A task graph threat update method and system based on time series convolution

Abstract

The application discloses a task graph threat updating method and system based on time sequence convolution, and belongs to the technical field of intelligent decision and time sequence prediction. The application takes a target node in a task graph as an object, obtains a target portrait, a historical dynamic state sequence, behavior stage features and task constraint information to form a prediction input; different time scale threat evolution features are extracted through a multi-scale time sequence convolution network, and portrait condition modulation and stage gate fusion are sequentially performed; credibility fusion is performed in combination with mechanism reference information to generate a threat curve, a peak threat, a peak arrival time and a high threat state in a future prediction window, and finally the task graph target node is written back. The application realizes conversion from static evaluation to dynamic prediction, enables the task graph to have a dynamically updatable threat attribute, and provides real-time decision support for task planning and resource scheduling.

Description

Technical Field

[0001] This invention belongs to the field of intelligent decision-making and time series prediction technology, specifically involving a task graph threat update method and system based on temporal convolution, which is particularly suitable for target threat evolution prediction, task graph node attribute update, and subsequent task chain generation and resource scheduling decision-making in dynamic adversarial environments. Background Technology

[0002] In dynamic adversarial environments, the threat status of mission targets typically evolves continuously with changes in time, spatial location, maneuvering behavior, firepower status, electronic countermeasures capabilities, and collaborative relationships. When generating plans, scheduling resources, or organizing mission chains, mission systems usually need to determine target priority, resource response timing, and mission execution windows based on the threat level of the targets. Therefore, accurately characterizing the dynamic trends of target threats and promptly translating the prediction results into computable node attributes in the mission graph is a crucial factor affecting the effectiveness of subsequent mission planning and resource allocation.

[0003] Existing threat assessment methods mostly employ rule-based models, index-weighted models, expert experience models, or static scoring models to calculate the target's current threat level based on information such as target type, target value, current distance, firepower capability, and mobility. While these methods are simple to implement and easy to interpret, they typically focus on the current state or static attributes, making it difficult to depict the trend of target threat changes over time or to identify potential high-threat windows in advance. When targets are in a state of rapid maneuver, covert approach, phase transition, or coordinated operation, relying solely on the current threat value for judgment can easily lead to delayed target priority updates, deviations in resource response timing, and insufficient task window matching.

[0004] With the development of deep learning and time series forecasting techniques, recurrent neural networks, long short-term memory networks, gated recurrent units, temporal convolutional networks, and Transformer-like models have been used for dynamic sequence forecasting tasks. These methods can learn temporal change patterns from historical dynamic state sequences and predict future states. Among them, temporal convolutional networks have strong parallel computing capabilities, scalable receptive fields, and stable long-sequence modeling, making them suitable for processing time series data such as target states, situational changes, and threat evolution. However, existing general-purpose time series forecasting models typically output a single future numerical sequence, focusing more on the prediction error itself, and struggling to simultaneously output summary attributes needed for task-oriented decision-making, such as peak threat, peak arrival time, and high-threat states.

[0005] Furthermore, in task planning and resource scheduling scenarios, objectives, resources, task roles, constraints, and collaborative relationships typically exhibit multi-type, multi-relationship, and dynamically changing characteristics. Heterogeneous task graphs can provide a unified representation of these elements, enabling task objectives, available resources, task roles, constraint nodes, and their relationships to form a computable structure. However, existing graph modeling methods often focus on entity relationship organization, path search, or static attribute representation, lacking a dynamic update mechanism to write time series prediction results back to task graph nodes in real time. In other words, the prediction model and the task graph structure are often disconnected, making it difficult for prediction results to directly participate in subsequent objective ranking, candidate chain scoring, time window matching, and resource scheduling.

[0006] On the other hand, the evolution of target threats is not only influenced by historical dynamic state sequences, but also by the target's own capability profile, current behavioral stage, and mechanistic constraints within the task scenario. Existing prediction methods typically use simple feature concatenation to fuse target attributes and historical dynamic states, which fails to fully express the differences in threat responses of different types of targets under similar motion states. Furthermore, the impact of short-term mutations, mid-term stage transitions, and long-term trends on future threats varies across different behavioral stages, and existing models often struggle to adaptively adjust the contribution of features at different time scales according to the target stage. Moreover, prediction models that rely entirely on data-driven approaches are prone to prediction fluctuations when there are insufficient samples, sudden state changes, or uncertain inputs, affecting the stability of subsequent task decisions.

[0007] Therefore, existing technologies still have the following shortcomings: First, static threat assessment methods are difficult to predict future threat change trends and high-threat windows; second, general time series prediction models are difficult to directly output multiple threat attributes applicable to task decision-making; third, there is a lack of effective connection between prediction results and task graph node attributes, making it difficult to support dynamic updates of the task graph; and fourth, existing models do not adequately integrate target profiles, behavioral stages, and mechanistic constraints, resulting in limited applicability and stability of threat prediction results in complex dynamic scenarios. Summary of the Invention

[0008] The purpose of this invention is to provide a task graph threat update method and system based on temporal convolution, addressing the problem that existing threat assessment methods mainly rely on current states or static indicators, making it difficult to characterize the future threat evolution trend of targets. This invention introduces the target's historical dynamic state sequence, target profile, and behavioral stage features. It utilizes a multi-scale temporal convolutional network to extract threat change features at different time scales, and combines profile conditional modulation, stage-gated fusion, and mechanistic reference information to achieve dynamic prediction of threat curves within the future prediction window.

[0009] This invention provides a task graph threat update method and system based on temporal convolution. Taking the target node in the task graph as the update object, it comprehensively utilizes the target's historical dynamic state sequence, target profile, target behavior stage characteristics, and mechanism reference information. It predicts the target's future threat curve through a multi-scale temporal convolutional network and further generates dynamic threat decision attributes such as peak threat, peak arrival time, and high threat status. Finally, the above prediction results are written back to the target node in the task graph, enabling the task graph to have the ability to update target node attributes in a dynamic environment.

[0010] This invention provides a task graph threat update method based on temporal convolution, comprising the following steps: Construct a task graph, which includes a set of nodes, a set of relationships between nodes, a set of node attributes, and a set of relationship attributes; The node set includes target nodes, resource nodes, task role nodes, and constraint nodes; The node attribute set includes target profile, historical dynamic state sequence, behavioral stage characteristics, and task constraint information; Select any target node and perform the following target node attribute update steps: Obtain the node attributes of the target node to form the threat prediction input; and extract the threat evolution characteristics of the target node; Based on the target profile and behavioral stage features, the threat evolution features are sequentially subjected to conditional modulation and stage-gated fusion to obtain the fused threat evolution features; Based on the fused threat evolution characteristics, a threat curve for the target node within a future prediction window is generated, and based on the threat curve within the future prediction window, peak threat, peak arrival time, and high threat status are generated. The threat curve, peak threat, peak arrival time, and high threat status within the future prediction window are written back to the corresponding target node in the task graph, thus completing the target node attribute update.

