A radar network observable behavior online prediction method, device and electronic equipment in a dynamic electronic countermeasure environment

Abstract

The application provides a radar network observable behavior online prediction method, device and electronic equipment in a dynamic electronic countermeasure environment. The prediction method comprises: obtaining observation records of a radar network in a dynamic electronic countermeasure process; constructing a discrete observation sequence of each radar node according to the observation records of the radar node in a preset historical observation window, and performing feature embedding on the discrete observation sequence to obtain a continuous feature sequence of each radar node; inputting the continuous feature sequence of each radar node into a trained radar behavior prediction model to obtain a working mode probability distribution and an anti-interference action probability distribution of each radar node at a next moment. The application can accurately predict the observable behavior of the radar network at the next moment without directly obtaining internal decision basis, physical situation or coordination graph of the radar network, and improves the applicability in a non-cooperative observation scene.

Description

Technical Field

[0001] This invention relates to the field of radar behavior cognition technology, specifically to an online prediction method, device, and electronic equipment for the observable behavior of a radar network under dynamic electronic countermeasures environments. It is applicable to scenarios where, under non-cooperative observation conditions where it is impossible to obtain information about the internal state of the radar network, the operating mode and anti-jamming actions of the radar network at the next moment are predicted online based on the historical interaction trajectory between the radar network and the jamming group. Background Technology

[0002] Networked multi-functional radar systems typically consist of multiple spatially distributed radar nodes capable of dynamically switching between various mission states, including search, acquisition, tracking, track maintenance, anti-jamming, and cooperative detection. Compared to traditional single, fixed-rule radars, modern radar networks possess stronger adaptability and coordination capabilities in complex electromagnetic environments. The operating modes and anti-jamming actions of different radar nodes are influenced not only by their own historical states but also by factors such as target motion, external interference conditions, inter-node coordination relationships, and changes in mission phases.

[0003] In a dynamic electronic warfare environment, jamming swarms can influence the detection reliability, tracking stability, and anti-jamming response of radar networks through different jamming patterns, target allocation, and variations in jamming intensity. Radar networks, in turn, may dynamically adjust their operating modes and anti-jamming actions based on the degree of interference, historical threat status, and collaborative auxiliary information. Therefore, the externally observable behavior of radar networks typically exhibits clear temporal, collaborative, uncertain, and phased variation characteristics.

[0004] For external observers, the internal state of a radar network, actual threat assessment results, resource scheduling criteria, collaborative decision-making mechanisms, and complete physical situation are typically difficult to obtain directly. The information that can be reliably obtained mainly includes the operating modes and anti-jamming actions exhibited by radar nodes, as well as external interaction records such as the jamming patterns and targets of jamming aircraft. Therefore, how to predict the observable behavior of a radar network in the next moment using limited historical trajectories under partially observable conditions is a key issue in radar behavior understanding, countermeasure situation inference, and electronic warfare-assisted decision-making.

[0005] Existing related technologies mainly include rule-based behavior inference methods, traditional supervised learning methods, behavior cloning methods, generative adversarial learning methods, and inverse reinforcement learning methods. Rule-based methods typically rely on expert experience to construct state transition logic, making it difficult to adapt to changes in radar behavior distribution in complex dynamic adversarial environments. Traditional supervised learning and behavior cloning methods can directly learn the mapping relationship between historical observations and the next action, but they usually tend to fit local state-action relationships and lack the ability to model the cooperative coupling relationships between multiple radar nodes, long-term dependencies, and low-frequency behavior transition samples. Generative adversarial learning methods can characterize behavioral patterns from the perspective of trajectory distribution matching, but the training process is complex, requires high sample stability, and is not easy to explicitly characterize the cooperative structure between radar network nodes. Inverse reinforcement learning methods can be used to infer potential reward functions or policy preferences, but most methods are more suitable for single agents or scenarios with relatively clear internal states, making it difficult to directly adapt to online behavior prediction tasks of radar networks under conditions of incomplete external observation.

[0006] Furthermore, in dynamic electronic warfare, jamming patterns, mission phases, and radar response patterns may change over time, leading to drift in the distribution of radar network behavior. Fixed offline models struggle to absorb newly emerging interaction samples in a timely manner, and their predictive performance is prone to decline during periods of frequent policy changes or behavioral transitions. Therefore, existing technologies still have the following shortcomings: First, it is difficult to model the implicit cooperative state and behavioral uncertainties of radar networks under partially observable conditions; second, it is difficult to simultaneously characterize the temporal evolution characteristics of a single radar node and the spatial cooperative relationships between multiple radar nodes; third, it is difficult to jointly predict operating modes and anti-jamming actions while maintaining the structural correlation between the two.

[0007] To address the aforementioned issues, this invention proposes an online prediction method and system for the observable behavior of radar networks under dynamic electronic countermeasures environments. Through spatiotemporal collaborative representation, conditional variational latent state modeling, joint prediction of operating modes and anti-jamming actions, and an online incremental update mechanism, it achieves accurate prediction and dynamic adaptation of the observable behavior of radar networks at the next moment. Summary of the Invention

[0008] In view of this, the present invention proposes an online prediction method, device, and electronic equipment for the observable behavior of radar networks in a dynamic electronic countermeasures environment, so as to accurately predict and dynamically adapt the observable behavior of radar networks in the next moment. Specifically, the present invention is achieved through the following technical solution:

[0009] According to a first aspect of the embodiments of this specification, an online prediction method for the observable behavior of a radar network in a dynamic electronic countermeasures environment is provided, comprising the following steps:

[0010] Acquire observation records of each radar node in the radar network and each jamming node in the jamming group during dynamic electronic countermeasures processes;

[0011] For each radar node, a discrete observation sequence is constructed based on its historical observation records and feature embedding is performed to obtain a continuous feature sequence.

[0012] The continuous feature sequences of each radar node are input into the trained radar behavior prediction model to obtain the probability distribution of the working mode and the probability distribution of anti-jamming actions of each radar node in the next time step. The radar behavior prediction model includes a spatiotemporal graph encoder, a conditional prior network, a posterior network, and a dual-head prediction network.

[0013] The spatiotemporal graph encoder is used to encode the continuous feature sequences of each radar node in the time dimension to obtain behavioral evolution features. Then, it aggregates each behavioral evolution feature in the spatial dimension according to the attention weight between nodes and fuses it to obtain the spatiotemporal features of each radar node. Finally, it aggregates all spatiotemporal features to obtain the global context features of the radar network.

[0014] The conditional prior network is used to map the global context features, output the distribution parameters of the conditional prior distribution, and obtain the conditional variational latent state features based on the distribution parameters. The conditional variational latent state features are used to characterize the implicit variables that are not reflected in the observation records and affect the subsequent behavior of the radar network.

[0015] The posterior network is used to output the distribution parameters of the variational posterior distribution during the training phase, taking global context features and the observation label of the next time step as inputs; during training, the conditional prior distribution approximates the variational posterior distribution through KL divergence constraints.

[0016] The dual-head prediction network is used to fuse the conditional variational latent state features with the spatiotemporal features of each radar node to obtain spatiotemporal enhancement features, and based on these features, predict the probability distribution of the working mode and the probability distribution of anti-jamming actions of each radar node at the next moment.

