Intelligent war game deduction simulation system based on digital twinborn fusion large model

Abstract

The invention relates to the technical field of intelligent war game deduction, in particular to an intelligent war game deduction simulation system based on a digital twin fusion large model. Comprising a digital twin modeling unit which adopts a dynamic precision adaptive modeling mechanism; a large model decision unit; a deduction execution unit; and a dynamic interaction unit. A dynamic precision self-adaptive modeling mechanism is adopted by the digital twin modeling unit, the model precision can be dynamically switched on demand, the detail simulation demand and the computing resource consumption are balanced, and meanwhile, the sub-second-level dynamic synchronization of a virtual model and a war game deduction scene is realized through a time sequence calibration algorithm; and space confrontation, time sequence decision and rule triggering features are fused through the multi-modal feature analysis module, and the Wargame rule verification module is combined to embed entity performance boundaries and scene rule constraints, so that compliance confrontation decision parameters conforming to Wargame deduction logic can be generated, and the problem of insufficient model precision adaptation and decision constraints is solved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent war game, in particular, to an intelligent war game simulation system based on digital twin fusion large model. BACKGROUND

[0002] War game is the core means of modern national defense strategy planning, tactical and operational method verification and command decision training. Its core is to pre-act the process and results of confrontation in virtual space through modeling and simulation technology. With the improvement of the complexity of military systems and the deepening of joint operation needs, traditional war game systems face serious challenges in terms of simulation fidelity, decision intelligence, and real-time continuity of the simulation process. How to build a simulation system that can simulate real battlefield with high fidelity, embed intelligent decision-making capabilities, and ensure efficient synchronization throughout the process has become a key direction for the development of the field.

[0003] In the prior art, related patents have explored digital twin applications, simulation process optimization, etc. For example, Chinese patent CN202311436752.8 discloses a digital twin war game weather data simulation and intelligent decision-making method and system. Its technical points are: combining digital twin technology with weather data simulation, building six sub-models of task target constraint, marine battlefield geographic information analysis, force and equipment deployment, enemy data analysis, weather comprehensive prediction, and action game, combining actual weather data with the model to simulate the influence of different weather on the simulation, and providing reference for military decision-making; and Chinese patent CN202211414513.8 discloses a war game system, which includes a simulation design module (receiving round setting instruction to generate round planning result), a simulation control module (dynamically managing simulation round information to generate round event information), a simulation simulation module (simulating to generate round result according to round event information), and a simulation command module (receiving command instruction to generate command information), which can solve the problems of fixed time and space resolution, weak rule adaptability, inflexible control, and difficulty in realizing multi-level linkage in traditional war game, and improve the efficiency and flexibility, scalability of the system.

[0004] Although the above technical solutions have corresponding design advantages, the above technical solutions still have the following technical defects: firstly, the model precision adaptation and decision constraint are insufficient: CN202311436752.8 does not dynamically adjust the model precision according to the frequency of chess entity interaction and spatial distance, cannot balance detailed simulation and resource consumption, and has no high-precision timing calibration mechanism; the decision only surrounds the weather data, does not fuse multi-modal features such as space and time, and does not embed entity performance boundaries and scene rule constraints, which is easy to produce illogical decisions. Secondly, virtual fidelity and deduction coherence are lacking: CN202211414513.8 has no high-fidelity virtual model of chess entity, which is difficult to reproduce entity attributes and real-time state; lacks a low-latency data interaction channel, and has no deduction deviation dynamic correction mechanism, which cannot adjust the deviation of the expected deduction path in time, affecting the coherence. In view of this, we propose an intelligent war game deduction simulation system based on digital twin fusion large model. SUMMARY

[0005] The purpose of the present application is to provide an intelligent war game deduction simulation system based on digital twin fusion large model to solve the problems of insufficient model precision adaptation and decision constraint and lack of virtual fidelity and deduction coherence in the background art.

[0006] To solve the above technical problems, the purpose of the present application is to provide an intelligent war game deduction simulation system based on digital twin fusion large model, which comprises: A digital twin modeling unit adopts a dynamic precision self-adaptive modeling mechanism to construct a hierarchical virtual model of a war game confrontation scene through multi-source heterogeneous data fusion. The hierarchical virtual model includes an "antagonistic entity mapping layer" and a "scene rule embedding layer", which can automatically adjust the model precision level according to the war game deduction granularity requirement, and realize sub-second dynamic synchronization of the virtual model and the war game deduction scene through a timing calibration algorithm; A large model decision unit is used to receive the war game scene state data output by the digital twin modeling unit, adopts a multi-modal feature fusion and war game rule constraint embedding mechanism, fuses the spatial confrontation features and time decision features of the war game scene, and embeds the performance boundaries of the war game confrontation entity and the scene rules into the pre-trained model as constraint conditions to generate confrontation decision parameters that conform to the war game deduction logic; wherein the large model decision unit comprises a multi-modal feature analysis module, a fusion reasoning module and a war game rule verification module; A deduction execution unit adopts a war game space-time coupled deduction mechanism, based on the war game virtual model of the digital twin modeling unit and the confrontation decision parameters of the large model decision unit, drives the confrontation process according to the war game deduction time sequence, simulates the war game confrontation result through space-time correlation calculation, and has a dynamic correction mechanism that can adjust the deduction path according to the deviation of the virtual model and the expected state of the war game deduction. The dynamic interaction unit is used to establish a two-way data channel for wargames between the digital twin modeling unit, the large model decision-making unit, and the simulation execution unit, so as to realize low-latency interactive transmission of wargame scene feature parameters, adversarial decision-making instructions, and simulation status data, and ensure the real-time performance and continuity of wargame simulation.

[0007] As a further improvement to this technical solution, the digital twin modeling unit includes a geometric feature mapping module, an attribute parameter binding module, and a state label synchronization module, wherein: The geometric feature mapping module is based on the 3D scanning data and parametric modeling rules of the wargame adversarial entities. It uses multi-detail level model loading technology to generate and store the 3D contour data and spatial coordinates of the wargame adversarial entities, and supports dynamic switching of precision from millimeter level to meter level. The attribute parameter binding module is based on the wargame simulation rule base and entity type configuration table. It uses dynamic parameter association technology to bind the inherent performance parameters of wargame adversaries to entity types and automatically updates the performance parameters when the state of the wargame adversaries changes. The state tag synchronization module uses real-time state triggering technology to mark the current state of the wargame adversarial entities based on the entity behavior result data output by the simulation execution unit, and the state switching delay is controlled in the sub-hundred millisecond range.

[0008] As a further improvement to this technical solution, the digital twin modeling unit also includes a precision level division module, a precision adjustment triggering module, and a timing calibration execution module, wherein: The precision level division module is based on the interaction frequency and spatial scale of the wargame adversaries and uses a distance-precision mapping rule to divide the model precision into three levels: Level 1 precision (millimeter level) is suitable for interaction scenarios where the distance between wargame adversaries is close, Level 2 precision (centimeter level) is suitable for scenarios where the distance between wargame adversaries is medium, and Level 3 precision (meter level) is suitable for scenarios where the distance between wargame adversaries is far. The precision adjustment triggering module is based on the relative motion data of the wargame adversaries and uses threshold monitoring technology. When the relative motion state between the wargame adversaries meets the preset triggering conditions, the precision is automatically upgraded; when there is no interaction between the wargame adversaries for a preset duration, the precision is automatically downgraded. The timing calibration execution module collects data from the physical scene with timestamps and uses time deviation correction technology. It calculates the time deviation between the virtual model and the physical scene through a timing calibration algorithm. If the time deviation exceeds the preset acceptable range, the state parameters of the virtual model are corrected through linear interpolation to ensure the synchronization between the virtual model and the war game scenario.

[0009] As a further improvement to this technical solution, the multimodal feature parsing module achieves structured extraction of wargame scene features through feature quantization and standardization processing, specifically including: For the spatial adversarial features in the wargame scenario state data, a coordinate system transformation method is used to convert the absolute coordinates of adversarial entities into quantitative parameters such as relative distance, azimuth angle, and terrain occlusion coefficient; For time-series decision characteristics, the historical action sequence of adversarial entities is extracted by using a sliding window to extract time-series parameters such as action interval and strategy switching frequency; For rule triggering features, convert the rule conditions such as resource threshold and round progress into binary feature codes; (1 if satisfied, 0 if not satisfied); The spatial adversarial features, temporal decision features, and rule-triggered feature parameters are normalized to a unified data dimension to form a multimodal fusion feature vector.