[0011] Furthermore, the threat prediction input is represented as: ; in, Represents the target node Threat prediction input, Indicates the target node number. Represents the characteristic organization function, Represents the target node The target profile Represents the target node Historical dynamic state sequence Represents the target node Behavioral stage characteristics, Indicates the relationship with the target node Relevant task constraint information; The task map is a forest fire emergency response task map. The target nodes in the task map include forest fire points, suspected fire areas, fire spread areas, or important facility areas threatened by fire. The resource nodes in the mission diagram include reconnaissance drones, firefighting drones, fire trucks, command centers, and communication relay drones. The task role nodes include fire reconnaissance role, communication relay role, and fire fighting role; The constraints include drone endurance constraints, communication coverage constraints, no-fly zone constraints, and disposal time limit constraints; The target profile includes vegetation density, type of combustible material, terrain slope, regional importance, and distribution of surrounding facilities; The historical dynamic state sequence includes temperature, smoke concentration, flame intensity, fire spread rate, wind speed, and wind direction at continuous observation times; The behavioral phases include the initial discovery phase, the fire expansion phase, the stable and controlled phase, and the reignition detection phase. Task constraint information includes the estimated arrival time of resources, the scope matching result, the communication status, and the allowed processing time window.

[0012] Furthermore, a temporal convolutional threat prediction module is used to extract the threat evolution features of the target node. The temporal convolutional threat prediction module includes at least two temporal convolutional branches with different time scales; the threat evolution features are threat evolution features under multiple time scales. Each temporal convolutional branch has the same structure, including an input mapping layer and at least one temporal convolutional block; The input mapping layer is used for threat prediction input. The feature dimension is transformed and mapped to a hidden feature representation; the temporal convolutional block is used to extract time-series features from the hidden feature representation to obtain threat evolution features at a single time scale.

[0013] Furthermore, the threat evolution characteristics are conditionally modulated, and the specific steps are as follows: First, create a target profile. Mapped to image embedding: ; in, This represents the image mapping function. Indicates the target profile The mapped embedding vector; Regenerate scaling parameters and offset parameters , represented as: ; in, This represents the parameter generation function; The threat evolution features are modulated channel by channel to obtain the threat evolution features conditionally modulated by the target profiling. , represented as: ; in, This represents element-wise multiplication. Indicates the first Threat evolution characteristics over time scales.

[0014] Furthermore, the evolutionary characteristics of threats after conditional modulation. The phased gating fusion is performed as follows: First, the characteristics of the behavioral stages Mapped to stage embedding: ; in, Represents the stage mapping function, Representing stage characteristics The mapped embedding vector; No. The gating weights corresponding to each time scale are: ; in, Indicates the first The gating weights corresponding to each time scale This represents the gating weight generation function. The nth element in the output vector of the gated weight generation function The components corresponding to each time scale This represents the component corresponding to the j-th time scale in the output vector of the gated weight generation function. Indicates the maximum values ​​of k and j; And satisfy ; The fusion is represented as: ; in, This indicates the threat evolution characteristics after stage-gated fusion.

[0015] Furthermore, based on the fused threat evolution characteristics, a threat curve for the target node within the future prediction window is generated, specifically as follows: Suppose that the future prediction window includes H prediction steps, based on the target node. Corresponding mechanism reference information , generate the first Mechanistic reference threat value corresponding to each prediction step , represented as: ; in, The mechanism reference generating function, The value is an integer, ranging from 1 to H; Generate target node In the The mechanism reliability coefficient corresponding to each prediction step , represented as: ; in, This represents the Sigmoid mapping function, used to map the confidence generation function. The output is mapped to the mechanism credibility coefficient. This indicates the threat evolution characteristics after stage-gated fusion. Indicates the first The prediction step embedding corresponding to each prediction step. This represents the credibility generation function. ; The threat decoding function generates the first Data-driven initial threat prediction values ​​for each prediction step , represented as: ; in, This represents the threat decoding function; No. Final threat prediction value for each prediction step , represented as: ; The final threat prediction values ​​of each prediction step constitute the threat curve of the target node within the future prediction window, as shown below: .

[0016] Further mechanistic reference information This includes at least one of the following: distance variation, scope of action, task rules, and stage constraints.

[0017] Furthermore, after writing the threat curve within the future prediction window, the peak threat, the peak arrival time, and the high-threat status back to the corresponding target node in the task graph, the updated target node attribute set is obtained. Represented as: ; in, Represents the target node initial properties, , Represents the target node The target profile Represents the target node Historical dynamic state sequence Represents the target node Behavioral stage characteristics, Indicates the relationship with the target node Relevant task constraint information, Indicates the target node number. This represents finding the union of sets; Represents the target node Dynamic threat decision-making attributes, writing, , Represents the target node Threat curve within the future forecast window, Indicates peak threat. Indicates the peak arrival time. This indicates a high-threat status.

[0018] Furthermore, the task graph contains multiple target nodes, and the target node attributes are updated sequentially for each target node until the target node attributes of all target nodes are updated.

[0019] This invention also provides a task graph threat update system based on temporal convolution, comprising: The task graph construction module is used to construct a task graph that includes a set of nodes, a set of relationships between nodes, a set of node attributes, and a set of relationship attributes. The set of nodes includes target nodes, resource nodes, task role nodes, and constraint nodes. The set of node attributes includes target profiles, historical dynamic state sequences, behavioral stage characteristics, and task constraint information. The module also determines the target nodes to be updated. The prediction input generation module is used to obtain the target profile, historical dynamic state sequence, behavioral stage characteristics and task constraint information of the target node to be updated, and form the threat prediction input; The temporal convolutional threat prediction module is used to extract the threat evolution features of the target node to be updated at various time scales through a temporal convolutional model. The feature fusion module is used to conditionally modulate the threat evolution features at each time scale based on the target profile, and to perform stage-gated fusion of the conditionally modulated threat evolution features based on the behavioral stage features to obtain the fused threat evolution features. The mechanism fusion module is used to generate a data-driven initial threat prediction value corresponding to the target node to be updated based on the fused threat evolution characteristics; generate a mechanism reference threat value and a mechanism confidence coefficient based on the mechanism reference information corresponding to the target node to be updated; and then fuse the data-driven initial threat prediction value and the mechanism reference threat value based on the mechanism confidence coefficient to obtain the final threat prediction value. The threat attribute generation module is used to generate a threat curve for the target node to be updated within a future prediction window based on the final threat prediction value, and to generate peak threat, peak arrival time and high threat status based on the threat curve within the future prediction window. The task graph update module is used to write the threat curve, the peak threat, the peak arrival time, and the high threat status within the future prediction window back to the target node to be updated in the task graph.

