteristics to form an electronic target sequence set.

[0082] Step S2, use a heterogeneous time series graph model to model and encode each electronic target sequence in the electronic target sequence set to obtain a heterogeneous time series graph sequence; as Figure 1 shown. The heterogeneous time series graph sequence is a graph structure for structurally characterizing each electronic target sequence in the electronic target sequence set;

[0083] In this embodiment, step S2 specifically includes:

[0084] Step S21, time slice division: According to a fixed time window, slice each electronic target sequence in the electronic target sequence set, and independently model each slice as a graph structure; the nodes of the graph structure represent platform targets or electronic devices, the node attributes are dynamic characteristics such as the electromagnetic parameters of the platform targets or electronic devices, the edges represent the cooperation relationships between targets (such as communication, association), and the edge attributes indicate that the type of the edge is dynamically updated with the slice;

[0085] Step S22, Non-numerical feature transformation: In the process of independently modeling each slice as a graph structure, non-numerical features are mapped to numerical values through a dynamically extended mapping table, so as to be input into the trained network for analysis. For example, the model number of a platform is "PTXH1", which can be transformed into the numerical value 1 through the mapping table; and in terms of edge attributes, 1 represents the communication relationship and 2 represents the association relationship;

[0086] Step S23, Dynamic mask alignment: To solve the difference in the number of nodes in different time slices, a mask matrix is introduced to unify the feature matrices of each slice to the maximum node dimension, retain the valid node information and fill the invalid nodes with zero values, forming a time series graph sequence after mask alignment, as Figure 4 shown;

[0087] Step S24, Spatiotemporal joint encoding: According to the Data class, the node feature matrix and the edge feature matrix of the time series graph sequence after mask alignment are concatenated to achieve the spatiotemporal joint encoding of the time series graph sequence, and a heterogeneous time series graph sequence is obtained. Among them: The Data class is used to represent graph data, which contains four optional attributes, and each attribute can be flexibly configured according to the specific electromagnetic scenario:

[0088] Node feature matrix (x): It is composed of the feature vectors of each node in the graph structure, and the dimension is [number of nodes, node feature dimension];

[0089] Edge index matrix (edge_index): Describes the connection relationship between each node in the graph structure, and is represented by a sparse matrix, with the dimension of [2, number of edges];

[0090] Edge feature matrix (edge_attr): Characterizes the attribute information of the edge, and the dimension is [number of edges, edge feature dimension];

[0091] Label data (y): Characterizes the node-level or graph-level labels, and the dimension is [number of nodes] or [number of graphs].

[0092] In the above technical solutions, considering that existing methods are mostly based on static graphs or complex networks, which can only describe the association relationships at fixed times and cannot capture the increase or decrease in the number of target objects, the evolution of cooperative behaviors, and the time-varying characteristics of attributes. The present invention has the capabilities of dynamic modeling and heterogeneous time series data processing: (1) Through the time series graph modeling framework, combined with the time slicing technology and the dynamic graph structure, the full life cycle tracking of the target behavior pattern is realized. (2) A dynamic mask alignment mechanism is proposed, and through zero value filling and mask identification, the dimension unification of the multi-time slice feature matrices is realized, and at the same time, the interference of invalid features is reduced.

[0093] Step S3: Based on the heterogeneous time-series graph sequence, perform time-series dynamic hidden relationship prediction and time-series dynamic feature prediction respectively to obtain the hidden relationship and predicted features. The hidden relationship includes hidden nodes and hidden edges.