[0017] According to a second aspect of the embodiments of this specification, an online prediction device for the observable behavior of a radar network in a dynamic electronic countermeasures environment is provided, the device comprising:

[0018] The data acquisition unit is used to acquire the observation records of each radar node in the radar network and each jamming node in the jamming group during the dynamic electronic countermeasures process.

[0019] The data processing unit is used to construct discrete observation sequences based on the historical observation records of each radar node and perform feature embedding to obtain continuous feature sequences.

[0020] The behavior prediction unit is used to input the continuous feature sequences of each radar node into the trained radar behavior prediction model to obtain the probability distribution of the working mode and the probability distribution of anti-jamming actions of each radar node in the next time step. The radar behavior prediction model includes a spatiotemporal graph encoder, a conditional prior network, a posterior network, and a dual-head prediction network.

[0021] The spatiotemporal graph encoder is used to encode the continuous feature sequences of each radar node in the time dimension to obtain behavioral evolution features. Then, it aggregates each behavioral evolution feature in the spatial dimension according to the attention weight between nodes and fuses it to obtain the spatiotemporal features of each radar node. Finally, it aggregates all spatiotemporal features to obtain the global context features of the radar network.

[0022] The conditional prior network is used to map the global context features, output the distribution parameters of the conditional prior distribution, and obtain the conditional variational latent state features based on the distribution parameters. The conditional variational latent state features are used to characterize the implicit variables that are not reflected in the observation records and affect the subsequent behavior of the radar network.

[0023] The posterior network is used to output the distribution parameters of the variational posterior distribution during the training phase, taking global context features and the observation label of the next time step as inputs; during training, the conditional prior distribution approximates the variational posterior distribution through KL divergence constraints.

[0024] The dual-head prediction network is used to fuse the conditional variational latent state features with the spatiotemporal features of each radar node to obtain spatiotemporal enhancement features, and based on these features, predict the probability distribution of the working mode and the probability distribution of anti-jamming actions of each radar node at the next moment.

[0025] According to a third aspect of the embodiments of this specification, an electronic device is provided, comprising: a processor; and a computer-readable storage medium storing computer program instructions that, when executed by the processor, cause the processor to perform the method as described in the first aspect.

[0026] The embodiments of the present invention have at least the following technical effects:

[0027] (1) The embodiments of the present invention are geared towards the external incomplete observation conditions in dynamic electronic countermeasures environments. Without directly obtaining the internal decision-making basis, physical situation or cooperative intention of the radar network, the observable behavior of the radar network at the next moment can be accurately predicted, which improves the applicability in non-cooperative observation scenarios.

[0028] (2) The embodiments of the present invention simultaneously model the behavioral evolution characteristics of a single radar node and the spatial collaborative correlation between different radar nodes by using a spatiotemporal graph encoder, thereby enhancing the ability to express the changes in the collaborative behavior of multiple nodes.

[0029] (3) The embodiments of the present invention introduce conditional variational latent states, which can model potential behavioral factors and behavioral uncertainties that are difficult to describe explicitly under some observable conditions, thereby improving the predictive ability of working mode transitions and anti-interference action switching. Attached Figure Description

[0030] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the accompanying drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Some specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings in an exemplary and non-limiting manner. The same reference numerals in the drawings indicate the same or similar parts or components. Those skilled in the art should understand that these drawings are not necessarily drawn to scale. In the drawings:

[0031] Figure 1 This is a flowchart illustrating an exemplary embodiment of the present invention of an online prediction method for the observable behavior of a radar network in a dynamic electronic warfare environment;

[0032] Figure 2 This is a schematic diagram of a dynamic electronic countermeasures scenario involving a radar network and a jamming swarm, as illustrated in an exemplary embodiment of the present invention.

[0033] Figure 3 This is a schematic diagram of the network structure of a radar behavior prediction model shown in an exemplary embodiment of the present invention;

[0034] Figure 4 This is an exemplary embodiment of the present invention illustrating a working mode prediction confusion matrix diagram of radar network observable behavior prediction results;

[0035] Figure 5 This is an exemplary embodiment of the present invention illustrating an anti-jamming action prediction confusion matrix diagram of radar network observable behavior prediction results;

[0036] Figure 6 This is an exemplary embodiment of the present invention illustrating a curve showing the change in overall sample prediction accuracy;

[0037] Figure 7 This is a curve illustrating the change in the accuracy of transition sample prediction, as shown in an exemplary embodiment of the present invention.

[0038] Figure 8 This is a block diagram illustrating an electronic device according to an exemplary embodiment of the present invention;

[0039] Figure 9 This is a block diagram illustrating an online prediction device for the observable behavior of a radar network under dynamic electronic countermeasures environment, as shown in an exemplary embodiment of the present invention. Detailed Implementation

[0040] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numerals in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the present invention. Rather, they are merely examples of apparatuses and methods consistent with some aspects of the invention as detailed in the appended claims.

[0041] The terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. The singular forms “a,” “the,” and “the” used in this invention and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used herein refers to and includes any or all possible combinations of one or more of the associated listed items.

[0042] To address the problem that existing technologies struggle to accurately predict the subsequent observable behavior of radar networks in dynamic electronic warfare environments based on incomplete external observation information, this invention proposes an online prediction method and system for the observable behavior of radar networks in dynamic electronic warfare environments. This method uses externally observable records generated during continuous confrontation between the radar network and jamming aircraft as input. Without relying on the radar network's internal decision-making basis, complete physical situation information, or genuine collaborative intentions, it learns the conditional mapping relationship between historical confrontation trajectories and the radar network's next-moment operating mode and anti-jamming actions, thereby achieving online prediction of the radar network's observable behavior.

[0043] The following detailed explanation of each step is provided in conjunction with the accompanying drawings.

[0044] Figure 1 This is a flowchart illustrating an exemplary embodiment of the present invention of an online prediction method for the observable behavior of a radar network in a dynamic electronic countermeasures environment, as shown below. Figure 1 As shown, the online prediction method includes the following steps:

[0045] Step S1: Obtain the observation records of each radar node in the radar network and each jamming node in the jamming group during the dynamic electronic countermeasures process. The observation records include at least the working mode and anti-jamming actions exhibited by each radar node, as well as the jamming patterns and jamming targets exhibited by each jamming node.

[0046] Step S2: For each radar node, a discrete observation sequence is constructed based on its historical observation records and feature embedding is performed to obtain a continuous feature sequence.

[0047] This step constructs a discrete observation sequence for each radar node based on its operating mode and anti-interference actions within a preset historical observation window, as well as the interference patterns and targets of all interfering nodes within the preset historical observation window. Then, it embeds features into the discrete observation sequence to obtain a continuous feature sequence for each radar node.

[0048] Step S3: Input the continuous feature sequences of each radar node into the trained radar behavior prediction model to obtain the probability distribution of the working mode and the probability distribution of anti-jamming actions of each radar node at the next time step; the radar behavior prediction model includes a spatiotemporal graph encoder, a conditional prior network, a posterior network, and a dual-head prediction network.