[0020] This invention integrates threat prediction and task graph node updating methods with historical dynamic states of targets, target profiles, behavioral stage characteristics, and mechanistic reference information. It not only obtains threat curves within future prediction windows but also generates dynamic threat decision attributes such as peak threat, peak arrival time, and high-threat status. These dynamic threat decision attributes are then written back to the target nodes in the task graph. This invention transforms target threat assessment from static evaluation to dynamic prediction and updating, enabling target nodes in the task graph to possess threat attributes that update with changing situations. Furthermore, by integrating target profiles, behavioral stages, and mechanistic reference information, it improves the adaptability of threat prediction results to different target types, different stage states, and complex dynamic environments, thus providing a dynamic data foundation for subsequent target sequencing, task chain generation, time window matching, and resource scheduling. It solves the problem that general time series prediction results are difficult to directly serve task planning.

[0021] Compared with the prior art, the present invention has the following beneficial effects: First, this invention uses the target's historical dynamic state, target profile, behavioral stage characteristics, and task constraint information as the basis for threat prediction, which can be extended from a single current threat assessment to a dynamic threat update oriented towards the future prediction window, thereby improving the ability to characterize the evolution trend of target threats.

[0022] Second, the present invention uses a temporal convolutional threat prediction module to extract target threat evolution features, which can simultaneously take into account short-term state changes, stage transition processes and long-term trend changes. Compared with static scoring or single-point prediction methods, it is more suitable for continuous threat judgment in dynamic environments.

[0023] Third, this invention enables the prediction process to consider the impact of different target types, different task stages and different scenario rules on threat changes through conditional modulation, stage feature fusion and mechanism reference correction, thereby improving the stability of prediction results under conditions of sudden state changes, insufficient samples or uncertain input.

[0024] Fourth, this invention not only outputs threat curves within the future prediction window, but also simultaneously generates dynamic threat decision attributes such as peak threat, peak arrival time, and high threat status, enabling the prediction results to directly serve target ranking, time window matching, and resource scheduling.

[0025] Fifth, this invention writes the generated dynamic threat decision attributes back to the target node of the task graph, transforming the task graph from a static descriptive structure into a dynamically updatable computational structure. This enables the temporal convolutional threat prediction module to connect with the subsequent task decision-making process, improving the real-time performance and applicability of task graph updates, candidate task chain scoring, and resource response decisions. Attached Figure Description

[0026] To more clearly illustrate the technical solution of the present invention, the accompanying drawings used in the embodiments of the present invention are briefly described below. It should be understood that the following drawings are only used to illustrate some embodiments of the present invention and do not constitute a limitation on the scope of protection of the present invention.

[0027] Figure 1 This is a flowchart of the overall process of a task graph threat update method based on temporal convolution according to the present invention.

[0028] Figure 2 This is a schematic diagram of the structure of a task graph threat update system based on temporal convolution according to the present invention.

[0029] Figure 3 This is a flowchart of the target node attribute update process in this invention. Detailed Implementation

[0030] The specific embodiments of the present invention will be further described below with reference to the accompanying drawings. It should be understood that the following embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit the scope of protection of the present invention. Equivalent substitutions, modifications, or improvements made to the present invention by those skilled in the art without departing from the concept of the present invention should all fall within the scope of protection of the present invention.

[0031] This invention provides a task graph threat update method and system based on temporal convolution. Its core is: updating objects with target node attributes in the task graph, integrating target profile, historical dynamic state sequence, behavioral stage features and task constraint information, predicting future threat changes of the target through a temporal convolution threat prediction module, generating dynamic threat decision attributes that can serve task decision-making, and then writing the dynamic threat decision attributes back to the target node in the task graph, so that the task graph has the ability to update target node attributes in a dynamic environment.

[0032] Example 1: A Task Graph Threat Update Method Based on Temporal Convolution like Figure 1 As shown, this embodiment provides a task graph threat update method based on temporal convolution, including the following steps: Step 1: Construct the task graph and determine the target node to be updated Building a task graph Represented as: ; in, Represents a set of nodes. Represents the set of relationships between nodes. This represents the set of node attributes and the set of relationship attributes.

[0033] Node set It should contain at least target nodes, resource nodes, task role nodes, and constraint nodes; target nodes are used to describe objects to be evaluated or processed; resource nodes are used to describe resource objects for executable tasks; task role nodes are used to describe reconnaissance, command, strike, support, or other task function roles; constraint nodes are used to describe constraints such as time, space, capability, connectivity, or task rules.

[0034] The set of relationships between nodes This includes the task requirement relationship between the target node and the task role node, the role adaptation relationship between the resource node and the task role node, the role or reachability relationship between the resource node and the target node, the communication or collaboration relationship between the resource nodes, the task dependency relationship between the task role nodes, and the constraint association relationship between the target node and the constraint node.

[0035] The node attribute set includes the target profile, historical dynamic state sequence, behavioral stage characteristics, and task constraint information related to the target node.

[0036] The set of relation attributes includes relation type, relation direction, fit, spatial distance, scope matching degree, reachability, communication quality, response time, time window, validity status, and confidence level. The corresponding relation attributes are configured according to the relation's function.

[0037] like Figure 3 As shown, select any target node in the task graph. The attribute is used as the update object, the target node The initial properties are represented as: ; in, Indicates the target node number. Represents the target node The target profile Represents the target node Historical dynamic state sequence Represents the target node Behavioral stage characteristics, Indicates the relationship with the target node Relevant task constraint information.

[0038] Step 2: Forming Threat Prediction Inputs for Target Nodes For the target node Obtain the target node Target profile Historical dynamic state sequence Characteristics of behavioral stages and task constraint information Forming threat prediction input , represented as: ; in, Represents the target node Threat prediction input, This represents a characteristic organization function.

[0039] Historical dynamic state sequence It can be composed of the target's motion state, distance state, velocity state, attitude state, task state, or situational state at consecutive moments. Target profile It can be composed of target capability, target category, target value, or target attributes. Characteristics of behavioral stages. It can be used to characterize a target's approach, penetration, entry into the effective range, withdrawal, or other preset phases. (Task constraint information) It can be used to characterize time windows, distance conditions, scope of action, rule conditions, or state confidence.