[0010] As a further improvement to this technical solution, the fusion reasoning module generates an initial adversarial decision-making scheme through feature-weighted fusion and knowledge graph reasoning, specifically including: An attention mechanism is used to dynamically weight the fused feature vector output by the multimodal feature parsing module, where the spatial feature weight increases linearly with the entity interaction intensity, and the temporal feature weight increases exponentially with the action coherence. A wargame adversarial knowledge graph is constructed based on a pre-trained model. The graph nodes contain entity types, action strategies, and rule clauses. The optimal decision logic for matching historical adversarial cases is calculated by the cosine similarity of the nodes. Based on the reasoning results, an initial adversarial decision-making scheme is output, which includes action priority, resource allocation ratio, and strategy alternative sequence. The scheme is stored in a structured data format, and the data fields include decision ID, target entity ID, and execution time sequence label.

[0011] As a further improvement to this technical solution, the wargame rule verification module ensures decision compliance through constraint matching and gradient correction, specifically including: Retrieve constraints that match the current scenario from the wargame rule base, including the performance boundaries of wargame adversaries and scenario rules; The parameters and constraints of the initial adversarial decision-making scheme are compared item by item to determine whether there are any violations such as parameter out-of-bounds or rule conflicts. The gradient adjustment method is used to correct violations. For performance boundary violations, the parameters are truncated according to the maximum allowable value. For scenario rule violations, resources are reallocated and the action sequence is adjusted according to the rule priority. After correction, the output adversarial decision parameters conform to the wargaming simulation logic.

[0012] As a further improvement to this technical solution, the simulation execution unit includes a time-series advancement module, a spatial interaction calculation module, and an adversarial result synthesis module, wherein: The timing progression module is used to drive the adversarial process according to the wargame simulation sequence, and to mark round-based simulations with round markers. As a time series unit, in the first At the start of each round, the adversarial decision parameters output by the large model decision unit are loaded; real-time simulations are performed using time steps. As a time series unit, in The state data of the wargame adversaries is updated synchronously within the interval; among which... Indicates the start time point of the real-time simulation. Indicates starting time point As a baseline, plus a time step The final time point obtained later; The spatial interaction computing module is used to process the spatial interaction of wargame adversaries, and obtains entity coordinates based on the wargame virtual model of the digital twin modeling unit. ,in Indicates the entity number and terrain parameters. That is, the terrain slope, used to calculate the attack effect value. and degree of movement obstruction Among them, attack effect value Related firepower parameters Values, entity spacing Terrain correction factor Degree of movement obstruction Correlation of terrain type coefficients Elevation difference ; The adversarial result synthesis module is used to integrate temporal data and spatial interaction results to generate wargame adversarial results, wherein entity life values ​​are updated in relation to attack effect values. Resource changes are correlated with interaction frequency, and rule triggering results are correlated with attack effectiveness values. Threshold determination, all results are marked with rounds. Stored as time-series tags.

[0013] As a further improvement to this technical solution, the dynamic correction mechanism of the inference execution unit also includes a deviation monitoring module, a correction strategy generation module, and a path adjustment execution module, wherein: The deviation monitoring module is used to calculate the deviation between the wargame virtual model and the expected state in the simulation, and compare the entity positions output by the digital twin modeling unit. Compared to the expected location Receive position deviation Compare resource values With expected resources Obtain resource deviation ; The correction strategy generation module is used to formulate a simulation path adjustment plan and set a deviation threshold. ,when and Maintain the current path if necessary; otherwise, generate a location compensation value. Resource compensation value and timing compensation step size ; The path adjustment execution module is used to apply correction parameters to adjust the inference path and update the entity coordinates. Resource values ​​updated to The subsequent simulation sequence was adjusted to The adjustment results are fed back to the digital twin modeling unit to update the wargame virtual model.

[0014] As a further improvement to this technical solution, the dynamic interaction unit includes a data acquisition module, a low-latency transmission module, and a state synchronization module, wherein: The data acquisition module is used to collect wargame scenario feature parameters from the digital twin modeling unit, collect adversarial decision instructions from the large model decision unit, and collect simulation status data from the simulation execution unit. The low-latency transmission module is used to construct a two-way transmission channel for wargame data. It adopts a lightweight encapsulation transmission technology based on the UDP protocol to achieve low-latency interactive transmission of wargame scene feature parameters, adversarial decision instructions, and simulation status data. The state synchronization module is used to synchronize the states of the wargame virtual scene of the digital twin modeling unit, the decision generation logic of the large model decision unit, and the simulation process of the simulation execution unit based on the data transmitted by the low-latency transmission module, so as to ensure the real-time performance and continuity of the wargame simulation.

[0015] As a further improvement to this technical solution, a wargaming result evaluation unit is also included. This unit is used to evaluate the effectiveness of the wargaming results output by the wargaming execution unit. It includes a result and benchmark acquisition module, a multi-dimensional evaluation module, and an evaluation result feedback module, wherein: The results and benchmark acquisition module is used to acquire wargame confrontation result data from the simulation execution unit, acquire wargame virtual model status data corresponding to the simulation stage from the digital twin modeling unit, and acquire data interaction delay records corresponding to the simulation stage from the dynamic interaction unit as evaluation benchmark data. The multi-dimensional evaluation module is used to judge the validity of the adversarial results by comparing them with existing standards. Specifically, it includes: judging the compliance of the deduction logic by the compliance standards output by the wargame rule verification module; judging the synchronization accuracy of the virtual scene by the sub-second synchronization requirements of the timing calibration execution module; and judging the real-time performance of data interaction by the low latency requirements of the dynamic interaction unit. If all three items meet the standards, the result is deemed valid. If any item does not meet the standards, it is marked as a result to be optimized and the non-compliant item is recorded. The evaluation result feedback module is used to store the effective evaluation results in the war game simulation result library, and to feed back the results to be optimized and the substandard items to the large model decision unit and the digital twin modeling unit. The information fed back to the large model decision unit is used to optimize the subsequent adversarial decision parameter generation logic, and the information fed back to the digital twin modeling unit is used to assist in adjusting the model accuracy level or the timing calibration frequency.

[0016] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention employs a dynamic precision adaptive modeling mechanism in the digital twin modeling unit, enabling on-demand dynamic switching of model precision to balance detailed simulation requirements with computational resource consumption. Simultaneously, a time-series calibration algorithm achieves sub-second dynamic synchronization between the virtual model and the wargame scenario. Furthermore, the large model decision unit integrates spatial adversarial, temporal decision-making, and rule-triggered features through a multimodal feature analysis module, and embeds entity performance boundaries and scenario rule constraints into the wargame rule verification module. This generates compliant adversarial decision parameters that conform to the wargame logic, solving the problems of insufficient model precision adaptation and decision constraints.

[0017] 2. This invention generates 3D contour data of wargame adversaries through a geometric feature mapping module in a digital twin modeling unit, associates and dynamically updates the entity's inherent performance parameters through an attribute parameter binding module, and marks the entity's real-time state through a status label synchronization module, thus constructing a high-fidelity wargame virtual model that accurately replicates the entity's geometric features, attribute parameters, and current state. Simultaneously, the dynamic interaction unit employs lightweight encapsulation and transmission technology based on the UDP protocol to construct a low-latency bidirectional data channel, and the simulation execution unit has a built-in dynamic correction mechanism to promptly adjust simulation paths that deviate from expectations, ensuring the real-time performance and continuity of the wargame simulation and solving the problem of insufficient virtual fidelity and simulation continuity. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of the system framework of the present invention; The meanings of the labels in the diagram are as follows: 100. Digital Twin Modeling Unit; 110. Geometric Feature Mapping Module; 120. Attribute Parameter Binding Module; 130. Status Label Synchronization Module; 140. Precision Level Division Module; 150. Precision Adjustment Trigger Module; 160. Timing Calibration Execution Module; 200. Large Model Decision Unit; 210. Multimodal Feature Analysis Module; 220. Fusion Inference Module; 230. Wargame Rule Verification Module; 300. Deduction Execution Unit; 310. Timing Advancement Module; 320. Spatial Interaction Calculation Module; 330. Adversarial Result Synthesis Module; 340. Deviation Monitoring Module; 350. Correction Strategy Generation Module; 360. Path Adjustment Execution Module; 400. Dynamic Interaction Unit; 410. Data Acquisition Module; 420. Low-Latency Transmission Module; 430. Status Synchronization Module; 500. Deduction Result Evaluation Unit; 510. Result and Benchmark Acquisition Module; 520. Multi-dimensional Evaluation Module; 530. Evaluation Result Feedback Module. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0020] like Figure 1 As shown, this embodiment provides an intelligent wargaming simulation system based on a digital twin fusion model, including: The digital twin modeling unit 100 adopts a dynamic precision adaptive modeling mechanism. It constructs a layered virtual model of the wargame scenario through the fusion of multi-source heterogeneous data. The layered virtual model includes an "adversarial entity mapping layer" and a "scenario rule embedding layer". It can automatically adjust the model precision level according to the granularity requirements of the wargame simulation and achieve sub-second dynamic synchronization between the virtual model and the wargame simulation scenario through a time-series calibration algorithm. In this embodiment, the digital twin modeling unit 100 includes a geometric feature mapping module 110, an attribute parameter binding module 120, and a status label synchronization module 130, wherein: The geometric feature mapping module 110 is based on the 3D scanning data and parametric modeling rules of the wargame adversarial entities. It adopts multi-detail level model loading technology to generate and store the 3D contour data and spatial coordinates of the wargame adversarial entities, and supports dynamic switching of precision from millimeter level to meter level. Specifically, a 3D laser scanner is first used to perform a full-angle scan of the wargame entities (such as combat equipment models, terrain sand table components, etc.) to acquire point cloud data. Then, according to parametric modeling rules (such as CAD-based parametric modeling standards), the point cloud data is converted into 3D model files (such as FBX format files) that support multi-level detail technology. The geometric feature mapping module 110, with the help of multi-level detail model loading technology, can quickly switch between millimeter-level, centimeter-level, and meter-level precision model instances according to the rendering requirements and computing resource status of the wargame simulation. When the wargame entity is in the center of the simulation interface and needs to be displayed in detail, a millimeter-level precision model is loaded; when the entity is at the edge of the field of view or when multiple entities need to be rendered in batches, it automatically switches to a meter-level precision model to balance the rendering effect and system performance.