[0049] The spatiotemporal graph encoder is used to encode the continuous feature sequences of each radar node along the time dimension to obtain behavioral evolution features. Then, based on the attention weights between nodes, it performs spatial dimension aggregation on each behavioral evolution feature and fuses it with the existing features to obtain the spatiotemporal features of each radar node. Finally, it aggregates all spatiotemporal features to obtain the global context features of the radar network. In other words, the spatiotemporal graph encoder encodes the continuous feature sequences of each radar node along the time dimension to obtain the behavioral evolution features of each radar node that characterize the changes of its discrete observation sequences over time. Based on the attention weights between nodes, it performs spatial dimension collaborative correlation aggregation on the behavioral evolution features of each radar node and fuses them with the existing behavioral evolution features to obtain the spatiotemporal features of each radar node. Finally, it aggregates all the spatiotemporal features of the radar nodes to obtain the global context features of the radar network.

[0050] The conditional prior network is used to map the global context features, output the distribution parameters of the conditional prior distribution, and obtain the conditional variational latent state features based on these distribution parameters. Specifically, the conditional prior network performs a nonlinear mapping on the global context features, outputting the distribution parameters of the conditional prior distribution, which describes the probability distribution of the latent state given the global context features; and obtains the conditional variational latent state features based on the distribution parameters of the conditional prior distribution, which characterize implicit variables not reflected in the observation records that influence subsequent changes in the radar network's behavior.

[0051] In this embodiment, the conditional variational latent state refers to a latent probability variable conditioned on historical observation information. This variable follows a conditional prior distribution output by a conditional prior network, including but not limited to implicit factors not directly reflected in external observation records, such as changes in the cooperative relationship between radar network nodes, task phase switching, disturbance assessment, and potential threat judgment. By introducing conditional variational inference, this invention can adaptively generate corresponding latent state representations based on different historical observation scenarios while modeling the uncertainty of subsequent radar network behavior, thereby improving the predictive ability for dynamic behavioral changes such as working mode transitions and anti-interference action switching.

[0052] The posterior network is used during the training phase to take global context features and the next-time observation label as input, and output the distribution parameters of the variational posterior distribution. During training, the conditional prior distribution approximates the variational posterior distribution through KL divergence constraints. That is, during the training phase, the posterior network takes the global context features of the training samples and the embedding features of the next-time observation label as input, and outputs the distribution parameters of the variational posterior distribution. During training, the conditional prior distribution approximates the variational posterior distribution output by the posterior network through KL divergence constraints, enabling the conditional prior network to learn the distribution characteristics of the latent states characterized by the posterior network.

[0053] The dual-head prediction network is used to fuse the conditional variational latent state features with the spatiotemporal features of each radar node to obtain spatiotemporal enhancement features, and based on these features, predict the probability distribution of the operating mode and the probability distribution of anti-jamming actions for each radar node at the next moment. In other words, the dual-head prediction network fuses the latent state features with the spatiotemporal features of each radar node to obtain the spatiotemporal enhancement features of each radar node, and based on these features, predicts the probability distribution of the operating mode and the probability distribution of anti-jamming actions for each radar node at the next moment.

[0054] Figure 1The online prediction method presented here, without relying on the internal state information of the radar network, simultaneously models the temporal evolution of a single radar node and the spatial collaborative correlation between multiple radar nodes through a spatiotemporal graph encoder. This allows the radar behavior prediction model to fully explore the impact of the collaborative coupling relationship between radar network nodes on subsequent behavior. By using conditional variational latent states with historical observation information as conditions, the implicit variables affecting subsequent changes in radar network behavior are probabilistically represented. KL divergence constraints are used to approximate the posterior network with the conditional prior network, enabling the generation of effective latent states containing implicit behavioral information without real labels during the inference stage. This effectively represents the uncertainty and multimodality of radar network behavior, improving the predictive ability for mode transitions and anti-jamming action switching. By jointly outputting the probability distribution of the working mode and the probability distribution of anti-jamming actions through a dual-head prediction network, the structural correlation between the two prediction tasks is maintained, avoiding logical conflicts caused by independent prediction. This invention only requires externally observable records to complete online prediction, providing effective technical support for radar network behavior understanding, adversarial situation assessment, and strategy prediction in non-cooperative observation scenarios.

[0055] Next, combined Figures 2 to 7 This invention details the process of online prediction of the observable behavior of radar networks in a dynamic electronic warfare environment.

[0056] First, step S1 is performed to obtain observation records of the radar network and jamming aircraft.

[0057] like Figure 2 As shown, this embodiment considers a dynamic electronic warfare scenario between a radar network and a jamming swarm. The radar network consists of multiple spatially distributed radar nodes, each with a relatively independent detection perspective and operating status, and can share local observation information or status information through a cooperative network. The jamming swarm consists of multiple jamming nodes, each capable of selecting appropriate jamming patterns and targets at different time steps, affecting the radar nodes' observation quality, operating mode transitions, and anti-jamming action selection.

[0058] In this embodiment, the radar network is assumed to include One radar node, jamming group includes Interference node. Each radar node at time The working mode is Anti-interference action is ;No. Each interference node at time The interference pattern is The target of interference is Then at that time Observation records Represented as:

[0059] (1)

[0060] Next, step S2 is executed to construct historical observation trajectory segments and embed features, which is used to convert continuous observation records into historical input sequences that the radar behavior prediction model can process. Specifically, this includes steps S21 to S23:

[0061] Step S21, using the sequence of events preceding the current time... The observation records at each time step are used as a historical observation window. The observation records are then slidably cropped to obtain historical observation trajectory segments. Let the current time be... Then, the historical observation trajectory fragments are represented as:

[0062] (2)

[0063] Step S22: For each radar node, extract the radar node's location within the historical observation window. The radar node's operating mode and anti-jamming actions are determined, and the jamming patterns and targets of the jamming group at the corresponding time steps are extracted. The discrete information is then correlated and encoded to obtain the node-related discrete observation sequence corresponding to the radar node.

[0064] For the Each radar node, at time The node-related discrete observations can be represented as:

[0065] , (3)

[0066] in, Indicates facing the first Association coding function for each radar node.

[0067] Step S23: Embedding mapping is performed on the node-related discrete observation sequences to transform discrete categorical variables into continuous feature representations, thereby forming the historical input feature sequence of the radar node.

[0068] (4)

[0069] in, Indicates the input embedded function, Indicates the first Each radar node at time The continuous input feature.

[0070] Therefore, the first The continuous characteristic sequence of each radar node within the historical observation window is as follows:

[0071] (5)

[0072] After obtaining the continuous feature sequence of each radar node within the historical observation window, step S3 is executed to input the continuous feature sequence of each radar node into the trained radar behavior prediction model to obtain the probability distribution of the working mode and the probability distribution of anti-jamming action of each radar node in the next moment.

[0073] like Figure 3 As shown, the radar behavior prediction model in this embodiment includes a spatiotemporal graph encoder, a conditional prior network, a posterior network, and a dual-head prediction network, wherein the conditional prior network and the posterior network constitute a conditional variational latent state construction module.

[0074] The spatiotemporal graph encoder in this embodiment includes a time encoding unit, a spatial collaborative modeling unit, a node-level fusion unit, and a global context aggregation unit.