[0040] Through the above processing, the target profile is created. Historical dynamic state sequence Characteristics of behavioral stages and task constraint information The organization is the input format that the temporal convolutional threat prediction module can handle.

[0041] Step 3: Extract threat evolution features using the temporal convolutional threat prediction module. Input threat prediction Input the temporal convolutional threat prediction module to extract target nodes. Threat evolution characteristics , represented as: ; in, This represents the temporal convolutional threat prediction module. Represents the target node Threat evolution characteristics.

[0042] The temporal convolutional threat prediction module is used to extract temporal dependencies from historical dynamic state sequences. Unlike static evaluation methods that rely solely on the current state, this temporal convolutional threat prediction module can extract temporal dependencies based on threat prediction input. By utilizing continuous state changes within a historical observation window, the trend of target threats changing over time can be learned.

[0043] The temporal convolutional threat prediction module includes single-scale temporal convolutional branches or multi-scale temporal convolutional branches.

[0044] (1) When the time-series convolutional threat prediction module includes a single-scale temporal convolutional branch, the threat evolution features are features of a single time scale; the single-scale temporal convolutional branch adopts the basic structure of the existing temporal convolutional network TCN. Specifically, the single-scale temporal convolutional branch includes an input mapping layer and one or more temporal convolutional blocks; when multiple temporal convolutional blocks are included, each temporal convolutional block is connected in sequence, and the output of the input mapping layer is connected to the input of the first temporal convolutional block; the input mapping layer is used to process the threat prediction input. The feature dimension is transformed and mapped to a hidden feature representation suitable for temporal convolution processing; the temporal convolution block is used to extract time-series features from the hidden feature representation to obtain threat evolution features at a single time scale.

[0045] Temporal convolutional blocks can include causal one-dimensional convolution, dilated convolution, nonlinear activation, and residual connections, and can be configured with normalization layers and random deactivation layers as needed. By pre-setting the convolutional kernel size, dilation coefficient, and number of network layers, this branch forms a predetermined time scale and extracts the corresponding single-time-scale threat evolution features. The basic processing such as causal one-dimensional convolution, dilated convolution, and residual connections all adopt existing TCN technology.

[0046] Among them, the causal one-dimensional convolutional layer performs convolution calculation on the hidden feature representation, so that the output of the current time step depends only on the input information of the current time step and the previous time steps, thereby maintaining the temporal causal relationship; the dilated convolution processing unit expands the convolution scale by setting the dilation coefficient, and then obtains the threat evolution features under a single time scale after nonlinear activation and residual connection.

[0047] (2) When the time-series convolutional threat prediction module includes multi-scale temporal convolutional branches, the threat evolution features are multi-time-scale features. The multi-scale temporal convolutional branches include at least two temporal convolutional branches with different time scales. Different time scales can be formed by setting different convolutional kernel sizes, dilation coefficients, number of network layers, time sampling intervals, or combinations thereof.

[0048] In this embodiment, the multi-scale temporal convolution branch includes three temporal convolution branches with different time scales: a first temporal convolution branch (short-scale temporal convolution branch), a second temporal convolution branch (medium-scale temporal convolution branch), and a third temporal convolution branch (long-scale temporal convolution branch). The first-scale temporal convolution branch has a first time scale and is used to extract short-scale threat evolution features corresponding to recent state abrupt changes and rapid local changes in the target. The second-scale temporal convolution branch has a second time scale larger than the first time scale and is used to extract medium-scale threat evolution features corresponding to target behavior stage transitions and continuous state changes. The third-scale temporal convolution branch has a third time scale larger than the second time scale and is used to extract long-scale threat evolution features corresponding to the target's long-term movement trends and cross-stage evolution.

[0049] The first-scale, second-scale, and third-scale temporal convolutional branches have the same structure, each including an input mapping layer and at least one temporal convolutional block. The temporal convolutional block includes a causal one-dimensional convolutional layer and residual connections, and may include at least one of dilated convolutional layers, normalization layers, nonlinear activation layers, and random deactivation layers. The difference lies in the different branches employing different kernel sizes, dilation coefficients, the number of temporal convolutional blocks, or combinations thereof to create different time scales, thus affecting the prediction of the input. After passing through the first-scale temporal convolution branch, the second-scale temporal convolution branch, and the third-scale temporal convolution branch respectively, the threat evolution features at the first time scale are obtained sequentially. The threat evolution characteristics at the second timescale are: The characteristics of threat evolution at the third timescale are: , will the Threat evolution characteristics over time scales are represented as ,in, The time scale is indicated by a number; in this embodiment, k = 1, 2, 3. Among them, threat evolution characteristics Used to reflect recent abrupt changes in the target's state and the characteristics of threat evolution. Used to reflect changes in the target's behavioral stages, threat evolution characteristics are: It is used to reflect the continuous approach, continuous suppression, continuous withdrawal, or other trend of a target.

[0050] It should be noted that this invention does not limit the specific number of layers, kernel size, number of channels, or parameter configuration of the temporal convolutional threat prediction module. As long as it can be based on historical dynamic state sequences... Extracting target threat evolution features and using them to generate subsequent dynamic threat decision attributes are all within the scope of protection of this invention.

[0051] Step 4: Conditional modulation and stage-gated fusion of threat evolution characteristics Based on the target profile Characteristics of behavioral stages The multi-scale threat evolution characteristics obtained in step 3 By performing conditional modulation and stage-gated fusion of profiles, the evolutionary characteristics of the fused threat are obtained. ; Specifically, based on the target profile Threat evolution characteristics at various time scales Perform conditional modulation. First, create a target profile. Mapped to image embedding: ; in, This represents the image mapping function. Indicates the target profile The mapped embedding vector; Then based on the embedding vector Generation and Threat Evolution Characteristics Corresponding scaling parameters and offset parameters : ; in, This represents the parameter generation function; Threat evolution characteristics Channel-by-channel modulation was performed to obtain the threat evolution characteristics after conditional modulation. , represented as: ; in , This represents element-wise multiplication. Through the above conditional modulation, target nodes with similar dynamic states but different target profiles can obtain differentiated threat feature representations.

[0052] Furthermore, based on the characteristics of the behavioral stages Threat evolution characteristics after conditional modulation Perform phased gating fusion.

[0053] First, characteristics of behavioral stages Mapped to stage embedding: ; in, Represents the stage mapping function, Representing stage characteristics The mapped embedding vector; Based on the embedding vector Generation Threat Evolution Characteristics Corresponding gating weights , represented as: ; in, This represents the gated weight generation function, used to generate weights based on the embedding vector. Generate gating weights corresponding to different time scales. This indicates that the output vector of the gated weight generation function is related to the first... The components corresponding to each time scale This represents the component corresponding to the j-th time scale in the output vector of the gated weight generation function. This represents the maximum values ​​of k and j.