[0021] The attribute parameter binding module 120 is based on the wargame simulation rule base and entity type configuration table. It adopts parameter dynamic association technology to bind the inherent performance parameters of wargame adversaries to entity types, and automatically updates performance parameters when the state of wargame adversaries changes, such as damage or resource consumption. Specifically, the attribute parameter binding module 120 pre-establishes a wargaming rule base (storing rules such as equipment range, movement speed, and damage threshold) and an entity type configuration table (defining attribute templates for different types of entities). After the geometric feature mapping module 110 generates the entity's 3D model, the attribute parameter binding module 120 uses parameter dynamic association technology to bind the inherent performance parameters of the corresponding entity type (retrieved from the configuration table) with the identification information of the 3D model (such as model ID), forming a "geometric feature-performance parameter" association data structure and storing it in the system database. During the simulation, if the simulation execution unit 300 outputs behavioral result data of entity damage (such as a signal that a piece of equipment has been hit), the attribute parameter binding module 120 will automatically update the performance parameters bound to the entity's 3D model according to the corresponding rules of "damage and performance degradation" in the wargaming rule base (such as proportionally reducing movement speed and shortening range), and synchronize the updated parameters to the status label synchronization module 130 and the precision adjustment trigger module 150.

[0022] Furthermore, the wargaming rule base data originates from publicly available military standards, equipment technical manuals, and authoritative wargaming datasets (such as the 2024 RAND Corporation's "Information Wargaming: Rulebook (2nd Edition)"). It is categorized by entity interaction, environmental impact, and temporal progression, and converted into a machine-readable format, verified and corrected through real-world wargaming cases. The entity type configuration table constructs a "major category-intermediate category-subcategory" system based on equipment classification standards, publicly available parameters, and general templates, binding inherent entity performance, state associations, and other parameters, and supports dynamic updates. Both are built upon publicly available data.

[0023] The state tag synchronization module 130 uses the entity behavior result data output by the simulation execution unit 300 and adopts real-time state triggering technology to mark the current state of the wargame adversary entities, with the state switching delay controlled in the sub-hundred millisecond level.

[0024] In this embodiment, the digital twin modeling unit 100 further includes a precision level division module 140, a precision adjustment triggering module 150, and a timing calibration execution module 160, wherein: The precision level division module 140 divides the model precision into three levels based on the interaction frequency and spatial scale of the wargame adversarial entities and adopts the distance-precision mapping rule. The first level of precision (millimeter level) is suitable for interaction scenarios where the distance between wargame adversarial entities is within a short range; the second level of precision (centimeter level) is suitable for scenarios where the distance between wargame adversarial entities is within a medium range; and the third level of precision (meter level) is suitable for scenarios where the distance between wargame adversarial entities is within a long range. Specifically, during the system initialization phase, based on typical wargame scenarios, a "distance-precision mapping rule" is predefined: the distance range between wargame entities is divided into short distance (e.g., 0-10 meters), medium distance (e.g., 10-50 meters), and long distance (e.g., over 50 meters), with corresponding levels of precision: Level 1 (millimeter-level), Level 2 (centimeter-level), and Level 3 (meter-level). During the simulation, the precision level division module 140 continuously obtains the real-time position information of each entity from the status tag synchronization module 130, calculates the relative distance between entities, determines the current precision level to be used, and sends the precision level information to the precision adjustment trigger module 150 and the geometric feature mapping module 110.

[0025] The precision adjustment trigger module 150 is based on the relative motion data of the wargame adversaries and uses threshold monitoring technology. When the relative motion state between the wargame adversaries meets the preset trigger conditions, the precision is automatically upgraded; when there is no interaction between the wargame adversaries for a preset duration, the precision is automatically downgraded. Specifically, the precision adjustment triggering module 150 obtains the relative motion data (such as relative speed, acceleration, and changes in motion direction) of the wargame adversaries from the status tag synchronization module 130, and combines it with the current precision level information sent by the precision level division module 140 to determine whether to trigger precision adjustment through threshold monitoring technology. For example, when two entities that were originally at a "far distance (level 3 precision)" move relative to each other, causing the distance to quickly shrink to the medium distance range and the relative speed to exceed a preset threshold (such as 5 meters / second), the precision upgrade (switching from level 3 precision to level 2 precision) is automatically triggered. If the wargame entities in a certain area do not interact for 5 minutes, the precision adjustment trigger module 150 will automatically trigger the precision downgrade to reduce the system's computational resource investment in the model of that area. Each time a precision adjustment is triggered, the precision adjustment triggering module 150 sends a precision adjustment command to the geometric feature mapping module 110, driving it to switch the model precision.

[0026] The timing calibration execution module 160 collects data from the physical scene with timestamps and uses time deviation correction technology. It calculates the time deviation between the virtual model and the physical scene through a timing calibration algorithm. If the time deviation exceeds the preset acceptable range, it corrects the state parameters of the virtual model through linear interpolation to ensure the synchronization between the virtual model and the war game scenario.

[0027] Specifically, the timing calibration execution module 160 continuously acquires timestamped physical scene data (such as time synchronization signals of the real wargaming sandbox collected by external sensors, and time stamps of entity position changes) through a dedicated data acquisition interface; simultaneously, it acquires time-related data of the virtual model from various modules within the digital twin modeling unit (such as the model state update time of the geometric feature mapping module 110 and the state mark time of the state label synchronization module 130). Subsequently, a time deviation correction technique is used to calculate the time deviation between the virtual model and the physical scene through a timing calibration algorithm (such as a time deviation fitting algorithm based on the least squares method). If the time deviation (such as the virtual model time being 200 milliseconds slower than the physical scene) exceeds a preset acceptable range (such as ±100 milliseconds), linear interpolation is used to correct the state parameters of the virtual model (such as entity position, velocity, and state label timestamps) to keep the time progression of the virtual model synchronized with the physical scene, ensuring the timing consistency between the virtual model and the real scene in the wargaming exercise.

[0028] The large model decision unit 200 receives the wargame scenario state data output by the digital twin modeling unit 100. It employs a multimodal feature fusion and wargame rule constraint embedding mechanism to fuse the spatial adversarial features and temporal decision features of the wargame scenario. It also embeds the performance boundaries of the wargame adversarial entities and scenario rules as constraints into the pre-trained model to generate adversarial decision parameters that conform to the wargame deduction logic. The large model decision unit 200 includes a multimodal feature parsing module 210, a fusion reasoning module 220, and a wargame rule verification module 230. In this embodiment, the multimodal feature parsing module 210 achieves structured extraction of wargame scene features through feature quantization and standardization processing, specifically including: For the spatial adversarial features in the wargame scenario state data, a coordinate system transformation method is used to convert the absolute coordinates of adversarial entities into quantitative parameters such as relative distance, azimuth angle, and terrain occlusion coefficient; For time-series decision characteristics, the historical action sequence of adversarial entities is extracted by using a sliding window to extract time-series parameters such as action interval and strategy switching frequency; For rule triggering features, convert the rule conditions such as resource threshold and round progress into binary feature codes; (1 if satisfied, 0 if not satisfied); The spatial adversarial features, temporal decision features, and rule-triggered feature parameters are normalized to a unified data dimension to form a multimodal fusion feature vector.

