Dynamic tactical knowledge graph construction and threat evaluation system and method based on environment embedded graph neural network

Abstract

The invention discloses a dynamic tactical knowledge graph construction and target attackable value analysis system and method based on an environment embedded graph neural network, and belongs to the technical field of military intelligent decision making and firepower command control. The system constructs a tactical knowledge graph fused with firepower application expert experience for tactical rule recessive and environmental constraint dynamic problems in target threat assessment. Based on an entity-relation-environment attribute triple structure, an ontology model covering a typical combat scene is established, and dynamic quantitative analysis of target attackable value is supported. An environment embedded graph neural network is innovatively designed, a battlefield environment is used as an internal structure variable of a graph, and deep coupling and self-adaptive reasoning of knowledge representation and a real-time situation are achieved. The system can comprehensively consider information such as enemy conditions, my conditions, typical scenes and the like, realizes accurate evaluation of multi-target threats and values, assists a commander in accurately capturing important high-threat high-value targets, and quickly forms a wave-order striking target list.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of artificial intelligence, military operations research, and command information systems technology. Specifically, it relates to a system and method for real-time analysis of target attack value using dynamic knowledge graphs and graph neural networks. It is particularly suitable for informationized and intelligent warfare, enabling automated and interpretable threat quantification and ranking of highly dynamic, highly adversarial, and heterogeneous battlefield targets, providing core decision support for commanders to quickly generate a target list. Background Technology

[0002] Modern warfare is characterized by an unprecedentedly accelerated pace, multi-dimensional battlefields, and a vast number of targets with diverse types and unpredictable behaviors. Against this backdrop, rapid, accurate, and automated target threat and value assessment has become crucial for achieving superiority in the Observation-Judgment-Decision-Action (OODA) cycle. However, existing mainstream threat assessment technologies face significant challenges and struggle to meet the demands of future operations:

[0003] (1) Static model is out of touch with dynamic battlefield: Existing assessment systems are mostly based on multi-attribute decision theory (such as AHP, TOPSIS, fuzzy comprehensive evaluation), the core of which relies on a static weight system set by experts in advance. Once these weights are set, they are difficult to adjust and cannot adapt to the drastic changes in battlefield environment from plain assault to urban street fighting, from electronic silence to full-frequency suppression, resulting in assessment results that are seriously deviated from the actual threat.

[0004] (2) Data and knowledge are disconnected, and decision-making is like a "black box": In recent years, data-driven methods based on deep learning have been introduced. Although they can learn complex patterns, their decision-making process lacks interpretability and is a typical "black box" model. Commanders cannot understand why the model judges a certain target as a high threat, and it is difficult to trust and adopt its results.

[0005] (3) Lack of ability to model complex relationships: Battlefield threats are not isolated attributes of targets, but rather stem from complex networks of interactions between targets. Traditional methods are unable to explicitly model and reason about such complex networks of "entity-relationship-environment", and can only perform shallow attribute weighting, ignoring the essence of system confrontation.

[0006] (4) Weak adaptability and anti-interference capability: When faced with deliberate deception, interference or sudden battlefield situations by the enemy, the existing system often requires manual intervention to adjust parameters or rules, resulting in a slow response. It lacks a mechanism that can quickly learn and self-adjust from continuous interaction, and its robustness is insufficient in a highly adversarial environment. Summary of the Invention

[0007] (a) Purpose of the invention

[0008] This invention aims to overcome the aforementioned technical bottlenecks and provide a system and method for constructing dynamic tactical knowledge graphs and analyzing target attackability value based on Environment Embedded Graph Neural Networks (EGNNs). The core objectives of this invention include:

[0009] (1) To achieve unified dynamic representation of multi-source heterogeneous information (enemy situation, our situation, environment);

[0010] (2) Construct a tactical knowledge graph that integrates expert experience to support rule-adaptive reasoning;

[0011] (3) To achieve accurate, explainable, and real-time assessment of the attackable value of the target;

[0012] (4) Meets the agile and adaptable needs of threat assessment in typical scenarios such as mountains and plains.

[0013] (II) Technical Solution

[0014] 1. System Overall Architecture

[0015] The system of this invention adopts a hierarchical and modular design, the core of which is to construct a closed loop of "perception-representation-reasoning-evaluation-evolution". The system consists of five core modules, in which data and knowledge flow, transform, and increase in value.