[0054] And it satisfies: ; Weighted fusion of modulated threat evolution features across different time scales: ; in, This represents the threat evolution characteristics after stage-gated fusion. Through stage-gated fusion, the contributions of features at different time scales are adaptively adjusted as the target's behavior changes through different stages.

[0055] Step 5: Generate data-driven initial threat prediction values, perform mechanism credibility fusion, and generate dynamic threat decision attributes. According to the task map In and target node By analyzing relevant node attributes, relationship attributes, and constraint attributes, mechanistic information related to the evolution of the target threat is extracted. This mechanistic information includes at least one of distance variation, scope of action, task rules, and stage constraints, forming a relationship between the target node and the target threat. Corresponding mechanism reference information ; Among them, the distance change in the mechanism information is obtained from the historical dynamic state sequence of the target node; the scope of action in the mechanism information is obtained based on the action or reachability relationship between the resource node and the target node (in the set of relationships between nodes), or based on the degree of matching of the scope of action or reachability relationship (in the set of relationship attributes). The task rules in the mechanism information are determined based on the task rules stored in the constraint nodes, the constraint relationship between the target node and the constraint nodes, and at least one of the corresponding time window, response time, or action condition. The stage constraints in the mechanistic information are determined based on the behavioral stage characteristics of the target node and the constraint nodes or constraint relationships associated with the behavioral stage. Suppose the future prediction window includes H prediction steps, based on mechanistic reference information. , generate the first Mechanistic reference threat value corresponding to each prediction step , is represented as: ; in, The mechanism reference generating function is represented.

[0056] Based on the characteristics of threat evolution after fusion Prediction step embedding Mechanism Reference Threat Value Generate target node In the The mechanism reliability coefficient corresponding to each prediction step , is represented as: ; in, This represents the Sigmoid mapping function. The first step is generated by the prediction step embedding function. The prediction step embedding corresponding to each prediction step. This represents the credibility generation function. Mechanism reference threat values ​​and mechanism confidence coefficients can be generated in parallel with data-driven initial threat prediction values.

[0057] The threat decoding process can be further combined with future prediction step information for step-size-aware decoding. This is based on the fused threat evolution characteristics. and prediction step embedding Generate the first threat decoding function Data-driven initial threat prediction values ​​for each prediction step , is represented as: ; in, This represents the threat decoding function.

[0058] Mechanism-based reference threat value and mechanism credibility coefficient Data-driven initial threat prediction values The data is then fused to obtain the final threat prediction value. , is represented as: ; in, Indicates the first The final threat prediction values ​​for each prediction step. The final threat prediction values ​​for each prediction step form the threat curve within the future prediction window: ; The threat curve within the future forecast window describes the trend of threat changes to a target over a future period. Based on the threat curve within the future forecast window... Further, dynamic threat decision attributes are generated: ; in, Indicates peak threat. Indicates the peak arrival time. Indicates a high-threat status. Represents the target node Dynamic threat decision-making attributes.

[0059] Among them, peak threat is used to characterize the highest level of threat that a target may reach within the future prediction window; peak arrival time is used to characterize the time and location at which the target reaches a higher threat state; and high threat state is used to determine whether the target has a high threat risk within the future prediction window. A high threat state is obtained by comparing the threat value at each moment in the threat curve within the future prediction window with a preset high threat threshold. A high threat state is defined as a situation where the threat value at any moment within the future prediction window is greater than or equal to the preset high threat threshold.

[0060] In an optional implementation, a high-threat window, threat growth trend, threat decay trend, threat duration, or other dynamic threat decision attributes can be further generated based on the threat curve within the future prediction window. These extended attributes can be set according to specific task requirements and do not affect the basic implementation of the present invention.

[0061] Step 6: Write the dynamic threat decision attributes back to the target node in the task graph. The dynamic threat decision attributes generated in step 5 Write back to the task diagram The corresponding target node Complete the dynamic update of target node attributes: ; After the update, the target node Not only includes target profile Historical dynamic state sequence Characteristics of behavioral stages and task constraint information It also includes threat curves within the future forecast window. Peak threat Peak arrival time and high threat status These dynamic threat decision-making attributes, This indicates finding the union of sets.

[0062] Through this step, the prediction results of the temporal convolutional threat prediction module are no longer presented as independent numerical outputs, but are transformed into a task graph. The target node attributes in the task graph. Therefore, the task graph... The descriptive structure, which only contains observed attributes or the current state, is transformed into a computational structure that can be dynamically updated as the environment changes.

[0063] The node attributes of the target node before the update in the task graph include the target profile, historical dynamic state sequence, behavioral stage characteristics, and task constraint information, all of which are observational attributes. These observational attributes can reflect the target's current or historical state, but they are difficult to express the target's threat change trend within the future prediction window.

[0064] Through the temporal convolutional threat prediction module of this invention, the target node obtains dynamic threat decision attributes. Dynamic threat decision attributes It should at least include the threat curve, peak threat, peak arrival time, and high-threat status within the future prediction window. Dynamic threat decision attributes should be included. After writing back to the target node, the target node's attributes are expanded, forming a dynamic node that includes both observed attributes and future prediction attributes.

[0065] In one implementation, the write-back process can employ overwrite updates, append updates, or versioned updates. Overwrite updates replace the target node's original dynamic threat decision attributes with the latest prediction results; append updates add new prediction results to the target node's attribute set to retain historical prediction records; and versioned updates generate corresponding attribute versions for different update times to support subsequent tracking, backtracking, or comparison.

[0066] In another implementation, the task graph can perform target node attribute updates based on preset time intervals, event triggering conditions, or state change thresholds. For example, target node attribute updates are triggered when the target state changes significantly; target node attribute updates are triggered when a task phase changes; and target node threat attribute updates are triggered when the target enters a preset area or meets preset rule conditions.

[0067] Through the aforementioned write-back mechanism, the task graph can continuously absorb the dynamic prediction results output by the temporal convolutional threat prediction module, transforming the target node from a static description node into a dynamic decision node, thereby providing real-time data support for subsequent task planning and resource scheduling.

[0068] The task graph threat update method of this invention can be executed periodically at fixed time intervals. For example, the attributes of target nodes in the task graph are updated every preset time interval to maintain the timeliness of dynamic threat decision attributes in the task graph.

[0069] In another alternative implementation, execution can be based on events. When the target state changes abruptly, the target enters a preset area, the behavior phase changes, the task constraints change, or the resource state changes, the corresponding target node attribute is updated.