[0094] In this embodiment, step S3 specifically includes:

[0095] Step S31: Based on the heterogeneous time-series graph sequence, perform time-series dynamic hidden relationship prediction

[0096] (1) Hidden node prediction

[0097] Based on the heterogeneous time-series graph sequence obtained in step S2, use the Network Completion Based on GCN (NCGCN) algorithm to perform hidden node prediction to obtain hidden nodes. Specifically, first randomly select n pairs of node pairs from the graph obtained by modeling the electronic target sequence; according to the selected node pairs, obtain the k-step adjacent nodes of the nodes in each node pair to form the k-hop neighborhood nodes of this node pair; according to the obtained k-hop neighborhood nodes of each node and the original graph, construct a closed subgraph for each node pair; then input the node feature matrix and edge feature matrix of the closed subgraph into the graph convolutional neural network; then calculate whether there are missing nodes between each pair of nodes according to the output of the graph convolutional neural network and whether the missing nodes of different node pairs are the same node; finally, predict the hidden nodes according to the calculation results.

[0098] (2) Hidden edge prediction

[0099] The present invention proposes a dynamic graph sequence edge prediction algorithm TransGCNE based on Transformer and GCN (i.e., the TransGCNE graph structure prediction model, hereinafter referred to as the TransGCNE model). Based on the heterogeneous time-series graph sequence, use the TransGCNE graph structure prediction model to perform edge prediction to obtain hidden edges.

[0100] Specifically, the TransGCNE graph structure prediction model includes an encoder and a decoder. The encoder includes a Transformer module and a GCN module (graph convolutional network module), and the decoder is composed of an inner product and a Sigmoid function. The core of the TransGCNE graph structure prediction model is that the Transformer module and the GCN module form a TransGCNE Layer. The structure diagram of the TransGCNE Layer is as shown in Figure 5As shown. The Transformer module is used to extract the temporal features of the graph sequence, and the output of the Transformer module updates the parameters of the GCN module. The specific process is as follows: First, the encoded electronic target sequence is input into the network and encoded through two TransGCNE layers to extract the spatio-temporal features of the graph sequence. Then, the output of the encoder is used as the input of the decoder for decoding to obtain the probability matrix of the existence of each edge. If the probability of an edge existing is greater than the threshold, the edge exists; otherwise, the edge does not exist. Usually, the threshold is set to 0.5. The flow of the hidden edge prediction algorithm is as Figure 6 shown.

[0101] Figure 6 In, the weight H of the GCN module is used as the input of the Transformer module. The temporal features are calculated and extracted through the Transformer module, and the hidden state is output. Then, the hidden state output by the Transformer module is used to update the weight of the GCN module. And the heterogeneous temporal graph sequence is input into the GCN module with updated weights for calculation and output.

[0102] The calculation formula of the encoder is as follows:

[0103]

[0104] The calculation formula of the decoder is as follows:

[0105]

[0106] In the formula, The dynamic weight matrix at the l-th layer at time step t, which is used to adjust the parameters of graph convolution;

[0107] Transformer(.): Processes the hidden states of the past K time steps to generate dynamic weights;

[0108] The sequence of hidden states within the time window [t - k, t - 1];

[0109] The static weight matrix of the l-th layer;

[0110] The node representation at the l-th layer at time step t;

[0111] The adjacency matrix with self-loops I is the identity matrix

[0112] The degree matrix of

[0113] H (l) : The output of the (l - 1)-th layer

[0114] σ: Activation function

[0115] P ij : Probability that there is an edge between node i and node j;

[0116] h i ,h j : Final representations of nodes i and j;

[0117] Sigmoid: Sigmoid function;

[0118] In the above technical solutions, considering that the existing combination of GCN and Transformer is a serial structure, the output of GCN is used as the input of Transformer for calculation to solve the graph sequence calculation with a fixed graph structure, and no mask is combined to solve the graph sequence data with different numbers of nodes. The designed TransGCNE graph structure prediction model of the present invention combines the Transformer module and the GCN module, and uses the Transformer module to update the parameters of the GCN module to better extract temporal features. The present invention has the following advantages:

[0119] a. Dynamically adjust the weights of GCN using Transformer to better adapt to temporal changes;

[0120] b. During the process of weight update, matrix dimension compression (d 2 -> d×d) is involved, reducing the computational complexity;

[0121] c. Combine masks to solve graph sequence data with different numbers of nodes.