[0075] The continuous feature sequences of each radar node are input into the time coding unit, and recursive coding or sequential coding is performed along the time order of the historical observation window to obtain the time hidden state of each radar node at each time step within the historical observation window.

[0076] Taking a gated loop unit as the time encoding unit as an example, its recursive encoding process is as follows:

[0077] (6)

[0078] in, This represents a gated loop unit function. Indicates the first Each radar node at time The time-hidden state.

[0079] In some examples, the temporal encoding unit may also be a sequence modeling structure employing a long short-term memory network, a temporal convolutional network, or a self-attention temporal encoder.

[0080] The temporal coding unit uses the temporal hidden state at the final moment, or the representation obtained by aggregating multiple temporal hidden states, as the behavioral evolution feature of the radar node.

[0081] For example, when the final hidden state is used as a behavioral evolution feature... At that time, there were:

[0082] (7)

[0083] Then, using the behavioral evolution characteristics of each radar node as node features, a radar node feature set under the current historical observation window is constructed. .

[0084] The attention weights between nodes in the spatial collaborative modeling unit adopt multi-head attention, and the spatiotemporal graph encoder is configured as follows:

[0085] For each attention head of each radar node, the behavioral evolution features of that radar node are mapped to query vectors and key vectors, and the behavioral evolution features of other radar nodes are mapped to value vectors. Based on the query vectors, key vectors, and value vectors, the attention weights of that radar node to other radar nodes are calculated. The value vectors are then weighted and aggregated using these attention weights to obtain the spatial collaborative features of that radar node. The behavioral evolution features and spatial collaborative features of each radar node are concatenated, and after linear mapping and nonlinear activation, the spatiotemporal features of each radar node are obtained.

[0086] Specifically, the spatial collaborative modeling unit, based on the radar node feature set, calculates the attention weight between any radar node and other radar nodes using an inter-node attention mechanism. This attention weight characterizes the strength of the collaborative association between different radar nodes. For the first... The first point of attention, the first Query vector of radar nodes , No. Key vectors of radar nodes Sum value vector They are represented as follows:

[0087] (8)

[0088] in, For the first The learnable parameter matrix corresponding to each attention head.

[0089] No. The radar node for the first Attention weights of each radar node Represented as:

[0090] (9)

[0091] in, Indicates the first Each attention head has a feature dimension. Used to characterize the first [number] in the current historical observation window. The radar node for the first The relative impact of the behavior of each radar node.

[0092] The spatial collaborative modeling unit performs weighted aggregation of node features in the radar node feature set to obtain the spatial collaborative representation corresponding to each radar node. For example, for the first... The first point of attention, the first Spatial Coordination Characteristics of Individual Radar Nodes for:

[0093] (10)

[0094] After concatenating and linearly mapping the outputs of multiple attention heads, we obtain the first... The spatial coordination characteristics of the radar nodes are:

[0095] (11)

[0096] in, Indicates the number of heads of attention. This represents a learnable linear mapping matrix. This indicates a feature splicing operation.

[0097] The node-level fusion unit fuses the behavioral evolution characteristics of each radar node with its corresponding spatial coordination characteristics to obtain the spatiotemporal characteristics of each radar node, where the first... Spatiotemporal characteristics of individual radar nodes It can be represented as:

[0098] (12)

[0099] in, and For learnable parameters, It is a non-linear activation function.

[0100] Spatiotemporal characteristics Simultaneously includes the The historical behavioral evolution information of each radar node and the collaborative context information from other radar nodes.

[0101] The global context aggregation unit performs attention-weighted aggregation of the spatiotemporal features of all radar nodes to obtain the global context features of the radar network. for:

[0102] (13)

[0103] in, These are learnable parameters.

[0104] The conditional variational latent state construction module includes a posterior network and a conditional prior network, used to construct the probability distribution of the latent state under given global context features and generate the corresponding latent state representation. The conditional variational latent state is used to characterize potential collaborative factors, potential threat judgment factors, and phased behavioral change factors that are difficult to obtain directly from external observation records under partially observable conditions. Specifically, it includes the following steps:

[0105] The posterior network is applied during the training phase. The global context features of the training samples and the embedding features of the observation labels at the next time step are input into the posterior network to obtain the mean and standard deviation parameters of the posterior distribution.

[0106] During the training phase, the radar network behavior label for the next time step is available. Therefore, this real behavior label is used to assist in constructing the posterior distribution, enabling the latent state to learn implicit factors related to changes in operating mode and anti-interference actions in the next time step. Let the real observable behavior label of the radar network at the next time step be... for:

[0107] (14)

[0108] During the training phase, construct the posterior distribution. :

[0109] (15)

[0110] in, Represents the conditional variational latent state. and Let represent the mean vector and standard deviation vector of the posterior distribution, respectively.

[0111] The posterior distribution parameters can be generated by the posterior network:

[0112] (16)

[0113] in, Represents the posterior network. This indicates that the feature embedding operation is performed on the next-time observation label of the radar network.

[0114] The posterior network is constructed using a multi-layer fully connected neural network, specifically including an input concatenation layer, a linear mapping layer, a non-linear activation layer, and a parameter output layer. The input concatenation layer is used to incorporate the global contextual features of the training samples. The embedded features of the next time-lapse observation label The system performs splicing; a linear mapping layer is used to perform linear transformation on the spliced ​​features to extract high-dimensional features related to the potential behavior of radar nodes; a nonlinear activation layer is used to perform nonlinear activation on the high-dimensional features; and a parameter output layer, connected after the nonlinear activation layer, includes a mean output header and a variance output header, wherein the mean output header is used to output the mean vector of the variational posterior distribution. The variance output header is used to output the standard deviation vector of the variational posterior distribution. .

[0115] Thus, the posterior network in this embodiment uses known real behavioral results from the training data to guide the latent state to learn the implicit factors behind the subsequent behavioral changes of the radar network, including changes in cooperative relationships, changes in task stages, changes in the degree of disturbance, and changes in the judgment of potential threats.

[0116] During the inference phase, the global context features of the radar network are input into the conditional prior network to obtain the mean and standard deviation parameters of the conditional prior distribution. Since the radar network behavior labels for the next time step are unavailable during the inference phase, a conditional prior distribution relying solely on historical observation trajectories is constructed. :

[0117] (17)

[0118] in, and Let represent the mean vector and standard deviation vector of the conditional prior distribution, respectively.

[0119] In this embodiment, the conditional prior distribution parameters are generated by a conditional prior network:

[0120] (18)

[0121] in, This represents a conditional prior network.

[0122] The conditional prior network is constructed using a multi-layer fully connected neural network, including an input layer, a linear mapping layer, a nonlinear activation layer, and a parameter output layer. The input layer receives the global context features; the linear mapping layer performs a linear transformation on the global context features to extract high-dimensional features related to subsequent changes in the radar network's behavior; the nonlinear activation layer performs nonlinear activation on the high-dimensional features; and the parameter output layer, connected after the nonlinear activation layer, includes a mean output header and a variance output header, the mean output header outputting the mean vector of the conditional prior distribution. The variance output header is used to output the standard deviation vector of the conditional prior distribution. .

[0123] Because the conditional prior network does not rely on the next-time observation label, it can be directly used in actual online prediction processes. Based on the global context vector formed by historical observation trajectories, the network estimates the potential cooperative states, potential threat judgment states, and phased behavioral change states that may exist at the current time, thus providing implicit state support for subsequent operational mode prediction and anti-interference action prediction.