[0070] In another alternative implementation, the method of the present invention can employ a combination of periodic updates and event triggering. For target nodes with stable states, updates are performed at a lower frequency; for target nodes with significant state changes or high predicted threats, the update frequency is increased. This ensures timely updates to dynamic threat decision attributes while reducing unnecessary computational overhead.

[0071] The updated task graph can be used for subsequent target priority ranking, candidate task chain scoring, time window matching, resource response timing selection, and task scheduling decisions. For example, target priority can be determined based on peak threat and target value in the target profile; resource response windows can be determined based on peak arrival time; whether resources need to be scheduled in advance can be determined based on high threat status; and candidate task chains can be dynamically scored based on threat curves within the future prediction window.

[0072] For example, in a multi-drone forest fire monitoring and emergency response application scenario, multiple drones with different functions and payload capabilities, ground fire-fighting resources, and a command center work together to perform tasks such as fire reconnaissance, fire confirmation, communication relay, fire extinguishing, and personnel evacuation guidance.

[0073] Target nodes in the task diagram are used to represent forest fire ignition points, suspected fire areas, fire spread areas, or important facility areas threatened by fire; this embodiment selects the first target node. As the target node to be updated.

[0074] The resource nodes in the mission diagram include reconnaissance drones, firefighting drones, fire trucks, command centers, and communication relay drones. The task role nodes include fire reconnaissance role, communication relay role, and fire fighting role; The constraints include drone endurance constraints, communication coverage constraints, no-fly zone constraints, and disposal time limits.

[0075] First target node The target profile includes vegetation density, combustible material type, terrain slope, regional importance, and distribution of surrounding facilities. In this embodiment, the normalized value of vegetation density is 0.82, the combustible material hazard level is 0.78, the terrain slope is 18°, the regional importance level is 0.90, and the first target node... There are important power facilities within a 500m radius.

[0076] The historical dynamic state sequence includes temperature, smoke concentration, flame intensity, fire spread rate, wind speed, and wind direction at continuous observation times.

[0077] The behavioral phase characteristics specifically include the initial discovery phase, the fire expansion phase, the stable and controlled phase, and the reignition detection phase.

[0078] For example, within six consecutive observation times before the prediction begins, the first target node The temperature in the corresponding area increased from 52℃ to 78℃, the normalized smoke concentration increased from 0.46 to 0.82, the normalized flame intensity increased from 0.40 to 0.76, the fire spread rate increased from 0.6 m / min to 1.4 m / min, the current wind speed is 6.5 m / s, and the wind direction is northeast. Based on the above changes in conditions, the first target node... The behavioral stage characteristics were identified as the fire expansion stage.

[0079] Task constraint information includes the estimated arrival time of the resource, the scope matching result, the communication status, and the allowed disposal time window, determined based on the interaction or reachability relationship between the resource node and the first target node and its relationship attributes. For example, a reconnaissance drone is already located at the first target node. The firefighting drone is expected to arrive in 2 minutes and its effective range will cover the first target node. The estimated arrival time of the fire truck is 4 minutes; the estimated arrival time of the communication relay drone is 2 minutes; the first target node The current communication quality normalization value with the command center is 0.68; the allowable response time window is 0 to 8 minutes after the start of the forecast.

[0080] The threat prediction input is formed by organizing target profiles, historical dynamic state sequences, behavioral stage characteristics, and task constraint information. This input is then fed into a multi-scale temporal convolutional threat prediction module to extract fire threat evolution features at different time scales, resulting in threat evolution features at the first, second, and third time scales. Specifically, the threat evolution features at the first time scale reflect recent abrupt changes in temperature and flame intensity, the features at the second time scale reflect the continuous expansion of the fire, and the features at the third time scale reflect the long-term trend of fire spread.

[0081] Based on the target profile, threat evolution features at various time scales are conditionally modulated, and stage gating weights are generated according to behavioral stage characteristics. In this embodiment, the stage gating weights corresponding to the threat evolution features at different time scales generated for the fire expansion stage are 0.35, 0.45, and 0.20, respectively. The conditionally modulated threat evolution features at each time scale are then weighted and fused to obtain the first target node. The evolutionary characteristics of the fusion threat.

[0082] Assuming a future prediction window of 6 minutes and a prediction time interval of 1 minute, the future prediction window comprises six prediction steps. Prediction steps 1 through 6 correspond to minutes 1 through 6 after the prediction begins, respectively. A corresponding prediction step embedding is generated based on the sequence number of each prediction step. And based on the fusion threat evolution characteristics and the prediction step embedding corresponding to each prediction step, The data for each prediction step is used to generate the initial threat prediction value.

[0083] In this embodiment, the data-driven initial threat prediction values ​​for each prediction step are as follows: ; Mechanism reference information is generated based on the fire spread rate, wind direction, effective range of firefighting resources, and estimated arrival time of resources. Based on this information, mechanism reference threat values ​​and mechanism confidence coefficients are then generated for each prediction step. The mechanism reference threat values ​​for each prediction step are as follows: ; The mechanism confidence coefficients for each prediction step are as follows: ;

[0084] Based on the mechanism confidence coefficient, the data-driven initial threat prediction value and the mechanism reference threat value for each prediction step are fused. Specifically, based on the first target node... , No. The final threat prediction value for each prediction step can be expressed as: ; in, express The initial threat prediction value is driven by data from each prediction step. Indicates the first The mechanism of each prediction step references the threat value. Indicates the first target node In the The mechanism confidence coefficient corresponding to each prediction step.

[0085] For example, in the fourth prediction step, the data-driven initial threat prediction value is 0.82, the mechanism-referenced threat value is 0.86, and the mechanism confidence coefficient is 0.65. Therefore, the final threat prediction value for the fourth prediction step is: ; After fusion, the final threat prediction values ​​for each prediction step are approximately as follows: ; The final threat prediction values ​​from each prediction step constitute the first target node. The fire threat curve within the future prediction window. Assuming a preset high-threat threshold of 0.80, the first target node... The threat values ​​exceeded the high threat threshold in both the 4th and 5th prediction steps. The peak threat was approximately 0.85, and the peak arrival time was 4 minutes after the prediction began; therefore, the first target node was identified. The area is in a high-threat state within the future forecast window, and the period around the 4th minute after the forecast begins is the key time window for intervention.