[0016] 2. Detailed Description of Core Modules

[0017] 2.1 Dynamic Battlefield Awareness Module

[0018] This module is the system's "sensory organs," responsible for extracting, cleaning, aligning, and structuring critical information from the noisy ocean of battlefield data.

[0019] Target attribute perception unit: Not only extracts basic states (position, velocity), but also constructs the target's capability vector, for example: Each dimension is rated from 0 to 1.

[0020] Tactical rule analysis unit: This is the key to making tacit knowledge explicit. It transforms military doctrine and expert experience into formalized rules, such as: "IF Target type = Command post AND Distance < 20km THEN Threat level = High WITH Confidence level = 0.9".

[0021] Dynamic environment sensing unit: It performs quantitative modeling of the environment. Terrain shading coefficient. Electromagnetic intensity was calculated using digital elevation modeling and line-of-sight analysis. It is obtained by interpolation of spectrum monitoring data.

[0022] 2.2 Tactical Knowledge Graph Construction Module

[0023] This module is the system's "Memory and Worldview," which organizes perceived information into a structured, dynamic knowledge network.

[0024] Ontology design: Defines the concepts, attributes, and relationships of the battlefield domain, including target types, combat units, environmental entities, etc.

[0025] Dynamic graph instantiation: in At any given moment, create a graph based on the perceived data. .

[0026] Node: Each battlefield entity is a node, with its initial characteristics... It is obtained from attribute encoding.

[0027] Edges: Relationships between entities constitute edges. The key innovation lies in edge properties. It encodes the environmental state connecting two entities, enabling the map to dynamically reflect changes in the battlefield environment.

[0028] 2.3 Context-Embedded Graph Neural Network (EGNN) Inference Module

[0029] This is the system's "brain," responsible for deep, interpretable reasoning on the dynamic graph. The core idea of ​​EGNN is that nodes update their cognition by exchanging information with their neighbors, and the efficiency (attention) and content (message) of each information exchange are profoundly influenced by the current environmental properties of the edge connecting them.

[0030] Assuming in the first Layers, nodes The characteristics are The update process is as follows:

[0031] (1) Generate environment-aware messages: for each neighbor to its characteristics and the environmental properties that connect them Together they generate a single message.

[0032]

[0033] (2) Computational environmentally sensitive attention: assessment in The environment of time Below, neighbor Information How important is it?

[0034]

[0035] in These are learnable parameters. The term directly injects environmental attributes into attention calculation, which makes the association importance of the same pair of entities different in different environments (such as clear sky vs. dense fog).

[0036] (3) Aggregation and Update: Nodes It aggregates all messages from its neighbors, weighted by environmental factors, and updates its own status.

[0037]

[0038] go through Layers are iterated in this way, and the features of each node All of them contain its The complex interaction results of all entities, relationships, and environmental information within the jump neighborhood form a deep contextual representation.

[0039] 2.4 Target Attack Value Analysis Module

[0040] High-order node features based on EGNN output Conduct a target threat value assessment.

[0041] (1) Global Situation Summary: First, obtain a macro-level situation summary of the entire battlefield through graph pooling (such as global average pooling):

[0042]

[0043] (2) Individual Threat Score: Then, the deep features of a single target are combined with the global situational features and fed into a multilayer perceptron (MLP) for threat scoring.

[0044]

[0045] This design allows the evaluation of a single target to take into account both its own characteristics and those of its local network, as well as its strategic position and role in the overall battlefield.

[0046] 2.5 Offline Training and Online Fine-Tuning Module

[0047] Offline pre-training: Using massive amounts of scenario data in a safe and controlled simulation environment. Training system. The loss function is a hybrid form:

[0048]

[0049] in The conflict between the output of the penalty model and the formalized tactical rules is addressed by injecting expert knowledge as a soft constraint into the model, thereby improving the rationality and interpretability of the decision-making.

[0050] Online fine-tuning: After deployment, the system continuously monitors its performance (such as the degree to which the evaluation results match the commander's final decision). When performance drift exceeds a threshold, lightweight online learning is initiated, using recent data to fine-tune the last few layers of the EGNN or the environmental encoder to achieve rapid adaptation.