[0029] Specifically, the multimodal feature parsing module 210 first receives the wargame scene state data (including entity coordinates, historical action records, resource reserves, round progress, etc.) output by the digital twin modeling unit 100, and then performs structured extraction of wargame scene features through feature quantization and standardization processing, specifically including: The transformation method of "WGS84 geographic coordinate system to local Cartesian coordinate system" is adopted to convert the absolute coordinates (longitude, latitude, elevation) of entities into local coordinates (x, y, z) with the center point of the simulation scene as the origin. The relative distance between entities is calculated by Euclidean distance formula, and the azimuth angle is solved by trigonometric functions (such as arctan2). At the same time, based on the 3D terrain model generated by the digital twin modeling unit 100, the terrain occlusion between two entities is determined by ray tracing to calculate the terrain occlusion coefficient (1 for no occlusion, 0 for complete occlusion, and partial occlusion is calculated according to the proportion of occlusion area), thus completing the quantification of spatial adversarial features. The sliding window size is determined according to the simulation type (a 10-second sliding window is used when the time step is 1 second in real-time simulation; a window period of 3 rounds is used in round-based simulation). The historical action sequence of entities within the window is extracted, and the time difference or round difference between two adjacent actions is extracted as the action interval. The ratio of the number of strategy changes within the window to the total duration of the window is calculated as the strategy switching frequency, forming a time-series decision feature parameter matrix. The resource thresholds (such as minimum tank fuel reserves and remaining ammunition) and round progress thresholds (such as the progress node of triggering the reinforcement rule in the 8th round of a round-based simulation) of the current scenario are retrieved from the wargame rule base. The cases where "the current resource value of the entity is greater than or equal to the resource threshold" and "the current round number is greater than or equal to the round progress threshold" are judged as satisfying the rule conditions (encoded as 1), and the other cases are judged as not satisfying the conditions (encoded as 0). A binary feature code of the rule trigger feature is generated. Furthermore, the min-max normalization method is adopted, which subtracts the minimum value of the feature among all entities from the actual value of the feature parameter, and then divides by the difference between the maximum and minimum values ​​of the feature. This maps the spatial adversarial feature, temporal decision feature, and rule triggering feature parameters to the interval between 0 and 1, and concatenates them to form a multimodal fusion feature vector (the vector dimension is the number of entities multiplied by the total feature dimension. For example, when the spatial adversarial feature has 3 dimensions, the temporal decision feature has 2 dimensions, and the rule triggering feature has 3 dimensions, the total feature dimension is 8), and transmits it to the fusion inference module 220 in real time.

[0030] In this embodiment, the fusion reasoning module 220 generates an initial adversarial decision scheme through feature weighted fusion and knowledge graph reasoning, specifically including: An attention mechanism is used to dynamically weight the fused feature vector output by the multimodal feature parsing module 210, wherein the spatial feature weight increases linearly with the entity interaction intensity, and the temporal feature weight increases exponentially with the action coherence. A wargame adversarial knowledge graph is constructed based on a pre-trained model. The graph nodes contain entity types, action strategies, and rule clauses. The optimal decision logic for matching historical adversarial cases is calculated by the cosine similarity of the nodes. Based on the reasoning results, an initial adversarial decision-making scheme is output, which includes action priority, resource allocation ratio, and strategy alternative sequence. The scheme is stored in a structured data format, and the data fields include decision ID, target entity ID, and execution time sequence label.

[0031] Specifically, in this embodiment, the fusion reasoning module 220 takes the multimodal fusion feature vector output by the multimodal feature parsing module 210 as input, and generates an initial adversarial decision scheme through feature weighted fusion and knowledge graph reasoning, specifically including: First, the interaction strength of entities is quantified (represented by the reciprocal of the relative distance between entities, i.e., the smaller the relative distance, the greater the interaction strength). This makes the weights of spatial features increase linearly with the increase of interaction strength. At the same time, the action coherence is obtained by calculating the ratio of the number of actions that use the same strategy consecutively within the sliding window to the total number of actions within the window. This makes the weights of temporal features increase exponentially with the increase of coherence. This completes the dynamic weighting of the multimodal fusion feature vector. A wargame adversarial knowledge graph is constructed based on a "military domain BERT pre-trained model" (fine-tuned and optimized using publicly available wargame simulation text datasets). Nodes in the graph include "entity types" (such as main battle tanks, fighter jets, infantry fighting vehicles, etc.), "action strategies" (such as frontal assault, flanking maneuver, defensive standby, etc.), and "rules" (such as "armored entities prioritize attacking enemy armored targets" and "air entities are prohibited from entering air defense fire coverage areas," etc.). Edges between nodes represent relationships (e.g., the "main battle tank" node and the "frontal assault" node are connected by an edge indicating "applicable strategy"). By calculating the cosine similarity between the current weighted feature vector and the feature vectors of historical adversarial cases in the knowledge graph (i.e., calculating the dot product of the two vectors and then dividing by the product of the magnitudes of the two vectors), the decision logic corresponding to the historical case with the highest similarity is selected as the optimal decision logic. The initial confrontation decision plan is output based on the optimal decision logic. The plan includes action priority (e.g., "Tank 1 should attack enemy Tank 2 first, and then perform defensive tasks"), resource allocation ratio (e.g., "Allocate 30% of the ammunition reserve to Tank 1 and 25% of the ammunition reserve to Tank 2"), and strategy alternative sequence (e.g., "Prioritize the flanking strategy, switch to the frontal assault strategy if the terrain is blocked, and switch to the defensive standby strategy if the frontal assault is blocked"). The plan is stored in JSON structured format, and the data fields include decision ID, target entity ID, and execution time sequence label, and is synchronously transmitted to the wargame rule verification module 230.

[0032] It is understandable that the "military domain BERT pre-trained model" in this embodiment is based on the general BERT architecture and is trained in two stages. The pre-training stage uses publicly available weapon and equipment documents from the Global Times military channel, wargaming specification texts from the NATO STANAG protocol, and operational simulation papers from the journal *Military Operations Research and Systems Engineering* as corpora. Unsupervised learning is used to enable the model to master military terminology and tactical logic. Simultaneously, the fine-tuning stage uses two types of publicly available datasets: one is wargaming training cases published by the National Defense University, containing tens of thousands of verified scenarios, situation descriptions, and decision records; the other is structured samples formed by extracting "rule-action-outcome" triples from GJB military standards and publicly available cases. Fine-tuning employs cross-validation to optimize parameters, and the model's decision logic is entirely generated through data induction. Meanwhile, fine-tuning of the BERT pre-trained model in the military field was carried out based on a publicly available wargaming text dataset of 100,000 texts. This dataset covers typical wargaming scenarios such as tactical commands, equipment interactions, and terrain effects, with a uniform character length of 512 characters per sample. A cosine annealing learning rate strategy was used during fine-tuning, with an initial learning rate of... The learning rate decays to its current value every 20 iterations. The total number of iterations is 100, the batch size is set to 32, the optimizer is AdamW, and the weight decay coefficient is [value missing]. .

[0033] Furthermore, the interaction strength of entities is represented by the reciprocal of the relative distance between the two entities. To avoid the interaction strength becoming unbounded when the relative distance approaches 0, it is constrained based on mathematical normalization rules: Let the relative distance be... (and (Interaction strength is considered to be 0 when the distance exceeds a certain limit; meters), defining normalized interaction strength. The calculation formula is (in Rice, if Then force take ,at this time ),make Constrained within the (0,1] interval, this process is based on the monotonicity and boundedness of mathematical functions, ensuring the stability of spatial feature weight calculation.

[0034] In this embodiment, the wargame rule verification module 230 ensures decision compliance through constraint matching and gradient correction, specifically including: Retrieve constraints that match the current scenario from the wargame rule base, including the performance boundaries of wargame adversaries and scenario rules; The parameters and constraints of the initial adversarial decision-making scheme are compared item by item to determine whether there are any violations such as parameter out-of-bounds or rule conflicts. The gradient adjustment method is used to correct violations. For performance boundary violations, the parameters are truncated according to the maximum allowable value. For scenario rule violations, resources are reallocated and the action sequence is adjusted according to the rule priority. After correction, the output adversarial decision parameters conform to the wargaming simulation logic.