[0124] It is worth noting that, in order to enable the conditional prior network in the inference phase to learn the latent behavioral factors characterized by the posterior network in the training phase, a KL regularization constraint is introduced between the variational posterior distribution and the conditional prior distribution during training, keeping them close in the latent space. Through this constraint, the latent behavioral information learned by the posterior network from the next-time observation labels during training can be approximately represented by the conditional prior network. Therefore, even if the next-time observation labels cannot be obtained during the inference phase, an effective latent state representation can still be generated through the conditional prior network.

[0125] After obtaining the distribution parameters of the conditional prior distribution, the conditional prior network is configured to: obtain the distribution parameters of the conditional prior distribution. The conditional variational latent state features are obtained by sampling from the middle, or the mean vector of the conditional prior distribution is directly used as the conditional variational latent state features. .

[0126] Through the above process, the posterior network in the training phase and the conditional prior network in the inference phase form a synergistic relationship: the posterior network uses the next-time observation label to enhance its ability to learn the latent state, and the conditional prior network learns to approximate the variational posterior distribution through KL regularization constraints, and independently generates the conditional variational latent state features in the inference phase.

[0127] After obtaining the conditional variational latent state features, the conditional variational latent state features and the spatiotemporal features of each radar node are input into the dual-head prediction network to obtain the probability distribution of the working mode and the probability distribution of anti-jamming actions of each radar node at the next moment.

[0128] The dual-head prediction network is a neural network module based on a shared feature fusion module and parallel dual classification heads. Specifically, the dual-head prediction network includes a shared feature fusion module, an operating mode prediction branch, and an anti-interference action prediction branch.

[0129] The shared feature fusion module is used to splice and nonlinearly map the spatiotemporal features of each radar node with the conditional variational latent state features to generate spatiotemporal enhancement features for each radar node.

[0130] In some embodiments, the shared feature fusion module adopts a multilayer perceptron structure, including a sequentially connected feature splicing layer, a first fully connected layer, a nonlinear activation layer, a normalization layer, a Dropout layer, and a second fully connected layer; the working mode prediction branch includes a sequentially connected fully connected layer, a nonlinear activation layer, and a Softmax output layer; the anti-interference action prediction branch includes a sequentially connected connected layer, a nonlinear activation layer, and a Softmax output layer.

[0131] The operating mode prediction branch is used to output the operating mode probability distribution of each radar node at the next moment based on the spatiotemporal enhancement features.

[0132] In some embodiments, the working mode prediction branch is a first classification head, which is composed of a fully connected layer, a nonlinear activation layer and a Softmax output layer. The number of neurons in its output layer is consistent with the number of preset working mode categories, and it is used to output the probability distribution of each working mode category at the next time step.

[0133] The anti-jamming action prediction branch is used to output the probability distribution of anti-jamming actions of each radar node at the next moment based on the spatiotemporal enhancement features.

[0134] In some embodiments, the anti-interference action prediction branch is a second classification head, which is composed of a fully connected layer, a nonlinear activation layer and a Softmax output layer. The number of neurons in its output layer is consistent with the number of preset anti-interference action categories, and it is used to output the probability distribution of each anti-interference action category at the next time step.

[0135] In this embodiment, the two prediction branches share the same augmentation node representation as input, but each has its own independent classification network parameters and output layer. This allows the working mode prediction and anti-interference action prediction to utilize the common spatiotemporal collaborative context and latent state information while maintaining their own independent class discrimination capabilities.

[0136] Specifically, the conditional variational latent state features Spatiotemporal representations of each radar node By merging, we obtain the first... Spatiotemporal enhancement features of individual radar nodes :

[0137] (19)

[0138] in, Indicating feature fusion networks, It contains both explicit spatiotemporal coordination information and implicit latent state information.

[0139] Based on the spatiotemporal enhancement features The probability distribution of the working mode is output through the working mode prediction branch and the anti-interference action prediction branch. and anti-interference action probability distribution :

[0140] (20)

[0141] in, Indicates the working mode prediction branch, This represents the anti-interference action prediction branch.

[0142] Therefore, the first Next-moment operating mode of each radar node Anti-interference action prediction results They are respectively:

[0143] (twenty one)

[0144] in, This represents a set of preset working mode categories. This represents a preset set of anti-interference action categories. Indicates the work mode category index, This represents the index of anti-interference action categories.

[0145] Thus, the prediction result of the radar network's observation behavior at the next moment can be expressed as:

[0146] (twenty two)

[0147] In some embodiments, the training process of the radar behavior prediction model includes:

[0148] Acquire training samples, which include continuous feature sequences of each radar node and observation labels for the next time step;

[0149] The training samples are input into the radar behavior prediction model to calculate the joint loss function, which includes the working mode prediction loss, anti-jamming action prediction loss, KL regularization term and consistency constraint term.

[0150] The network parameters of the radar behavior prediction model are optimized based on the joint loss function.

[0151] The predicted loss for the working mode is determined based on the cross-entropy between the probability distribution of the working mode and the actual working mode in the observed label at the next time step.

[0152] The anti-interference action prediction loss is determined based on the cross-entropy between the anti-interference action probability distribution and the actual anti-interference action in the observation label at the next moment.

[0153] The KL regularization term is the KL divergence between the variational posterior distribution and the conditional prior distribution;

[0154] The consistency constraint term, based on the working mode probability distribution, penalizes the probability of action categories in the anti-jamming action probability distribution of each radar node that do not correspond to the working mode.

[0155] Specifically, in order to maintain the structural correlation between working mode prediction and anti-interference action prediction, and to avoid the formation of hard constraint error cascades in working mode misjudgment and anti-interference action prediction, consistency constraints on working mode and anti-interference action are introduced during training.

[0156] According to the Construct a candidate operating mode set from the predicted probability distribution of the operating modes of each radar node. :

[0157] (twenty three)

[0158] in, This indicates the preset threshold for candidate working modes.

[0159] In practical applications, the Top-K method can also be used to select several working modes with high prediction probabilities to construct a candidate working mode set.

[0160] set up Indicates working mode The set of anti-interference actions that are allowed or more likely to occur is the set of candidate anti-interference actions induced by the set of candidate working modes. for:

[0161] (twenty four)

[0162] Based on this, the consistency constraint term Defined as:

[0163] (25)

[0164] This constraint penalizes the probability quality outside the candidate anti-interference action set, but does not set it to zero directly, thus maintaining the relative independence of the two prediction branches while keeping the working mode and anti-interference action structure in harmony.

[0165] The working mode prediction loss in this embodiment Defined as:

[0166] (26)

[0167] Anti-interference action prediction loss Defined as:

[0168] (27)

[0169] in, and The first The operating mode label and anti-jamming action label of each radar node in the next moment.

[0170] KL regularization term between variational posterior distribution and conditional prior distribution Defined as:

[0171] (28)

[0172] When both the variational posterior distribution and the conditional prior distribution adopt a diagonal Gaussian distribution, the KL regularization term can be analytically calculated as follows:

[0173] (29)

[0174] in, This represents the index of the potential state dimension.