[0122] Step S32, perform temporal dynamic feature prediction according to the heterogeneous temporal graph sequence

[0123] According to the heterogeneous temporal graph sequence, use the TransCGAT feature parameter prediction model to perform temporal dynamic feature prediction to obtain predicted features;

[0124] Specifically, the TransCGAT feature parameter prediction model is a hybrid deep learning network for spatio-temporal graph sequence prediction, which integrates the triple feature learning mechanisms of the GAT processing module, the CNN local temporal encoding module, and the Transformer encoder to achieve fine-grained spatio-temporal feature extraction and prediction.

[0125] GAT processing module: Dynamically weight neighbor nodes through attention coefficients to strengthen key relationship extraction;

[0126] CNN local temporal encoding module: Capture short-term temporal patterns through a sliding window;

[0127] Transformer Encoder: The self-attention mechanism models long-range dependencies across time slices.

[0128] The comparison between the TransCGAT model and existing methods on multiple datasets is shown in Table 1.

[0129] Table 1

[0130]

[0131] The structure of the TransCGAT feature parameter prediction model (abbreviated as TransCGAT model) is as Figure 7 shown. The specific process is as follows: Input the heterogeneous time-series graph sequence into the TransCGAT feature parameter prediction model. First, it passes through the GAT processing module, and the structure of the GAT processing module is as Figure 8 shown. Inside the GAT processing module, it passes through three GATConv layers to aggregate node features and obtain a new feature matrix. Then, the output of the GAT processing module is reshaped to meet the input format of the CNN local time-series encoding module, and the structure of the CNN local time-series encoding module is as Figure 9 shown. Inside the CNN local time-series encoding module, the feature matrix first passes through a one-dimensional convolution (K = 3), then through a max-pooling layer (K = 2), then through another one-dimensional convolution (K = 3), and finally through a max-pooling layer (K = 2) for output. Finally, the output of the CNN local time-series encoding module is reshaped to meet the input format of the Transformer encoder, and the output of the Transformer encoder is output through a fully connected layer. In particular, the model retains the feature slices of the last Pre_len time steps (Pre_len is the prediction step length) for output.

[0132] The self-attention formula of the Transformer encoder is:

[0133]

[0134] where Q, K, and V are the query, key, and value matrices, and d k is the dimension scaling factor.

[0135] In the above technical solution, considering that the existing similar models are the serial combination of GCN (Graph Convolutional Network), TCN (Temporal Convolutional Neural Network) or GCN (Graph Convolutional Network), Transformer, which are used to predict sequences with a fixed graph structure and do not combine masks to solve graph sequence data with different numbers of nodes. Among them, TCN needs multiple layers of dilated convolution to extract temporal features, with a relatively high computational complexity, and GCN does not combine the attention mechanism and cannot highlight some important nodes. In contrast, the GAT processing module, 1D-CNN local temporal encoding module, and Transformer of the present invention are serially combined, having the following advantages:

[0136] a. By using GAT to introduce the attention mechanism, the weight coefficients between nodes can be calculated, and larger weights can be given to important nodes when dynamically updating node features. When aggregating features, important nodes can be prominently highlighted, and spatial features can be better extracted;

[0137] b. Using 1D-CNN can extract local temporal features, and the time dimension can be compressed to 1 / 4 of the original length through downsampling, significantly reducing the computational amount of the subsequent Transformer;

[0138] c. Using Transformer to extract global temporal features, adopting CNN+Transformer, through local temporal feature extraction + global temporal feature extraction, can better extract temporal features and better handle the situation where short-term mutations and global cycles coexist, without being biased towards either the local or global aspect;

[0139] d. Combining masks to solve graph sequence data with different numbers of nodes.

[0140] Step S4: Construct the graph data at the next moment according to the hidden relationship and prediction features; based on the graph data at the next moment, perform intention recognition through the graph attention convolutional network.