[0035] The specific wargame rule verification module 230 ensures decision compliance through constraint matching and gradient correction, specifically including: Based on the current scenario labels (such as "urban street fighting scenario", "high-altitude mountain warfare scenario", "maritime collaborative confrontation scenario", etc.), the constraints matching the current scenario are retrieved from the wargame rule base. Among them, the performance boundaries of the wargame adversaries include "maximum ammunition capacity of main battle tanks: 12 rounds", "maximum combat radius of fighter jets: 800 kilometers", "maximum speed of infantry fighting vehicles: 60 kilometers per hour", etc. The scenario rules include "low-altitude assault by air units is prohibited in urban street fighting scenarios", "tank movement speed is calculated at 80% of the baseline speed in high-altitude mountain warfare scenarios", "destroyers must maintain a coordination distance of 5-10 kilometers with frigates in maritime collaborative confrontation scenarios", etc. Each parameter in the initial adversarial decision-making scheme is compared with the retrieved constraints to determine whether there are any violations such as parameter over-limit (e.g., "the ammunition allocation of tank No. 1 in the initial scheme is 4 rounds, which exceeds its maximum ammunition carrying capacity of 3 rounds") or rule conflict (e.g., "the fighter jet performs a low-altitude assault mission in the initial scheme, which conflicts with the prohibition rule of urban street fighting scenario"). The specific type of violation (performance boundary violation or scenario rule violation) and related parameters are recorded. For violations, a gradient adjustment method is used for correction. For performance boundary violations, the violation parameters are truncated according to the maximum allowable value of the entity's performance boundary (e.g., adjusting the ammunition allocation of Tank No. 1 from 4 rounds to its maximum allowable carrying capacity of 3 rounds). For scenario rule violations, they are sorted according to rule priority (e.g., scenario-specific rules have higher priority than general adversarial rules), and relevant resources are reallocated and the action sequence is adjusted (e.g., "because the urban street fighting scenario rules have higher priority than the air assault strategy, the fighter jet's action strategy is changed from low-altitude assault to high-altitude support, and the 15% of fuel resources originally planned for low-altitude assault are simultaneously adjusted to the fuel configuration for high-altitude support missions"). After correction, adversarial decision parameters that conform to the wargaming simulation logic are output and transmitted in real time to the simulation execution unit 300, while also being fed back to the dynamic interaction unit 400 for state synchronization between units.

[0036] Furthermore, the wargaming rule base achieves automatic updates through a "scenario tag - rule subset" association mechanism: each simulation scenario (such as urban warfare, plateau and mountainous terrain, and naval formation) is assigned a unique tag, which is bound to scenario features such as terrain, climate, and equipment type; the rule base is divided into corresponding subsets according to the scenario tag (such as the "urban warfare" subset including building occlusion judgment and urban warfare movement speed correction rules, and the "plateau and mountainous terrain" subset including the impact of oxygen content on personnel physical strength and mountain off-road maneuver rules); when the digital twin modeling unit 100 detects changes in scenario features through terrain data analysis, environmental parameter input, and other technical means, the digital twin modeling unit 100 transmits the scenario feature change information to the simulation execution unit 300. The simulation execution unit 300 extracts the core features of the new scenario and matches the corresponding tag through its built-in rule calling function module, then closes the calling interface of the original scenario rule subset, opens the calling interface of the new scenario rule subset, and loads the new rule subset synchronously. This mechanism is based on data matching and signal transmission logic between modules, and relies on computer technology to achieve dynamic adaptation of scenarios and rules, ensuring that simulation behavior conforms to the current scenario constraints.

[0037] It should be added that, in the large model decision unit 200 of this embodiment, the multimodal feature parsing module 210, the fusion reasoning module 220, and the wargame rule verification module 230 achieve real-time interaction through the high-speed data bus interface inside the system. The multimodal feature parsing module 210 updates the multimodal fusion feature vector once every one simulation time unit (one time unit is 1 second in real-time simulation and one round in turn-based simulation). The fusion reasoning module 220 outputs new decision logic based on the updated feature vector. The wargame rule verification module 230 performs compliance verification and correction on the new decision logic in real time to ensure that the generated adversarial decision parameters are always consistent with the wargame scenario state and wargame rule requirements, and fully play its supporting role as the "intelligent decision core" of wargame simulation.

[0038] The simulation execution unit 300 adopts a wargame spatiotemporal coupling simulation mechanism. Based on the wargame virtual model of the digital twin modeling unit 100 and the adversarial decision parameters of the large model decision unit 200, it drives the adversarial process according to the wargame simulation sequence, simulates the wargame adversarial results through spatiotemporal correlation calculation, and has a built-in dynamic correction mechanism that can adjust the simulation path according to the deviation between the virtual model and the expected state of the wargame simulation. In this embodiment, the deduction execution unit 300 includes a timing progression module 310, a spatial interaction calculation module 320, and an adversarial result synthesis module 330, wherein: The timing progression module 310 is used to drive the adversarial process according to the wargame simulation timing, and to mark the turn-based simulation with turn markers. As a time series unit, in the first At the start of the round, the adversarial decision parameters output by the large model decision unit 200 are loaded; the real-time simulation is performed at time steps. As a time series unit, in The state data of the wargame adversaries is updated synchronously within the interval; among which... Indicates the start time point of the real-time simulation. Indicates starting time point As a baseline, plus a time step The final time point obtained later; Specifically, the timing progression module 310 includes operations that drive the adversarial process according to the wargame simulation timing. During implementation, the simulation type is first determined based on the simulation requirements. For turn-based simulations, use turn markers. (like (representing rounds 1, 2, and 3) is used as a time unit, in the... At the start of the round, the adversarial decision parameters (including action priority and resource allocation ratio) output by the large model decision unit 200 are retrieved through the system's internal data interface and loaded into the status control module of each wargame adversarial entity to initialize the entity's action objectives and resource quotas for this round. For real-time simulations, with time step (Based on the simulation precision setting, such as 1 second or 0.5 seconds) is the time series unit, within the time interval. Within the system, real-time status data (including location coordinates, resource consumption, and action progress) of each entity is synchronously collected through a high-frequency data sampling interface and updated to the system status database to ensure that the entity status matches the simulation time in real time. This is the starting point for the real-time simulation. The end time of this time step, the synchronization update frequency and Maintain consistency.

[0039] The spatial interaction computing module 320 is used to process the spatial interaction of wargame adversaries, and obtains entity coordinates based on the wargame virtual model of the digital twin modeling unit 100. ,in Indicates the entity number and terrain parameters. That is, the terrain slope, used to calculate the attack effect value. and degree of movement obstruction Among them, attack effect value Related firepower parameters Values, entity spacing Terrain correction factor Degree of movement obstruction Correlation of terrain type coefficients Elevation difference ; Specifically, the spatial interaction computing module 320 handles the spatial interaction operations of the wargame adversaries. During implementation, it first retrieves the entity coordinates from the geometric feature mapping module 110 of the digital twin modeling unit 100. ( (Each entity has a unique identifier, corresponding to different adversarial entities) and terrain parameters (Topographic slope, range of values) Then, the attack effect value is calculated using a formula. With the degree of movement obstruction Calculate the attack effect value The formula is ,in This represents the attack effect value (the larger the value, the stronger the attack effect). The firepower intensity parameters for the attacking entity (retrieved from the entity type configuration table, such as main battle tank main gun firepower intensity 5-10, infantry firearms firepower intensity 1-3). The protection coefficient of the defending entity (retrieved from the entity type configuration table, such as armored vehicle protection coefficient 3-8, infantry protection coefficient 0.5-1.5). The entity spacing (through the coordinates of the two entities) and Calculated using the Euclidean distance formula: ), The terrain correction coefficients (retrieved from the status tag synchronization module 130 of the digital twin modeling unit 100, e.g., 1.0 for open terrain, 0.8 for semi-occluded terrain, and 0.3 for fully occluded terrain) are calculated based on the attack effect and... Proportional to, with , Inversely proportional, then Adjustment; Furthermore, calculate the degree of movement obstruction. The formula is ,in The degree of obstruction to movement (values ​​range from 0.2 to 5.0, with higher values ​​indicating greater difficulty in movement). This is a terrain type coefficient (unitless, retrieved from the terrain parameter database, e.g., plains 0.2, hills 0.5, mountains 0.7, swamps 1.5). The elevation difference between the current position and the target position of the entity (calculated using the following formula): , For the current elevation of the entity, The target location elevation, (absolute value of elevation difference) For reference elevation difference (using the industry-standard benchmark of 10 meters), the calculation logic is that the movement obstacle is determined by the basic terrain obstacle and the additional obstacle of elevation difference. The larger the elevation difference, the stronger the obstacle.