[0051] (III) Beneficial Effects

[0052] (1) Significantly improved assessment accuracy: By explicitly modeling the dynamic coupling between the target, relationship and environment, this invention can capture complex threat emergence patterns that are ignored by traditional methods.

[0053] (2) Real-time decision-making meets tactical requirements: Thanks to the high efficiency of parallelism of graph neural networks, on an embedded command platform equipped with domestic processors, a full-map reasoning and threat ranking of the battlefield map is completed, which fully meets the tactical real-time response requirements.

[0054] (3) Provides deep interpretability and enhances commanders’ trust: The system can output a “threat attribution report”, which visually displays high attention edges and key reasoning paths, enabling commanders to understand the basis of model decision-making.

[0055] (4) It has strong adaptability and robustness: the online fine-tuning mechanism enables the system to quickly adapt to new opponents and new environments, and the evaluation accuracy remains stable when switching between mountainous and plain scenes.

[0056] (5) Supports rapid generation of multi-target threat lists: The system can generate threat / value lists for 10 targets in a timely manner, meeting the system operation requirements and assisting commanders in quickly forming a wave-style strike target list. Attached Figure Description

[0057] Figure 1 System overall architecture diagram

[0058] Figure 2 Example of a dynamic tactical knowledge graph

[0059] Figure 3 Schematic diagram of single-layer message passing and aggregation in an Context Embedded Graph Neural Network (EGNN)

[0060] Figure 4 : System reasoning process. Detailed Implementation

[0061] Implementation Example: Threat Assessment of Mobile Operational Targets in Plains Areas

[0062] 1. Scene characteristics

[0063] Environmental characteristics: Open plains with good visibility, but susceptible to electromagnetic interference.

[0064] Target characteristics: high speed, strong formation coordination, including high-speed vehicles and concealed fortifications.

[0065] The core challenge is to accurately assess the threat value of targets and identify key nodes in the context of rapid target mobility.

[0066] 2. System Adaptive Adjustment and Implementation

[0067] For plains scenarios, the system is configured accordingly:

[0068] Enhanced Environmental Perception: Environmental Vectors The weighting of dimensions such as "electromagnetic interference intensity" and "line-of-sight distance" is strengthened.

[0069] Temporal modeling enhancement: The EGNN input is fused with a map sequence of 3 consecutive frames to enhance the capture of target maneuver trends;

[0070] Rule base expansion: Add tactical rules specific to plains areas, such as "armored groups in open areas should be prioritized for attack".

[0071] 3. Implementation Results

[0072] In a simulated "plains maneuver warfare" scenario:

[0073] The system successfully identified the enemy command vehicle and air defense system as the highest threat targets.

[0074] The time allocation for 10 different types of targets meets the requirements for operational use.

[0075] The threat list output by the system is consistent with the expert assessment results.

Claims

1. A system for constructing dynamic tactical knowledge graphs and analyzing the attackability value of targets based on environment-embedded graph neural networks, characterized in that, include: (1) A dynamic battlefield perception module, used to acquire and fuse multi-source heterogeneous battlefield data in real time, including target static / dynamic attributes, formalized tactical rules and dynamic environmental parameters; (2) A tactical knowledge graph construction module, used to instantiate the perception data into a dynamic temporal knowledge graph containing "entity-relationship-environment attribute" triples based on a predefined domain ontology model. ,in To attach to the edge (3) An environment-embedded graph neural network inference module for receiving time-series dynamic knowledge graph sequences. Through multi-layered environment-sensitive message passing and aggregation, it outputs context-enhanced feature representations of all entity nodes at the current moment. (4) Target attack value analysis module, based on Global situation features obtained by graph pooling It calculates the comprehensive threat value of all enemy targets in parallel using a learnable evaluation function and generates a real-time ranked list. (5) Offline training and online fine-tuning module, used to pre-train the system using historical or simulation data, and to achieve incremental adaptive optimization of the model based on online interactive feedback.