[0175] Thus, the joint loss function described in this embodiment for:

[0176] (30)

[0177] in, , , These represent the weights of the corresponding loss terms.

[0178] During training, annealing weights are applied to the KL regularization term, gradually increasing its weight from its initial value to a preset upper limit. During round training, the weights of the KL regularization term can be expressed as:

[0179] (31)

[0180] in, This indicates that the KL weight has a preset upper limit. Indicates the number of preheating cycles.

[0181] This embodiment uses the KL annealing strategy to reduce the risk that the latent state representations will be ignored by the prediction network in the early stages of training.

[0182] In some embodiments, after acquiring new trajectory segments during online adversarial processes, the sample pool is updated, and a small batch of historical observation samples is constructed from the updated sample pool for incremental training to adapt to behavioral distribution drift caused by changes in jamming patterns, mission phases, or radar response patterns.

[0183] Let the first The sample pool before the next online update was The newly obtained set of trajectory segments is The sample pool update process can be represented as follows:

[0184] (32)

[0185] in, This indicates a sample pool update operation, which may include direct merging, sliding retention, priority retention of low-frequency samples, or retention in chronological order.

[0186] From the updated sample pool Constructing small batches of historical observation samples The radar behavior prediction model parameters are then incrementally optimized based on the joint loss function.

[0187] (33)

[0188] in, Indicates the first Model parameters before the last online update Indicates the first The model parameters after the second online update This represents the learning rate.

[0189] In addition, the present invention also conducted the following experiments:

[0190] To verify the effectiveness of the method of this invention, this embodiment constructs a discrete-time dynamic electronic warfare simulation environment consisting of three radars and three jammers. The three radars form a cooperative radar network, and the three jammers form a cooperative jammer swarm. At each time step, each jammer selects a jamming pattern from a preset jamming pattern set and a jamming target from a radar node set, forming the external jamming behavior at that moment. The jamming behavior further affects the jammed state, operating mode transition, and anti-jamming action selection of each radar node.

[0191] In this simulation environment, when target observation quality is good and the level of disturbance is low, radar nodes tend to transition from search or acquisition states to tracking or track-holding states. However, when external interference intensifies or deceptive interference persists, radar nodes may experience state-holding failures or operational mode degradation. Simultaneously, nodes in a higher detection state within the radar network can provide cooperative assistance to nodes in lower states, thus influencing the probability of lower-state nodes entering higher states such as acquisition or tracking. Regarding anti-jamming action generation, the environment determines candidate anti-jamming actions based on the dominant interference pattern, current operational mode, and historical disturbance states, and generates the final action probabilistically, ensuring that the generated trajectory simultaneously possesses state-holding, behavior transition, and action switching characteristics.

[0192] In this experiment, the historical observation window length was set to 6 steps, the number of training rounds was set to 200, the maximum number of steps was set to 100, the batch size was set to 64, the number of test trajectories was set to 80, and the learning rate was set to... The update mechanism employs online trajectory sample updates and incremental model training.

[0193] During the evaluation process, if the operating mode of a radar node changes between two adjacent time steps, the sample is recorded as an operating mode transition sample; if the anti-jamming action category of a radar node changes between two adjacent time steps, the sample is recorded as an anti-jamming action transition sample. Evaluation metrics include overall operating mode prediction accuracy, overall anti-jamming action prediction accuracy, operating mode transition sample prediction accuracy, and anti-jamming action transition sample prediction accuracy. A confusion matrix is ​​used to analyze prediction performance under conditions of uneven category distribution.

[0194] Experimental results are as follows Figures 4 to 7 As shown. Figure 4 The confusion matrix for radar operating mode prediction, such as Figure 4 The experiments shown demonstrate that the model can effectively distinguish between the main operational states, such as capture and tracking. The prediction accuracy for category 0 is 52.9%, with misjudgments mainly concentrated in category 1. Category 0 corresponds to the search state, which often appears in the initial stage of the simulation trajectory. Due to insufficient effective contextual information within the fixed-length historical window, the transition boundary between the search state and subsequent capture states is difficult to accurately determine.

[0195] Figure 5 The confusion matrix for predicting radar anti-jamming actions, such as Figure 5 The experiments show that the model has a good ability to identify the main anti-interference actions. The misclassification is mainly concentrated between categories 1 and 2, 3 and 4, and 6 and 7, indicating that some actions have similar historical observation characteristics during the switching phase.

[0196] Figure 6 The prediction accuracy of radar operating modes and anti-jamming actions on the overall sample is shown as a function of the number of training epochs, such as... Figure 6 As shown, with the increase of the number of training rounds, the prediction accuracy of both types of predictions showed a significant upward trend, indicating that the model has formed a relatively stable overall behavior prediction capability.

[0197] Figure 7 The prediction accuracy of radar operating mode transition samples and anti-jamming action switching samples is shown to change with the number of training rounds, such as... Figure 7 As shown, the prediction accuracy of both types of transition samples improves rapidly in the early stages of training and stabilizes at around 0.8 after about 100 rounds, indicating that the model can identify behavioral changes using historical observation sequences. Compared with the overall sample, the transition sample curves show more obvious local fluctuations, reflecting stronger uncertainty in the boundaries of state transitions and action switching. However, the model still maintains relatively stable prediction performance in the later stages of training.

[0198] in, Figure 6 and Figure 7The lighter-colored curves represent the original curves, while the darker-colored curves represent the trend curves after the original curves have undergone moving average processing, in order to more clearly show the overall trend of the original curves.

[0199] Overall, the confusion matrix results indicate that the method of this invention still possesses relatively stable classification capabilities under class imbalance conditions. For some classes with similar state boundaries or response mechanisms, a small amount of cross-prediction may still occur, but the overall confusion level is low. The method of this invention can not only learn the state-preserving patterns of radar network behavior but also capture dynamic behavioral changes such as operating mode switching and anti-interference action adjustments, and has good class discrimination capabilities.

[0200] Based on the above embodiments of the present invention, it can be seen that the online prediction method for observable behavior of radar networks of the present invention has at least the following advantages:

[0201] (1) The embodiments of the present invention are geared towards the external incomplete observation conditions in dynamic electronic countermeasures environments. Without directly obtaining the internal decision-making basis, physical situation or cooperative intention of the radar network, the observable behavior of the radar network at the next moment can be accurately predicted, which improves the applicability in non-cooperative observation scenarios.

[0202] (2) The embodiments of the present invention simultaneously model the behavioral evolution characteristics of a single radar node and the spatial collaborative correlation between different radar nodes by using a spatiotemporal graph encoder, thereby enhancing the ability to express the changes in the collaborative behavior of multiple nodes.

[0203] (3) The embodiments of the present invention introduce conditional variational latent states, which can model potential behavioral factors and behavioral uncertainties that are difficult to describe explicitly under some observable conditions, thereby improving the predictive ability of working mode transitions and anti-interference action switching.

[0204] (4) The present invention uses a working mode prediction branch and an anti-interference action prediction branch for joint prediction, and introduces consistency constraints on working mode and anti-interference action to maintain structural coordination between the two at the probability distribution level and avoid error cascading caused by hard constraints.