[0040] The adversarial result synthesis module 330 integrates temporal data and spatial interaction results to generate wargame adversarial results, in which entity life values ​​are updated in relation to attack effect values. Resource changes are correlated with interaction frequency, and rule triggering results are correlated with attack effectiveness values. Threshold determination, all results are marked with rounds. Stored as time-series tags.

[0041] Specifically, for the entity's life value update dimension, the formula is as follows: ,in HPinit represents the updated health value (no units, ≥0, 0 indicates the entity is "destroyed"), while HPinit represents the initial health value at the start of the time series (retrieved from the entity type configuration table, such as 100 for armored vehicles and 50 for infantry). The attack effect value output by the spatial interaction computing module 320. This is the health decay coefficient (unitless, 0.8 for armored entities, 1.2 for infantry entities), calculated based on health decay as health decreases. Increase and decrease, decay rate from distinguish; For the dimension of resource change, the formula is: ,in For updated resource values ​​(such as ammunition "rounds" and fuel "liters"), The current resource value is the starting point of the time series. This refers to the frequency of entity interactions within this time series (such as the number of attacks and the number of moves, which are counted by the time series advancement module 310). Resource consumption per unit interaction (retrieved from the entity type configuration table, e.g., each attack consumes 1 ammunition). (Send / time), the calculation logic is that resource consumption is proportional to the interaction frequency; For the rule triggering result dimension, retrieve from the wargame rule base. Trigger threshold (e.g.) Triggering "entity devastation", Trigger "Entity Destruction"), compare with the current The value determines whether the rule is triggered; all results are marked with a round number. (Turn-based) or time-based (Real-time) is a time-series label, stored in JSON format.

[0042] In this embodiment, the dynamic correction mechanism of the deduction execution unit 300 further includes a deviation monitoring module 340, a correction strategy generation module 350, and a path adjustment execution module 360, wherein: Deviation monitoring module 340 is used to calculate the deviation between the wargame virtual model and the expected state of the simulation, and compare the entity positions output by digital twin modeling unit 100. Compared to the expected location Receive position deviation Compare resource values With expected resources Obtain resource deviation ; Specifically, during implementation, the expected state parameters of the entities (including expected locations) are first retrieved from the simulation scheme database. Expected Resources Then, the current position of the entity is obtained from the digital twin modeling unit 100. Compared with current resource values Calculate the positional deviation. The formula is ,in This is a spatial deviation measure, calculated using Euclidean distance to quantify the degree of spatial deviation. The larger the deviation, the more severe the deviation; calculate the resource deviation. The formula is ,in This is the absolute deviation of resources (units consistent with resource type). The calculation logic is to quantify resource differences through absolute values. The larger the difference, the more severe it is.

[0043] The correction strategy generation module 350 is used to formulate a simulation path adjustment plan and set a deviation threshold. ,when and Maintain the current path if necessary; otherwise, generate a location compensation value. Resource compensation value and timing compensation step size ; Specifically, first set a deviation threshold based on the entity type, then compare the current deviation with the threshold: if and If the current path is maintained, the current path will be maintained; if the threshold is exceeded, a position compensation value will be calculated. (respectively) , , Axis compensation amount (positive values ​​are adjusted in the positive direction, negative values ​​are adjusted in the negative direction), resource compensation value. (Unit and resource type are consistent; positive values ​​replenish resources, negative values ​​reduce resources), time-series compensation step size , hour, The larger the deviation, the better the corrected timing step size. The smaller the value, the higher the adjustment accuracy.

[0044] The path adjustment execution module 360 ​​is used to apply correction parameters to adjust the inference path and update the entity coordinates. Resource values ​​updated to The subsequent simulation sequence was adjusted to The adjustment results are fed back to the digital twin modeling unit 100 to update the wargame virtual model.

[0045] Specifically, update the entity's current coordinates to The current resource value has been updated to The subsequent time series unit was adjusted to The time interval becomes The adjusted entity status data is fed back to the status tag synchronization module 130 of the digital twin modeling unit 100 in real time to update the war game virtual model and ensure that the virtual model is consistent with the expected state of the simulation.

[0046] It should be added that, in the simulation execution unit 300 of this embodiment, the timing advancement module 310, the spatial interaction calculation module 320, and the adversarial result synthesis module 330 are linked in real time according to the process of "timing-driven-spatial interaction-result generation". The deviation monitoring module 340, the correction strategy generation module 350, and the path adjustment execution module 360 ​​of the dynamic correction mechanism monitor the simulation deviation and dynamically adjust it throughout the process. All modules realize data interaction through the system's high-speed data bus to ensure the continuity, accuracy and controllability of the wargame adversarial process and give full play to the role of the "process-driven core".

[0047] The dynamic interaction unit 400 is used to establish a two-way wargame data channel between the digital twin modeling unit 100, the large model decision-making unit 200 and the simulation execution unit 300, so as to realize the low-latency interactive transmission of wargame scene feature parameters, adversarial decision-making instructions and simulation status data, and ensure the real-time performance and continuity of wargame simulation.

[0048] In this embodiment, the dynamic interaction unit 400 includes a data acquisition module 410, a low-latency transmission module 420, and a status synchronization module 430, wherein: The data acquisition module 410 is used to acquire war game scene feature parameters from the digital twin modeling unit 100, acquire adversarial decision instructions from the large model decision unit 200, and acquire simulation status data from the simulation execution unit 300. Specifically, multi-source data acquisition is completed through standardized internal interfaces: scene feature parameters such as entity coordinates, terrain data, and entity status labels are collected from the digital twin modeling unit 100; adversarial decision-making instructions such as action priority, resource allocation ratio, and compliance decision parameters are collected from the large model decision-making unit 200; and inference status data such as current time sequence markers, attack effect values, entity life values, and resource values ​​are collected from the inference execution unit 300. The acquisition adopts a trigger-based strategy, automatically triggering acquisition when data in each unit is updated, with the frequency synchronized with the unit data update frequency. The acquired data is uniformly converted into a standardized format and temporarily stored in a local cache.

[0049] The low-latency transmission module 420 is used to build a two-way transmission channel for wargame data. It adopts a lightweight encapsulation transmission technology based on the UDP protocol to achieve low-latency interactive transmission of wargame scene feature parameters, adversarial decision instructions, and simulation status data. Specifically, the low-latency transmission module 420 employs a lightweight transmission mechanism based on the UDP protocol: data encapsulation uses a "protocol header (including data type, sending unit identifier, and timestamp) + compressed data body" structure to reduce transmission volume; transmission uses a priority queue, with adversarial decision-making instructions set as the highest priority to ensure critical data is transmitted first; packet loss handling is achieved through "critical data retransmission + non-critical data redundancy," triggering retransmission when critical data is missing, and adding redundancy to static data to reduce the impact of packet loss. The channel communicates via a preset port, automatically switching to a backup path when latency exceeds a threshold.

[0050] The state synchronization module 430 is used to synchronize the state of the wargame virtual scene of the digital twin modeling unit 100, the decision generation logic of the large model decision unit 200, and the simulation process of the simulation execution unit 300 based on the data transmitted by the low-latency transmission module 420, so as to ensure the real-time performance and continuity of the wargame simulation.

[0051] The specific state synchronization module 430 implements multi-unit state coordination based on the transmitted data. During implementation, it follows the process of "data parsing - conflict determination - synchronous execution": parsing the transmitted data and distributing it to the corresponding buffer according to the unit type; conflict determination is based on the data of the simulation execution unit 300 (after dynamic correction and calibration) and the results of the wargame rule verification; during the synchronous execution phase, corresponding data is pushed to each unit to update the virtual scene display, decision logic and simulation process, ensuring that the state of each unit is consistent.

[0052] It should be added that in the dynamic interaction unit 400 of this embodiment, the data acquisition module 410, the low-latency transmission module 420, and the state synchronization module 430 work together in a closed-loop process of "acquisition-transmission-synchronization". Through standardized data format and priority transmission mechanism, real-time data interaction is realized between the digital twin modeling unit 100, the large model decision-making unit 200, and the simulation execution unit 300, ensuring the consistency and coherence of the scenario, decision, and execution status during the wargaming simulation, and giving full play to the role of the "data interaction hub".