[0141] Combining the prediction features and the hidden relationship, construct the graph data at the next moment, and realize intention recognition through the graph attention convolutional network (GAT). The specific process is as follows: First, input the encoded graph data into the trained intention recognition network, use two GATConv layers to aggregate the node features to obtain a new feature matrix, and then input the new feature matrix into a linear classifier to obtain the intention label of the electronic target sequence. The core idea of GAT is that in each graph attention layer, it is necessary to calculate the attention weight coefficient between node i and neighbor j to highlight the proportion of important nodes. The intention recognition process is as Figure 10 shown.

[0142] (1) Attention weight calculation: The attention coefficient between node i and neighbor j is:

[0143]

[0144] Among them, W is a learnable weight matrix, and a is an attention vector;

[0145] (2) Feature aggregation and classification: The node features are updated as:

[0146]

[0147] Finally, the probability of the intent label is output through a linear classifier, and the maximum value is selected as the recognition result.

[0148] Embodiment 2

[0149] As Figure 11 shown, the difference between this embodiment and Embodiment 1 is that this embodiment provides a dynamic electronic target sequence intent recognition system based on a temporal graph structure, and this system corresponds one-to-one with the dynamic electronic target sequence intent recognition method based on a temporal graph structure in Embodiment 1; this system includes:

[0150] An acquisition unit, configured to acquire multi-moment target data for a target object that dynamically changes in a real-time scenario;

[0151] A sequence partitioning unit, configured to partition the multi-moment target data by using the FP-Growth algorithm to obtain an electronic target sequence set;

[0152] A temporal graph modeling and encoding unit, configured to model and encode each electronic target sequence in the electronic target sequence set by using a heterogeneous temporal graph model to obtain a heterogeneous temporal graph sequence; the heterogeneous temporal graph sequence is a graph structure for structurally characterizing each electronic target sequence in the electronic target sequence set;

[0153] A prediction unit, which respectively performs temporal dynamic hidden relationship prediction and temporal dynamic feature prediction according to the heterogeneous temporal graph sequence to obtain a hidden relationship and prediction features; the hidden relationship includes hidden nodes and hidden edges;

[0154] An intent recognition unit, configured to construct graph data for the next moment according to the hidden relationship and prediction features; and perform intent recognition through a graph attention convolutional network based on the graph data for the next moment.

[0155] As a further implementation, the temporal graph modeling and encoding unit includes:

[0156] A slicing subunit, which is used to slice each electronic target sequence in the set of electronic target sequences according to a fixed time window, and independently model each slice as a graph structure; the nodes of the graph structure represent platform targets or electronic devices, the node attributes are dynamic characteristics such as the electromagnetic parameters of the platform targets or electronic devices, the edges represent the collaboration relationships between targets (such as communication, association), and the edge attributes indicate that the type of the edge is dynamically updated with the slice;

[0157] A mask alignment subunit, which is used to introduce a mask matrix to unify the feature matrices of each slice to the maximum node dimension, and form a time-series graph sequence after mask alignment;

[0158] A spatio-temporal joint encoding subunit, which is used to splice the node feature matrix and the edge feature matrix of the time-series graph sequence after mask alignment according to the Data class, realize the spatio-temporal joint encoding of the time-series graph sequence, and obtain a heterogeneous time-series graph sequence.

[0159] Among them, the execution processes of each unit can be carried out according to the flow steps of the method for identifying the intention of the dynamic electronic target sequence based on the time-series graph structure in Embodiment 1, and will not be elaborated one by one in this embodiment.

[0160] Meanwhile, the present invention also provides a computer-readable storage medium, which stores a computer program, and when the computer program is executed by a processor, it implements the above-mentioned method for identifying the intention of the dynamic electronic target sequence based on the time-series graph structure.

[0161] Those skilled in the art should understand that the embodiments of the present application can be provided as a method, a system, or a computer program product. Therefore, the present application can take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present application can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk memories, CD-ROMs, optical memories, etc.) containing computer-usable program codes.