2. The system according to claim 1, characterized in that: The dynamic battlefield awareness module further includes: (1) Target attribute perception unit, used to extract the target type (including infantry, transport vehicle, wheeled armored vehicle, tracked vehicle, light fortification, command post, etc.), identity, location, speed, heading, formation to which it belongs and combat capability vector; (2) Tactical rule parsing unit, which uses semantic parsing and logic programming technology to transform unstructured military doctrines and expert experience into machine-executable probabilistic logic rules or production rules, in the form of IF <condition> THEN <conclusion> WITH <confidence / weight>; (3) Dynamic environment sensing unit, used to synchronously process multi-dimensional environmental flow data from geographic information system, meteorological sensor and electronic support measures, and output standardized environmental feature vector. It supports environmental modeling for typical combat scenarios such as mountains and plains.

3. The system according to claim 1, characterized in that: In the tactical knowledge graph construction module, environmental attributes It is a time Related dimensional vector, through a trainable environment encoder generate: Entity node initial features Its static attributes and dynamic states are concatenated and then processed by a physical encoder. The mapping yields: .

4. The system according to claim 1, characterized in that: The core of the environment-embedded graph neural network inference module is the environment-sensitive message passing layer. Layer to node The operation is defined as follows: in, The function will connect neighboring nodes Features With edge environment attributes Merge to generate messages; Function computation relies on the normalized attention weights of nodes for features and environmental attributes. ; The function aggregates the weighted messages and updates its own characteristics. (Stacking) The final node features are obtained after layering. .

5. The system according to claim 4, characterized in that: The The function employs an additive attention mechanism, specifically calculated as follows: in, For learnable parameter matrix, This represents vector concatenation. This mechanism enables the model to dynamically evaluate in different environments. Below, neighbor For the central node The importance of.

6. The system according to claim 1, characterized in that: The evaluation function of the target attack value analysis module is: in, For graph-level pooling functions, For the overall situational awareness, and For learnable parameters, Constrain the output of the Sigmoid activation function to interval, That is the goal The normalized threat score.

7. The system according to claim 1, characterized in that: The offline training phase of the offline training and online fine-tuning modules employs a hybrid loss function combining supervised learning and rule consistency constraints. in, Penalty model prediction results and formalized tactical rules The degree of conflict, and These are hyperparameters. During the online fine-tuning phase, a sliding window-based online learning algorithm is employed, using recent interaction data to quickly and in small batches update the gradients of some model parameters.

8. A method for dynamic tactical target attack value analysis based on the system described in any one of claims 1-7, characterized in that, Includes the following steps: (1) System initialization. Load the tactical knowledge graph ontology, initialize the EGNN model parameters, and configure the environment encoder, entity encoder, and parameters of each functional module; (2) Offline pre-training. Using large-scale historical battle examples or high-fidelity simulation datasets. By minimizing the loss function Train the model end-to-end until it converges; (3) Online deployment and dynamic perception. The system accesses real-time battlefield data streams, and the dynamic battlefield perception module operates at a frequency of... Periodically output structured perception results; (4) Knowledge Graph Construction and Update. The tactical knowledge graph construction module will... Real-time awareness of result instantiation, updating or generating dynamic knowledge graphs And call the environment encoder to calculate the properties of all edges. ; (5) Graph Neural Network Inference. Input the trained EGNN, perform forward propagation, and then... Layer message passing yields the deep feature representations of all nodes. ; (6) Target Attack Value Calculation and Ranking. The target attack value analysis module calculates the threat value of each enemy target in parallel. and generate according to A real-time threat list sorted in descending order; (7) Output and Decision Support. The threat list and key reasoning evidence (such as high-attention edges) are pushed to the command and decision-making system, and a visual interface is provided; (8) Online feedback and adaptation. Collect feedback signals such as commander confirmation and target damage effect. If the system performance evaluation is lower than the threshold, trigger the online fine-tuning process, use recent data to incrementally update the model, and return to step S3.

9. The method according to claim 8, characterized in that: In step S4, for complex maneuvering targets, a spatiotemporal trajectory prediction module is used to predict their short-term future state, and the predicted state is incorporated into the form of "virtual nodes" or node future attribute vectors. This allows threat assessment to have a certain degree of foresight.

10. The method according to claim 8, characterized in that: The method supports multi-level threat assessment by defining a hierarchical relationship of "target-subsystem-platform-formation" in the knowledge graph and designing a hierarchical aggregation mechanism in EGNN to synchronously output cross-granularity threat assessment reports from single equipment to campaign units.

Images