[0086] Write the fire threat curve, peak threat, peak arrival time, and high threat status back to the first target node. The task decision module, based on the updated target node attributes, will set the first target node... The target was identified as a priority target, and the following dispatch measures were taken: The reconnaissance drone was controlled to continuously acquire the first target node. The system receives information on fire developments; immediately dispatches firefighting drones to the target area within 2 minutes of the forecast's start to initiate firefighting efforts; dispatches fire trucks to the target area near the peak arrival time to conduct continuous firefighting operations; when communication quality falls below the preset communication quality threshold of 0.70, it dispatches a communication relay drone to establish the first target node. The communication relay link with the command center; simultaneously, based on northeasterly wind conditions, priority is given to monitoring and protecting the first target node. Important facilities area downwind.

[0087] Through the above processing, the system is able to identify the first target node. The threat level is highest around the 4th minute, and the dispatch of firefighting drones and communication relay drones is completed before the threat peak is reached, thus providing a dynamic decision-making basis for the early handling of fire threats.

[0088] The above values ​​are only used to illustrate the specific implementation process of the present invention and do not constitute a limitation on the prediction window length, prediction time interval, threat threshold, input parameters or resource response time.

[0089] Example 2: A Task Graph Threat Update System Based on Temporal Convolution

[0090] Based on the task graph threat update method based on temporal convolution in Embodiment 1, this invention also provides a task graph threat update system based on temporal convolution, such as... Figure 2 As shown, it includes:

[0091] The task graph construction module is used to construct a task graph, which includes a set of nodes, a set of relationships between nodes, node attributes, and a set of relationship attributes. The node set includes at least target nodes, resource nodes, task role nodes, and constraint nodes. The node attribute set includes target profile, historical dynamic state sequence, behavioral stage characteristics, and task constraint information related to the target node. Among them, the target profile is used to describe the attribute differences of the target, the historical dynamic state sequence is used to describe the dynamic behavior of the target over time, and the behavioral stage characteristics are used to describe the stage the target is in.

[0092] The prediction input generation module is used to obtain the target profile, historical dynamic state sequence, behavioral stage characteristics and task constraint information of the target node to be updated, and form the threat prediction input.

[0093] The temporal convolutional threat prediction module extracts threat evolution features of the target node to be updated based on the threat prediction input. This module can employ a multi-scale temporal convolutional structure to extract threat change features across different time ranges. The multi-scale temporal convolutional structure can also be implemented in parallel, either serially or hierarchically. This invention does not limit the specific structure, as long as it can extract the target threat evolution features.

[0094] The feature fusion module conditionally modulates the threat evolution features at various time scales based on the target profile, and performs stage-gated fusion on the conditionally modulated threat evolution features based on behavioral stage features to obtain the fused threat evolution features.

[0095] The mechanism fusion module is used to generate a data-driven initial threat prediction value corresponding to the target node to be updated based on the fused threat evolution characteristics; generate a mechanism reference threat value and a mechanism credibility coefficient based on the mechanism reference information corresponding to the target node to be updated; and fuse the data-driven initial threat prediction value and the mechanism reference threat value based on the mechanism credibility coefficient to obtain the final threat prediction value.

[0096] The threat attribute generation module is used to generate threat curves within the future prediction window based on the final threat prediction value, and to determine the peak threat, peak arrival time, and high threat status based on the threat curves within the future prediction window.

[0097] The task graph update module is used to write the threat curve, peak threat, peak arrival time and high threat status within the future prediction window back to the target node to be updated in the task graph.

[0098] The task graph threat update system based on temporal convolution of the present invention also includes a task decision interface module, which is used to provide updated target node attributes to the target sorting module, task chain scoring module, time window matching module or resource scheduling module.

[0099] Through the above system structure, the present invention can realize the complete process from target state acquisition, temporal threat prediction, dynamic attribute generation to task graph node update, so that the prediction results can be directly embedded into the task graph calculation structure, instead of remaining at the level of independent prediction output.

[0100] The temporal convolutional threat prediction module is trained using threat curves within real windows from historical samples before being used for task graph threat updates. Let the number of training samples be... Future forecast windows include One prediction step, target node In the The final threat prediction value for each prediction step is The corresponding real threat value is The threat curve predicts the loss as follows: ; in, The error function representing the difference between the predicted and actual values ​​can be at least one of absolute error, squared error, or smoothed absolute error. By minimizing the threat curve prediction loss, trainable parameters are determined in the temporal convolutional threat prediction module, the profile modulation structure, the stage gating structure, the threat decoding structure, and the mechanism credibility generation structure.

[0101] It should be noted that the temporal convolutional threat prediction module functional structure in this embodiment is only one optional implementation. This invention does not limit the specific neural network parameters, training data size, training epochs, loss function form, or threshold settings. Any implementation that uses temporal convolution to extract target threat evolution features and combines them with target profiles, stage features, or mechanistic reference information to generate dynamic threat decision attributes should fall within the protection scope of this invention.

[0102] This invention relates to a task graph threat update method and system based on temporal convolution, which can be applied to scenarios requiring dynamic task decisions based on changes in target state. By writing the threat curve, peak threat, peak arrival time, and high-threat status within the future prediction window back to the target node in the task graph, subsequent task decision-making processes can directly invoke the dynamic threat decision attributes of the target node.

[0103] The target prioritization module can determine target priority based on peak threat, high threat status, and target value.

[0104] The candidate task chain scoring module can evaluate the degree of matching of candidate task chains within a time window based on the threat curve and peak arrival time within the future prediction window.

[0105] The resource scheduling module can determine whether to allocate resources in advance and when to respond to resources based on the high threat status and peak arrival time.

[0106] Therefore, this invention enables dynamic threat prediction results to directly serve the task graph calculation, task chain generation, and resource scheduling processes, thereby improving the timeliness, adaptability, and stability of task decisions in dynamic scenarios.

Claims

1. A task graph threat update method based on temporal convolution, characterized in that, Includes the following steps: Construct a task graph, which includes a set of nodes, a set of relationships between nodes, a set of node attributes, and a set of relationship attributes; The node set includes target nodes, resource nodes, task role nodes, and constraint nodes; The node attribute set includes target profile, historical dynamic state sequence, behavioral stage characteristics, and task constraint information; Select any target node and perform the following target node attribute update steps: Obtain the node attributes of the target node to form the threat prediction input, and extract the threat evolution characteristics of the target node; Based on the target profile and behavioral stage features, the threat evolution features are sequentially subjected to conditional modulation and stage-gated fusion to obtain the fused threat evolution features; Based on the fused threat evolution characteristics, a threat curve for the target node within a future prediction window is generated, and based on the threat curve within the future prediction window, peak threat, peak arrival time, and high threat status are generated. Write the threat curve within the future prediction window, the peak threat, the peak arrival time, and the high threat status back to the corresponding target node in the task graph to complete the target node attribute update.