[0053] It also includes a simulation result evaluation unit 500, which is used to evaluate the effectiveness of the wargame results output by the simulation execution unit 300. This unit includes a result and benchmark acquisition module 510, a multi-dimensional evaluation module 520, and an evaluation result feedback module 530, wherein: The results and benchmark acquisition module 510 is used to acquire wargame confrontation result data from the simulation execution unit 300, acquire wargame virtual model status data corresponding to the simulation stage from the digital twin modeling unit 100, and acquire data interaction delay records corresponding to the simulation stage from the dynamic interaction unit 400 as evaluation benchmark data. Specifically, the results and benchmark acquisition module 510 collects data in a targeted manner according to the simulation stage through the system's internal data interface: the wargame combat result data collected from the simulation execution unit 300 includes entity life value update records, resource change details, rule trigger results, and corresponding time series labels; the virtual model state data corresponding to the simulation stage collected from the digital twin modeling unit 100 includes entity coordinates, terrain parameters, and entity state labels (strictly corresponding to the time series labels of the combat result data); the data interaction delay records collected from the dynamic interaction unit 400 include the timestamp difference of data transmission between units (the difference between the sending timestamp and the receiving timestamp) and the time series nodes where the delay occurred. The acquisition timing is synchronized with the result output of the simulation execution unit 300 (at the end of each round in round-based systems, and at the end of each time step in real-time systems). The acquired data is uniformly associated with time series labels and temporarily stored in the evaluation benchmark database in the form of structured tables.

[0054] The multi-dimensional evaluation module 520 is used to judge the validity of the adversarial results by comparing them with existing standards. Specifically, it includes: judging the compliance of the deduction logic by the compliance standards output by the wargame rule verification module 230; judging the synchronization accuracy of the virtual scene by the sub-second synchronization requirements of the timing calibration execution module 160; and judging the real-time performance of data interaction by the low latency requirements of the dynamic interaction unit 400. If all three items meet the standards, the result is deemed valid. If any item fails to meet the standards, it is marked as a result to be optimized and the non-compliant item is recorded. Specifically, the multi-dimensional evaluation module 520 compares the effectiveness of the adversarial results against the standard, and the evaluation is carried out according to three core indicators, including: The logical compliance assessment uses the compliance standards output by the wargame rule verification module 230 as a benchmark to compare the entity actions, resource consumption, and rule triggers in the confrontation results one by one to see if they conform to the preset rules (such as "whether the attack range of armored units is within the distance limited by the rules" and "whether the resource consumption matches the interaction frequency"). The specific content and corresponding timing of non-compliant items are recorded. Synchronization accuracy assessment: Based on the sub-second synchronization requirements of the timing calibration execution module 160, the time deviation (difference between state update time and result generation time) between the virtual model state data of the digital twin modeling unit 100 and the adversarial result data of the simulation execution unit 300 under the same timing label is calculated to determine whether the deviation is within the sub-second range. The real-time performance evaluation of the interaction is based on the low latency requirement of the dynamic interaction unit 400. The number and percentage of times the interaction latency exceeds the preset latency threshold are statistically analyzed to determine whether the data transmission meets the real-time requirements. Furthermore, the evaluation conclusion is generated according to the principle that "if all three items meet the standard, it is judged as a valid result, and if any item fails to meet the standard, it is marked as a result to be optimized". The specific parameters of the non-compliant items are recorded simultaneously (such as non-compliant rules and clauses, excessive time deviation values, and the number of transmissions with delays exceeding the threshold).

[0055] The evaluation result feedback module 530 is used to store the effective evaluation results in the war game simulation result library, and to feed back the results to be optimized and the substandard items to the large model decision unit 200 and the digital twin modeling unit 100. The information fed back to the large model decision unit 200 is used to optimize the subsequent adversarial decision parameter generation logic, and the information fed back to the digital twin modeling unit 100 is used to assist in adjusting the model accuracy level or the timing calibration frequency.

[0056] Specifically, the evaluation result feedback module 530 is used to classify and process evaluation results and implement closed-loop feedback. During implementation, results are processed separately according to their type, including: Effective evaluation results are stored in the wargaming results database in the format of "time series label + evaluation conclusion + key indicator data" as reference cases for subsequent simulations. Results to be optimized and non-compliant items are fed back in a targeted manner according to unit type—information fed back to the large model decision-making unit 200 includes the decision parameters corresponding to the non-compliant items (such as "the specific value of the resource allocation ratio exceeding the rule threshold") and an analysis of the impact of interaction delay on decision generation, used by the fusion reasoning module 220 to optimize the parameter weights in the decision logic; information fed back to the digital twin modeling unit 100 includes time series nodes with excessive synchronization deviations and corresponding virtual model state data, used to assist the geometric feature mapping module 110 in adjusting the model accuracy level (such as improving grid accuracy in areas of concentrated deviation) or the time series calibration execution module 160 in optimizing the calibration frequency. Feedback data is pushed through the low-latency transmission channel of the dynamic interaction unit 400 to ensure that relevant units obtain optimization basis in a timely manner.

[0057] It should be added that in the inference result evaluation unit 500 of this embodiment, the results, benchmark acquisition module 510, multi-dimensional evaluation module 520, and evaluation result feedback module 530 work together in accordance with the process of "data acquisition - multi-dimensional evaluation - result feedback". By associating time series labels, the evaluation and inference process are accurately matched. The targeted feedback mechanism promotes the continuous optimization of the large model decision-making unit 200 and the digital twin modeling unit 100, giving full play to the role of the "core of effect verification".

[0058] Those skilled in the art will understand that the process of implementing all or part of the steps of the above embodiments can be carried out by hardware or by a program instructing the relevant hardware.

[0059] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. An intelligent wargaming simulation system based on a digital twin fusion model, characterized in that, include: The digital twin modeling unit (100) adopts a dynamic precision adaptive modeling mechanism, and constructs a hierarchical virtual model of the war game confrontation scenario through the fusion of multi-source heterogeneous data. The hierarchical virtual model includes an "adversarial entity mapping layer" and a "scenario rule embedding layer". It can automatically adjust the model precision level according to the granularity requirements of the war game simulation, and realize sub-second dynamic synchronization between the virtual model and the war game simulation scenario through a time-series calibration algorithm. The large model decision unit (200) is used to receive the wargame scene state data output by the digital twin modeling unit (100), and to perform fusion processing on the spatial adversarial features and temporal decision features of the wargame scene by adopting a multimodal feature fusion and wargame rule constraint embedding mechanism. The performance boundary and scene rules of the wargame adversarial entities are embedded as constraints into the pre-trained model to generate adversarial decision parameters that conform to the wargame deduction logic. The large model decision unit (200) includes a multimodal feature parsing module (210), a fusion reasoning module (220), and a wargame rule verification module (230). The simulation execution unit (300) adopts a wargame spatiotemporal coupling simulation mechanism. Based on the wargame virtual model of the digital twin modeling unit (100) and the adversarial decision parameters of the large model decision unit (200), it drives the adversarial process according to the wargame simulation sequence, calculates the simulation wargame adversarial results through spatiotemporal correlation, and has a built-in dynamic correction mechanism that can adjust the simulation path according to the deviation between the virtual model and the expected state of the wargame simulation. The dynamic interaction unit (400) is used to establish a two-way wargaming data channel between the digital twin modeling unit (100), the large model decision-making unit (200) and the simulation execution unit (300), so as to realize the low-latency interactive transmission of wargaming scene feature parameters, adversarial decision-making instructions and simulation status data, and ensure the real-time performance and continuity of wargaming simulation.

2. The intelligent wargaming simulation system based on a digital twin fusion model according to claim 1, characterized in that, The digital twin modeling unit (100) includes a geometric feature mapping module (110), an attribute parameter binding module (120), and a state label synchronization module (130), wherein: The geometric feature mapping module (110) is based on the three-dimensional scanning data and parametric modeling rules of the wargame adversarial entity. It adopts multi-detail level model loading technology to generate and store the three-dimensional contour data and spatial coordinates of the wargame adversarial entity, and supports dynamic switching of precision from millimeter level to meter level. The attribute parameter binding module (120) is based on the war game simulation rule base and entity type configuration table. It adopts parameter dynamic association technology to bind the inherent performance parameters of the war game confrontation entity to the entity type, and automatically updates the performance parameters when the state of the war game confrontation entity changes. The state tag synchronization module (130) uses real-time state triggering technology to mark the current state of the wargame adversarial entities based on the entity behavior result data output by the simulation execution unit (300).