[0162] The present application is described with reference to the flowcharts and / or block diagrams of the method, device (system), and computer program product according to the embodiments of the present application. It should be understood that each process and / or block in the flowchart and / or block diagram, and the combination of the processes and / or blocks in the flowchart and / or block diagram 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 means for realizing the functions specified in one process Figure 1 one process or multiple processes and / or blocks Figure 1 one block or multiple blocks.

[0163] These computer program instructions can also be stored in a computer-readable memory that can direct a computer or other programmable data processing apparatus to operate in a particular manner, such that the instructions stored in the computer-readable memory produce a manufacture including an instruction device that implements the functions specified in one or more of the processes Figure 1 one or more processes and / or blocks Figure 1 specified in one or more of the blocks.

[0164] These computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process, whereby the instructions executed on the computer or other programmable apparatus provide steps for implementing the functions specified in one or more of the processes Figure 1 one or more processes and / or blocks Figure 1 specified in one or more of the blocks.

[0165] The specific embodiments described above further elaborate on the objectives, technical solutions, and beneficial effects of the present invention. It should be understood that the above descriptions are only specific embodiments of the present invention and are not intended to limit the protection scope of the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present invention shall be included within the protection scope of the present invention.

Claims

1. A method for identifying the intention of a dynamic electronic target sequence based on a timing diagram structure, characterized in that The method includes: For a target object that dynamically changes in a real-time scenario, obtaining target data at multiple moments; using the FP-Growth algorithm to perform sequence partitioning on the target data at multiple moments to obtain a set of electronic target sequences; Using a heterogeneous temporal graph model to model and encode each electronic target sequence in the set of electronic target sequences to obtain a heterogeneous temporal graph sequence; According to the heterogeneous temporal graph sequence, performing temporal dynamic hidden relationship prediction and temporal dynamic feature prediction respectively to obtain a hidden relationship and predicted features; the hidden relationship includes hidden nodes and hidden edges; According to the hidden relationship and predicted features, constructing graph data for the next moment; based on the graph data for the next moment, performing intent recognition through a graph attention convolutional network.

2. The method for identifying the intention of a dynamic electronic target sequence based on a timing diagram structure according to claim 1, characterized in that, Using the FP-Growth algorithm to perform sequence partitioning on the target data at multiple moments to obtain a set of electronic target sequences, including: According to the target data at multiple moments, traversing the target data at all moments, extracting frequently co-occurring electronic targets and taking them as frequent items, and all frequently occurring targets form a frequent item set; Based on the electronic targets, generating a compressed prefix tree based on the frequent item set to locate highly correlated electronic target sequences; According to the highly correlated electronic target sequences, through pruning and backtracking strategies, dividing out target formations with strong spatio-temporal cooperation characteristics to form a set of electronic target sequences.

3. The dynamic electronic target sequence intention recognition method based on the timing diagram structure according to claim 1, characterized in that Using a heterogeneous temporal graph model to model and encode each electronic target sequence in the set of electronic target sequences to obtain a heterogeneous temporal graph sequence, including: Slicing each electronic target sequence in the set of electronic target sequences according to a fixed time window, and independently modeling each slice as a graph structure; Introducing a mask matrix to unify the feature matrices of each slice to the maximum node dimension to form a masked-aligned temporal graph sequence; According to the Data class, splicing the node feature matrix and edge feature matrix of the masked-aligned temporal graph sequence to realize spatio-temporal joint encoding of the temporal graph sequence and obtain a heterogeneous temporal graph sequence.

4. The method for identifying the intention of a dynamic electronic target sequence based on a timing diagram structure according to claim 3, wherein The node of the graph structure represents a platform target or an electronic device, the node attribute is the dynamic electromagnetic parameter feature of the platform target or the electronic device, the edge represents the cooperation relationship between targets, and the edge attribute indicates that the type of the edge is dynamically updated with the slice.

5. The method for identifying the intention of a dynamic electronic target sequence based on a timing diagram structure according to claim 1, wherein Using a heterogeneous temporal graph model to model and encode each electronic target sequence in the set of electronic target sequences to obtain a heterogeneous temporal graph sequence, further including: In the process of independently modeling each slice as a graph structure, mapping non-numerical features to numerical values through a dynamically extended mapping table.