2. The method according to claim 1, characterized in that, The threat prediction input is represented as follows: ， in, Represents the target node Threat prediction input, Indicates the target node number. Represents the characteristic organization function, Represents the target node The target profile Represents the target node Historical dynamic state sequence Represents the target node Behavioral stage characteristics, Indicates the relationship with the target node Relevant task constraint information; The task map is a forest fire emergency response task map. The target nodes in the task map include forest fire points, suspected fire areas, fire spread areas, or important facility areas threatened by fire. The resource nodes in the mission diagram include reconnaissance drones, firefighting drones, fire trucks, command centers, and communication relay drones. The task role nodes include fire reconnaissance role, communication relay role, and fire fighting role; The constraints include drone endurance constraints, communication coverage constraints, no-fly zone constraints, and disposal time limit constraints; The target profile includes vegetation density, type of combustible material, terrain slope, regional importance, and distribution of surrounding facilities; The historical dynamic state sequence includes temperature, smoke concentration, flame intensity, fire spread rate, wind speed, and wind direction at continuous observation times; The behavioral phases include the initial discovery phase, the fire expansion phase, the stable and controlled phase, and the reignition detection phase. Task constraint information includes the estimated arrival time of resources, the scope matching result, the communication status, and the allowed processing time window.

3. The method according to claim 1, characterized in that, The threat evolution features of the target node are extracted using a temporal convolutional threat prediction module, which includes at least two temporal convolutional branches with different time scales; the threat evolution features are threat evolution features under multiple time scales. Each temporal convolutional branch has the same structure, including an input mapping layer and at least one temporal convolutional block; The input mapping layer is used for threat prediction input. The feature dimension is transformed and mapped to a hidden feature representation; the temporal convolutional block is used to extract time-series features from the hidden feature representation to obtain threat evolution features at a single time scale.

4. The method according to claim 1, characterized in that, The specific steps for conditionally modulating the threat evolution characteristics are as follows: First, create a target profile. Mapped to image embedding: ， in, This represents the image mapping function. Indicates the target profile The mapped embedding vector; Regenerate scaling parameters and offset parameters , represented as: ， in, This represents the parameter generation function; The threat evolution features are modulated channel by channel to obtain the threat evolution features conditionally modulated by the target profiling. , represented as: ， in, This represents element-wise multiplication. Indicates the first Threat evolution characteristics at various time scales.

5. The method according to claim 1, characterized in that, Threat evolution characteristics after conditional modulation The phased gating fusion is performed as follows: First, the characteristics of the behavioral stages Mapped to stage embedding: ， in, Represents the stage mapping function, Indicating behavioral stage characteristics The mapped embedding vector; No. The gating weights corresponding to each time scale are: ， in, Indicates the first The gating weights corresponding to each time scale This represents the gating weight generation function. The nth element in the output vector of the gated weight generation function The components corresponding to each time scale This represents the component corresponding to the j-th time scale in the output vector of the gated weight generation function. Indicates the maximum values ​​of k and j; And satisfy ； The fusion is represented as: ， in, This indicates the threat evolution characteristics after stage-gated fusion.

6. The method according to claim 1, characterized in that, Based on the fused threat evolution characteristics, a threat curve for the target node within the future prediction window is generated, specifically as follows: Suppose that the future prediction window includes H prediction steps, based on the target node. Corresponding mechanism reference information , generate the first Mechanistic reference threat value corresponding to each prediction step , represented as: ， in, The mechanism reference generating function, The value is an integer, ranging from 1 to H; Generate target node In the The mechanism reliability coefficient corresponding to each prediction step , represented as: ， in, This represents the Sigmoid mapping function. This indicates the threat evolution characteristics after stage-gated fusion. Indicates the first The prediction step embedding corresponding to each prediction step. This represents the credibility generation function. ; The threat decoding function generates the first Data-driven initial threat prediction values ​​for each prediction step , represented as: ， in, This represents the threat decoding function; No. Final threat prediction value for each prediction step , represented as: ， The final threat prediction values ​​of each prediction step constitute the threat curve of the target node within the future prediction window. , represented as: 。 7. The method according to claim 6, characterized in that, Mechanism Reference Information This includes at least one of the following: distance variation, scope of action, task rules, and stage constraints.

8. The method according to claim 1, characterized in that, After writing the threat curve within the future prediction window, the peak threat, the peak arrival time, and the high-threat status back to the corresponding target node in the task graph, the updated target node attribute set is obtained. Represented as: ， in, Represents the target node initial properties, , Represents the target node The target profile Represents the target node Historical dynamic state sequence Represents the target node Behavioral stage characteristics, Indicates the relationship with the target node Relevant task constraint information, Indicates the target node number. This represents finding the union of sets; Represents the target node Dynamic threat decision-making attributes, , Represents the target node Threat curve within the future forecast window, Indicates peak threat. Indicates the peak arrival time. This indicates a high-threat status.

9. The method according to claim 1, characterized in that, The task graph contains multiple target nodes. The target node attributes are updated sequentially for each target node until the target node attributes of all target nodes are updated.

10. A task graph threat update system based on temporal convolution, characterized in that, include: The task graph construction module is used to construct a task graph that includes a set of nodes, a set of relationships between nodes, a set of node attributes, and a set of relationship attributes. The set of nodes includes target nodes, resource nodes, task role nodes, and constraint nodes. The set of node attributes includes target profiles, historical dynamic state sequences, behavioral stage characteristics, and task constraint information. The module also determines the target nodes to be updated. The prediction input generation module is used to obtain the target profile, historical dynamic state sequence, behavioral stage characteristics and task constraint information of the target node to be updated, and form the threat prediction input; The temporal convolutional threat prediction module is used to extract the threat evolution features of the target node to be updated at various time scales through a temporal convolutional model. The feature fusion module is used to conditionally modulate the threat evolution features at various time scales based on the target profile, and to perform stage-gated fusion of the conditionally modulated threat evolution features based on the behavioral stage features to obtain the fused threat evolution features. The mechanism fusion module is used to generate a data-driven initial threat prediction value corresponding to the target node to be updated based on the fused threat evolution characteristics. Generate a mechanism reference threat value and mechanism credibility coefficient based on the mechanism reference information corresponding to the target node to be updated; Based on the mechanism credibility coefficient, the data-driven initial threat prediction value and the mechanism reference threat value are then fused to obtain the final threat prediction value. The threat attribute generation module is used to generate a threat curve of the target node to be updated within a future prediction window based on the final threat prediction value, and to generate peak threat, peak arrival time and high threat status based on the threat curve within the future prediction window. The task graph update module is used to write the threat curve, the peak threat, the peak arrival time, and the high threat status within the future prediction window back to the target node to be updated in the task graph.

Images