3. The intelligent wargaming simulation system based on a digital twin fusion model according to claim 2, characterized in that, The digital twin modeling unit (100) further includes a precision level division module (140), a precision adjustment trigger module (150), and a timing calibration execution module (160), wherein: The precision level division module (140) divides the model precision into three levels based on the interaction frequency and spatial scale of the wargame adversarial entities and using a distance-precision mapping rule. The precision adjustment trigger module (150) is based on the relative motion data of the wargame adversaries and uses threshold monitoring technology. When the relative motion state between the wargame adversaries meets the preset trigger conditions, it automatically triggers a precision upgrade; when there is no interaction between the wargame adversaries for a preset duration, it automatically triggers a precision downgrade. The timing calibration execution module (160) collects data from the physical scene with timestamps, uses time deviation correction technology, and calculates the time deviation between the virtual model and the physical scene through a timing calibration algorithm. If the time deviation exceeds the preset acceptable range, the state parameters of the virtual model are corrected by linear interpolation to ensure the synchronization between the virtual model and the war game scenario.

4. The intelligent war game simulation system based on a digital twin fusion model according to claim 3, characterized in that, The multimodal feature parsing module (210) achieves structured extraction of wargame scene features through feature quantization and standardization processing, specifically including: For the spatial adversarial features in the wargame scenario state data, a coordinate system transformation method is used to convert the absolute coordinates of adversarial entities into quantitative parameters such as relative distance, azimuth angle, and terrain occlusion coefficient; For time-series decision characteristics, the historical action sequence of adversarial entities is extracted by using a sliding window to extract time-series parameters such as action interval and strategy switching frequency; For rule triggering features, convert the rule conditions such as resource thresholds and round progress into binary feature codes; The spatial adversarial features, temporal decision features, and rule-triggered feature parameters are normalized to a unified data dimension to form a multimodal fusion feature vector.

5. The intelligent wargaming simulation system based on a digital twin fusion model according to claim 4, characterized in that, The fusion reasoning module (220) generates an initial adversarial decision scheme through feature weighted fusion and knowledge graph reasoning, specifically including: An attention mechanism is used to dynamically weight the fused feature vector output by the multimodal feature parsing module (210), wherein the spatial feature weight increases linearly with the entity interaction intensity, and the temporal feature weight increases exponentially with the action coherence. A wargame adversarial knowledge graph is constructed based on a pre-trained model. The graph nodes contain entity types, action strategies, and rule clauses. The optimal decision logic for matching historical adversarial cases is calculated by the cosine similarity of the nodes. Based on the reasoning results, an initial adversarial decision-making scheme is output, which includes action priority, resource allocation ratio, and strategy alternative sequence. The scheme is stored in a structured data format, and the data fields include decision ID, target entity ID, and execution time sequence label.

6. The intelligent war game simulation system based on a digital twin fusion model according to claim 5, characterized in that, The wargame rule verification module (230) ensures decision compliance through constraint matching and gradient correction, specifically including: Retrieve constraints that match the current scenario from the wargame rule base, including the performance boundaries of wargame adversaries and scenario rules; The parameters and constraints of the initial adversarial decision-making scheme are compared item by item to determine whether there are any violations such as parameter out-of-bounds or rule conflicts. The gradient adjustment method is used to correct violations. For performance boundary violations, the parameters are truncated according to the maximum allowable value. For scenario rule violations, resources are reallocated and the action sequence is adjusted according to the rule priority. After correction, the output adversarial decision parameters conform to the wargaming simulation logic.

7. The intelligent war game simulation system based on a digital twin fusion model according to claim 6, characterized in that, The simulation execution unit (300) includes a time-series advancement module (310), a spatial interaction calculation module (320), and an adversarial result synthesis module (330), wherein: The timing progression module (310) is used to drive the adversarial process according to the wargame simulation timing, and to mark the turn-based simulation with turn markers. As a time series unit, in the first At the start of the round, the adversarial decision parameters output by the large model decision unit (200) are loaded; the real-time simulation is performed with time steps. As a time series unit, in The state data of the wargame adversaries is updated synchronously within the interval; among which... Indicates the start time point of the real-time simulation. Indicates starting time point As a baseline, plus a time step The final time point obtained later; The spatial interaction computing module (320) is used to process the spatial interaction of wargame adversaries, and obtains entity coordinates based on the wargame virtual model of the digital twin modeling unit (100). ,in Indicates the entity number and terrain parameters. That is, the terrain slope, used to calculate the attack effect value. and degree of movement obstruction Among them, attack effect value Related firepower parameters Values, entity spacing Terrain correction factor Degree of movement obstruction Correlation of terrain type coefficients Elevation difference ; The adversarial result synthesis module (330) is used to integrate temporal data and spatial interaction results to generate wargame adversarial results, wherein entity life values ​​are updated in relation to attack effect values. Resource changes are correlated with interaction frequency, and rule triggering results are correlated with attack effectiveness values. Threshold determination, all results are marked with rounds. Stored as time-series tags.

8. The intelligent war game simulation system based on a digital twin fusion model according to claim 7, characterized in that, The dynamic correction mechanism of the simulation execution unit (300) further includes a deviation monitoring module (340), a correction strategy generation module (350), and a path adjustment execution module (360), wherein: The deviation monitoring module (340) is used to calculate the deviation between the wargame virtual model and the expected state of the simulation, and compare the entity positions output by the digital twin modeling unit (100). Compared to the expected location Receive position deviation Compare resource values With expected resources Obtain resource deviation ; The correction strategy generation module (350) is used to formulate a simulation path adjustment plan and set a deviation threshold. ,when and Maintain the current path if necessary; otherwise, generate a location compensation value. Resource compensation value and timing compensation step size ; The path adjustment execution module (360) is used to apply correction parameters to adjust the inference path and update the entity coordinates. Resource values ​​updated to The subsequent simulation sequence was adjusted to The adjustment results are fed back to the digital twin modeling unit (100) to update the wargame virtual model.

9. The intelligent war game simulation system based on a digital twin fusion model according to claim 8, characterized in that, The dynamic interaction unit (400) includes a data acquisition module (410), a low-latency transmission module (420), and a status synchronization module (430), wherein: The data acquisition module (410) is used to acquire war game scene feature parameters from the digital twin modeling unit (100), acquire adversarial decision instructions from the large model decision unit (200), and acquire simulation status data from the simulation execution unit (300). The low-latency transmission module (420) is used to construct a two-way transmission channel for wargame data. It adopts a lightweight encapsulation transmission technology based on the UDP protocol to realize low-latency interactive transmission of wargame scene feature parameters, adversarial decision instructions, and simulation status data. The state synchronization module (430) is used to synchronize the state of the wargame virtual scene of the digital twin modeling unit (100), the decision generation logic of the large model decision unit (200), and the simulation process of the simulation execution unit (300) according to the data transmitted by the low-latency transmission module (420), so as to ensure the real-time performance and continuity of the wargame simulation.

10. The intelligent wargaming simulation system based on a digital twin fusion model according to claim 9, characterized in that, It also includes a simulation result evaluation unit (500), which is used to evaluate the effectiveness of the wargame confrontation results output by the simulation execution unit (300), including a result and benchmark acquisition module (510), a multi-dimensional evaluation module (520), and an evaluation result feedback module (530), wherein: The result and benchmark acquisition module (510) is used to acquire war game confrontation result data from the simulation execution unit (300), acquire war game virtual model status data corresponding to the simulation stage from the digital twin modeling unit (100), and acquire data interaction delay records corresponding to the simulation stage from the dynamic interaction unit (400) as evaluation benchmark data. The multi-dimensional evaluation module (520) is used to judge the validity of the adversarial results by comparing with existing standards. Specifically, it includes: judging the compliance of the deduction logic by the compliance standard output by the war game rule verification module (230), judging the synchronization accuracy of the virtual scene by the sub-second synchronization requirement of the timing calibration execution module (160), and judging the real-time performance of data interaction by the low latency requirement of the dynamic interaction unit (400). If all three items meet the standard, the result is judged to be valid. If any item does not meet the standard, it is marked as a result to be optimized and the non-compliant item is recorded. The evaluation result feedback module (530) is used to store the effective evaluation results in the war game simulation result library, and to feed back the results to be optimized and the substandard items to the large model decision unit (200) and the digital twin modeling unit (100). The information fed back to the large model decision unit (200) is used to optimize the subsequent adversarial decision parameter generation logic, and the information fed back to the digital twin modeling unit (100) is used to assist in adjusting the model accuracy level or the timing calibration frequency.

Images