6. The method for identifying the intention of a dynamic electronic target sequence based on a timing diagram structure according to claim 1, characterized in that, According to the heterogeneous temporal graph sequence, performing temporal dynamic hidden relationship prediction, including: According to the heterogeneous temporal graph sequence, using a node completion algorithm based on a graph convolutional network to predict hidden nodes to obtain hidden nodes; According to the heterogeneous temporal graph sequence, using a TransGCNE graph structure prediction model to perform edge prediction to obtain hidden edges; The TransGCNE graph structure prediction model includes an encoder and a decoder. The encoder includes a Transformer module and a GCN module. The weights of the GCN module are used as the input of the Transformer module. The Transformer module calculates and extracts temporal features and outputs hidden states. Then, the hidden states output by the Transformer module are used to update the weights of the GCN module. And the heterogeneous temporal graph sequence is input into the GCN module with updated weights for calculation and output.

7. The method for identifying the intention of a dynamic electronic target sequence based on a timing diagram structure according to claim 1, wherein According to the heterogeneous temporal graph sequence, temporal dynamic feature prediction is performed, including: According to the heterogeneous temporal graph sequence, a TransCGAT feature parameter prediction model is used to perform temporal dynamic feature prediction to obtain predicted features. The TransCGAT feature parameter prediction model includes a GAT processing module, a CNN temporal encoding module, and a Transformer encoder connected in sequence. The heterogeneous temporal graph sequence is used as the input of the GAT processing module. The GAT processing module aggregates the node features to obtain a new feature matrix. Then, the new feature matrix output by the GAT processing module is input into the CNN temporal encoding module for shape transformation, and finally, it is encoded and output by the Transformer encoder.

8. A dynamic electronic target sequence intention recognition system based on a timing diagram structure, characterized in that, The system includes: An acquisition unit for acquiring multi-moment target data for a target object that dynamically changes in a real-time scenario. A sequence division unit for dividing the multi-moment target data using the FP-Growth algorithm to obtain an electronic target sequence set. A temporal graph modeling and encoding unit for modeling and encoding each electronic target sequence in the electronic target sequence set using a heterogeneous temporal graph model to obtain a heterogeneous temporal graph sequence; the heterogeneous temporal graph sequence is a graph structure that structurally represents each electronic target sequence in the electronic target sequence set. A prediction unit that respectively performs temporal dynamic hidden relationship prediction and temporal dynamic feature prediction according to the heterogeneous temporal graph sequence to obtain hidden relationships and predicted features; the hidden relationships include hidden nodes and hidden edges. An intention recognition unit for constructing graph data for the next moment according to the hidden relationships and predicted features; based on the graph data for the next moment, intention recognition is performed through a graph attention convolutional network.

9. The method for identifying the intention of a dynamic electronic target sequence based on a timing diagram structure according to claim 8, wherein The temporal graph modeling and encoding unit includes: A slicing subunit for slicing each electronic target sequence in the electronic target sequence set according to a fixed time window and independently modeling each slice as a graph structure; the node of the graph structure represents a platform target or an electronic device, the node attribute is the dynamic electromagnetic parameter feature of the platform target or the electronic device, the edge represents the cooperation relationship between targets, and the edge attribute indicates that the type of the edge is dynamically updated with the slice. A mask alignment subunit for introducing a mask matrix to unify the feature matrices of each slice to the maximum node dimension to form a masked-aligned temporal graph sequence. A spatio-temporal joint encoding subunit, which is used to splice the node feature matrix and the edge feature matrix of the time series graph sequence after mask alignment according to the Data class, realize the spatio-temporal joint encoding of the time series graph sequence, and obtain a heterogeneous time series graph sequence.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the method for identifying the intention of a dynamic electronic target sequence based on a time series graph structure according to any one of claims 1 to 7.

Images