[0205] (5) The present invention introduces an online sample pool update and incremental training mechanism, which enables the prediction model to continuously absorb new trajectory segments, alleviate the behavior distribution drift problem in the dynamic adversarial process, and improve the model's continuous adaptability.

[0206] In summary, this invention, through the technical approach of acquiring externally observable records, constructing historical trajectory fragments, spatiotemporal collaborative representation, conditional variational latent state modeling, node-level joint prediction, and online incremental updates, realizes online prediction of the observable behavior of radar networks in the next moment under dynamic electronic warfare environments, and can provide technical support for understanding radar network behavior, assessing the confrontation situation, and predicting strategies.

[0207] Figure 8 This is a schematic diagram of an electronic device illustrated in this specification according to an exemplary embodiment. Please refer to... Figure 8 At the hardware level, the device includes a processor 810, an internal bus 820, a network interface 830, memory 840, a hardware acceleration device 850, and non-volatile memory 860, and may also include other hardware required for its functions. One or more embodiments of the present invention can be implemented in software, for example, the processor 810 reads the corresponding computer program from the non-volatile memory 860 into the memory 840 and then runs it. Of course, in addition to software implementation, one or more embodiments of the present invention do not exclude other implementation methods, such as logic devices or a combination of hardware and software, etc. That is to say, the execution subject of the above processing flow is not limited to each logic unit, but can also be hardware or logic devices.

[0208] Figure 9 This is a structural block diagram of an online prediction device for the observable behavior of a radar network in a dynamic electronic countermeasures environment, as illustrated in an exemplary embodiment of the present invention. The online prediction device can be applied to applications such as... Figure 8 The electronic device shown implements the technical solution of the present invention. The online prediction device includes: a data acquisition unit 910, a data processing unit 920, and a behavior prediction unit 930, wherein:

[0209] The data acquisition unit 910 is used to acquire the observation records of each radar node in the radar network and each jamming node in the jamming group during the dynamic electronic countermeasures process. The observation records include at least the working mode and anti-jamming actions exhibited by each radar node, as well as the jamming patterns and jamming targets exhibited by each jamming node.

[0210] The data processing unit 920 is used to construct a discrete observation sequence based on the historical observation records of each radar node and perform feature embedding to obtain a continuous feature sequence.

[0211] The behavior prediction unit 930 is used to input the continuous feature sequences of each radar node into the trained radar behavior prediction model to obtain the probability distribution of the working mode and the probability distribution of anti-jamming actions of each radar node in the next moment. The radar behavior prediction model includes a spatiotemporal graph encoder, a conditional prior network, a posterior network, and a dual-head prediction network.

[0212] The spatiotemporal graph encoder is used to encode the continuous feature sequences of each radar node in the time dimension to obtain behavioral evolution features. Then, it aggregates each behavioral evolution feature in the spatial dimension according to the attention weight between nodes and fuses it to obtain the spatiotemporal features of each radar node. Finally, it aggregates all spatiotemporal features to obtain the global context features of the radar network.

[0213] The conditional prior network is used to map the global context features, output the distribution parameters of the conditional prior distribution, and obtain the conditional variational latent state features based on the distribution parameters.

[0214] The posterior network is used to output the distribution parameters of the variational posterior distribution during the training phase, taking global context features and the observation label of the next time step as inputs; during training, the conditional prior distribution approximates the variational posterior distribution through KL divergence constraints.

[0215] The dual-head prediction network is used to fuse the conditional variational latent state features with the spatiotemporal features of each radar node to obtain spatiotemporal enhancement features, and based on these features, predict the probability distribution of the working mode and the probability distribution of anti-jamming actions of each radar node at the next moment.

[0216] In some embodiments, the apparatus further includes a model training unit for acquiring training samples, the training samples including continuous feature sequences of each radar node and observation labels for the next time step; inputting the training samples into the radar behavior prediction model, calculating a joint loss function, the joint loss function including operating mode prediction loss, anti-jamming action prediction loss, KL regularization term, and consistency constraint term; optimizing the network parameters of the radar behavior prediction model based on the joint loss function; wherein, the operating mode prediction loss is determined based on the cross-entropy between the operating mode probability distribution and the actual operating mode in the observation labels for the next time step; the anti-jamming action prediction loss is determined based on the cross-entropy between the anti-jamming action probability distribution and the actual anti-jamming action in the observation labels for the next time step; the KL regularization term is the KL divergence between the variational posterior distribution and the conditional prior distribution; the consistency constraint term, based on the operating mode probability distribution, penalizes the probability of action categories in the anti-jamming action probability distribution of each radar node that do not correspond to the operating mode.

[0217] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of the present invention according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0218] Accordingly, the present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method described in any of the above embodiments.

[0219] Accordingly, embodiments of the present invention also provide a computer program product configured to perform the methods described in any of the above embodiments.

[0220] The systems, devices, modules, or units described in the above embodiments can be implemented by computer chips or entities, or by products with certain functions. A typical implementation device is a computer, which can take the form of a personal computer, laptop computer, cellular phone, camera phone, smartphone, personal digital assistant, media player, navigation device, email sending and receiving device, game console, tablet computer, wearable device, or any combination of these devices.

[0221] In a typical configuration, a computer includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0222] Memory may include non-persistent storage in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM. Memory is an example of computer-readable media.

[0223] Computer-readable media, including both permanent and non-permanent, removable and non-removable media, can store information using any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random-access memory (SRAM), dynamic random-access memory (DRAM), other types of random-access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, disk storage, quantum memory, graphene-based storage media or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.

[0224] While this specification contains numerous specific implementation details, these should not be construed as limiting the scope of any invention or the scope of the claims, but rather are primarily intended to describe features of specific embodiments of a particular invention. Certain features described in the various embodiments herein may also be implemented in combination in a single embodiment. Conversely, various features described in a single embodiment may also be implemented separately in various embodiments or in any suitable sub-combination. Furthermore, while features may function in certain combinations as described above and even initially claimed in this way, one or more features from a claimed combination may be removed from that combination in some cases, and a claimed combination may refer to a sub-combination or a variation thereof.

[0225] Similarly, although the operations are depicted in a specific order in the accompanying drawings, this should not be construed as requiring these operations to be performed in the specific order shown or sequentially, or requiring all illustrated operations to be performed to achieve the desired result. In some cases, multitasking and parallel processing may be advantageous. Furthermore, the separation of various system modules and components in the above embodiments should not be construed as requiring such separation in all embodiments, and it should be understood that the described program components and systems can generally be integrated together in a single software product or packaged into multiple software products.

[0226] Thus, specific embodiments of the subject matter have been described. Other embodiments are within the scope of the appended claims. In some cases, the actions recited in the claims may be performed in a different order and still achieve the desired result. Furthermore, the processes depicted in the drawings are not necessarily shown in a specific order or sequence to achieve the desired result. In some implementations, multitasking and parallel processing may be advantageous.

[0227] It should be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0228] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for online prediction of the observable behavior of a radar network under dynamic electronic countermeasures environment, characterized in that, Includes the following steps: Acquire observation records of each radar node in the radar network and each jamming node in the jamming group during dynamic electronic countermeasures processes; For each radar node, a discrete observation sequence is constructed based on its historical observation records and feature embedding is performed to obtain a continuous feature sequence. The continuous feature sequences of each radar node are input into the trained radar behavior prediction model to obtain the probability distribution of the operating mode and the probability distribution of anti-jamming actions of each radar node in the next moment; the radar behavior prediction model includes: The spatiotemporal graph encoder is used to encode the continuous feature sequences of each radar node in the time dimension to obtain behavioral evolution features. Then, based on the attention weights between nodes, the behavioral evolution features are aggregated in the spatial dimension and fused with them to obtain the spatiotemporal features of each radar node. Finally, all spatiotemporal features are aggregated to obtain the global context features of the radar network. The conditional prior network is used to map the global context features, output the distribution parameters of the conditional prior distribution, and obtain the conditional variational latent state features based on the distribution parameters. The posterior network is used to output the distribution parameters of the variational posterior distribution during the training phase, taking global context features and the observation label of the next time step as inputs; during training, the conditional prior distribution is approximated by the variational posterior distribution through KL divergence constraints. The dual-head prediction network is used to fuse the conditional variational latent state features with the spatiotemporal features of each radar node to obtain spatiotemporal enhancement features, and predict the probability distribution of the operating mode and the probability distribution of anti-jamming actions of each radar node at the next moment based on these features; the dual-head prediction network includes a shared feature fusion module, an operating mode prediction branch and an anti-jamming action prediction branch; The shared feature fusion module adopts a multilayer perceptron structure, including a sequentially connected feature stitching layer, a first fully connected layer, a nonlinear activation layer, a normalization layer, a Dropout layer, and a second fully connected layer. The shared feature fusion module is used to stitch and nonlinearly map the spatiotemporal features of each radar node with the conditional variational latent state features to generate spatiotemporal enhancement features for each radar node. The operating mode prediction branch includes a fully connected layer, a nonlinear activation layer and a Softmax output layer connected in sequence. The operating mode prediction branch is used to output the operating mode probability distribution of each radar node at the next moment according to the spatiotemporal enhancement features. The anti-jamming action prediction branch includes a sequentially connected connection layer, a nonlinear activation layer, and a Softmax output layer. The anti-jamming action prediction branch is used to output the probability distribution of anti-jamming actions of each radar node at the next moment based on the spatiotemporal enhancement features.

2. The method according to claim 1, characterized in that, The attention weights between nodes employ multi-head attention, and the spatiotemporal graph encoder is configured as follows: For each attention head of each radar node, the behavioral evolution features of that radar node are mapped into query vectors and key vectors, and the behavioral evolution features of other radar nodes are mapped into value vectors. Based on the query vectors, key vectors, and value vectors, the attention weights of that radar node to other radar nodes are calculated. The value vectors are then weighted and aggregated using the attention weights to obtain the spatial collaborative features of that radar node. The behavioral evolution characteristics and spatial coordination characteristics of each radar node are spliced ​​together, and then linearly mapped and nonlinearly activated to obtain the spatiotemporal characteristics of each radar node.

3. The method according to claim 1, characterized in that, The posterior network includes: The input concatenation layer is used to concatenate the global context features of the training samples with the embedded features of the observation labels at the next time step to obtain the concatenated features; A linear mapping layer is used to perform a linear transformation on the spliced ​​features to extract high-dimensional features related to the potential behavior of radar nodes. A nonlinear activation layer is used to perform nonlinear activation on the high-dimensional features; The parameter output layer, connected after the nonlinear activation layer, includes a mean output header and a variance output header. The mean output header is used to output the mean vector of the variational posterior distribution, and the variance output header is used to output the standard deviation vector of the variational posterior distribution.

4. The method according to claim 1, characterized in that, The conditional prior network includes: The input layer is used to receive the global context features; A linear mapping layer is used to perform linear transformations on the global context features and extract high-dimensional features related to subsequent changes in the radar network behavior. A nonlinear activation layer is used to perform nonlinear activation on the high-dimensional features; The parameter output layer, connected after the nonlinear activation layer, includes a mean output header and a variance output header. The mean output header is used to output the mean vector of the conditional prior distribution, and the variance output header is used to output the standard deviation vector of the conditional prior distribution.

5. The method according to claim 1, characterized in that, The distribution parameters include the mean vector and standard deviation vector of the conditional prior distribution, and the conditional prior network is configured as follows: The conditional variational latent state features are obtained by sampling from the conditional prior distribution, or by directly using the mean vector of the conditional prior distribution as the conditional variational latent state features.

6. The method according to any one of claims 1 to 5, characterized in that, The training process of the radar behavior prediction model includes: Acquire training samples, which include continuous feature sequences of each radar node and observation labels for the next time step; The training samples are input into the radar behavior prediction model to calculate the joint loss function, which includes the working mode prediction loss, anti-jamming action prediction loss, KL regularization term and consistency constraint term. The network parameters of the radar behavior prediction model are optimized based on the joint loss function. The predicted loss for the working mode is determined based on the cross-entropy between the probability distribution of the working mode and the actual working mode in the observed label at the next time step. The anti-interference action prediction loss is determined based on the cross-entropy between the anti-interference action probability distribution and the actual anti-interference action in the observation label at the next moment. The KL regularization term is the KL divergence between the variational posterior distribution and the conditional prior distribution; The consistency constraint term, based on the working mode probability distribution, penalizes the probability of action categories in the anti-jamming action probability distribution of each radar node that do not correspond to the working mode.

7. An online prediction device for the observable behavior of a radar network under dynamic electronic countermeasures environment, characterized in that, The apparatus is used to perform the method as described in claim 1, the apparatus comprising: The data acquisition unit is used to acquire the observation records of each radar node in the radar network and each jamming node in the jamming group during the dynamic electronic countermeasures process. The data processing unit is used to construct discrete observation sequences for each radar node based on its historical observation records and perform feature embedding to obtain continuous feature sequences. A behavior prediction unit is used to input the continuous feature sequences of each radar node into the trained radar behavior prediction model to obtain the probability distribution of the operating mode and the probability distribution of anti-jamming actions of each radar node in the next moment; the radar behavior prediction model includes: The spatiotemporal graph encoder is used to encode the continuous feature sequences of each radar node in the time dimension to obtain behavioral evolution features. Then, based on the attention weights between nodes, the behavioral evolution features are aggregated in the spatial dimension and fused with them to obtain the spatiotemporal features of each radar node. Finally, all spatiotemporal features are aggregated to obtain the global context features of the radar network. The conditional prior network is used to map the global context features, output the distribution parameters of the conditional prior distribution, and obtain the conditional variational latent state features based on the distribution parameters. The posterior network is used to output the distribution parameters of the variational posterior distribution during the training phase, taking global context features and the observation label of the next time step as inputs; during training, the conditional prior distribution is approximated by the variational posterior distribution through KL divergence constraints. The dual-head prediction network is used to fuse the conditional variational latent state features with the spatiotemporal features of each radar node to obtain spatiotemporal enhancement features, and based on these features, predict the probability distribution of the working mode and the probability distribution of anti-jamming actions of each radar node at the next moment.

8. An electronic device, characterized in that, include: processor; as well as A computer-readable storage medium storing computer program instructions that, when executed by the processor, cause the processor to perform the method as described in any one of claims 1 to 6.

Images