Military simulation knowledge graph generation method based on large language model

Abstract

The application relates to the technical field of knowledge graphs, and discloses a military simulation knowledge graph generation method based on a large language model, which comprises the following steps: performing transient constraint perception preprocessing on military simulation data, and extracting simulation element tuples by using a military large language model to calculate element weights; performing cross-layer entity alignment by using multi-dimensional semantic similarity, and constructing a multi-dimensional time sequence coordination graph; extracting a coordination trigger sparse attribute for a multi-element coordination edge, establishing a time sequence effectiveness function, and calculating a trigger density and concentration; then performing relationship fusion and dynamic updating, reorganizing the multi-dimensional time sequence coordination graph to determine an action chain allocation proportion; and finally generating a deduction configuration file according to the proportion to control scene loading of a simulation engine. The application can accurately retain the transient burst attribute of tactical constraints and guarantee efficient execution of a scene file.

Description

Technical Field

[0001] This invention relates to the field of knowledge graph technology, and more specifically, to a method for generating military simulation knowledge graphs based on large language models. Background Technology

[0002] With the evolution of artificial intelligence technology, knowledge graphs have been widely used in battlefield environment cognition and strategic planning. Early battlefield knowledge modeling typically employed triples to link various intelligence and entities. However, this representation method resulted in severe fragmentation of battlefield environment knowledge. While subsequent models incorporating a time dimension and dynamic evolution have emerged, they generally lack high-order interaction modeling.

[0003] In real military simulation discussions, simulation scenarios are not simply collections of entities and relationships, but rather a complex tapestry woven together by various elements such as agents, tasks, rules, action start and end states, temporal relationships, and triggering rules. Different simulation levels and modes have varying requirements for scenario scripts. During the generation process, what truly determines the executability of simulated actions is often a high-order constraint, such as: a combat unit, a task, an engagement rule, terrain concealment conditions, a weather window, an electromagnetic state, and a time slice.

[0004] Current common graph generation processes still tend to forcibly decompose the aforementioned high-order constraints into several binary edges, and then use graph neural networks for local alignment and relation fusion. This approach increases connectivity on the graph surface, but it severely fragments the execution trigger bundles that originally had to be valid within a short time window. When dynamic updates of the simulation log are subsequently superimposed, the local edge weights and relation names will change, leading to a hidden and troublesome trigger drift phenomenon: although the generated simulation scenario configuration file can be loaded normally by the simulation engine, the start time, trigger order, and environmental dependencies of the action chain have deviated significantly from the original tactical constraints. Summary of the Invention

[0005] This invention provides a method for generating military simulation knowledge graphs based on large language models, which solves the technical problems raised in the background.

[0006] This invention provides a method for generating military simulation knowledge graphs based on large language models, including:

[0007] Transient constraint perception preprocessing is performed on the collected military simulation data to extract semantic units, and the data quality score and static field ratio of the semantic units are calculated.

[0008] The semantic unit is input into the military big language model to extract simulation element tuples, and the element weights of the simulation element tuples are calculated based on the data quality scores.

[0009] Multidimensional semantic similarity is used to perform cross-layer entity alignment on the simulation element tuples to construct a multidimensional temporal collaborative graph containing multi-element collaborative edges;

[0010] For the multi-element collaborative edges, collaborative trigger sparsity attributes are extracted, and a temporal validity function is established to calculate the mean trigger density and transient trigger concentration of the multi-element collaborative edges, thereby obtaining a sparsity determination value.

[0011] The constraint-driven relationship fusion and dynamic update are performed, combining the element weights, the mean trigger density and the transient trigger concentration into a collaborative trigger vector, and combining the update gain to calculate the rule offset amplitude.

[0012] Based on the static field ratio and the collaborative triggering vector, the multidimensional time-series collaborative graph is recombined in two channels to calculate the initial state score and action chain score to determine the action chain allocation ratio.

[0013] Based on the action chain allocation ratio, a simulation configuration file is generated, and a weighted loading coverage rate is calculated based on the pattern mapping completeness to control the scene loading of the simulation engine.

[0014] The beneficial effects of this invention are as follows: by jointly extracting simulation element tuples through transient constraint perception preprocessing and large language models, the high-order collaborative relationships in complex battlefield environments are fully preserved; further, by utilizing multi-dimensional temporal collaborative graphs and the extraction of collaborative triggering sparse attributes, the transient concentrated burst patterns of multiple elements within a specific short time window are accurately captured, thereby overcoming the action chain trigger drift problem caused by excessive decomposition of binary relationships in traditional methods; combined with constraint state-driven dynamic relationship updates and dual-channel recombination mechanisms, the final generated simulation configuration file can accurately reproduce the initial state of the battlefield and ensure the rigorous temporal dependence of the action chain, greatly improving the accuracy of scenario loading in the simulation engine and the reliability of campaign-level simulations. Attached Figure Description

[0015] Figure 1 This is a flowchart of the military simulation knowledge graph generation method based on a large language model according to the present invention. Detailed Implementation

[0016] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed only to enable those skilled in the art to better understand and implement the subject matter described herein, and changes may be made to the function and arrangement of the elements discussed without departing from the scope of this specification. Various processes or components may be omitted, substituted, or added as needed in the examples. Furthermore, features described in some examples may be combined in other examples.

[0017] like Figure 1As shown, the method for generating military simulation knowledge graphs based on large language models includes:

[0018] Transient constraint perception preprocessing is performed on the collected military simulation data to extract semantic units, and the data quality score and static field ratio of the semantic units are calculated.

[0019] The semantic unit is input into the military big language model to extract simulation element tuples, and the element weights of the simulation element tuples are calculated based on the data quality scores.

[0020] Multidimensional semantic similarity is used to perform cross-layer entity alignment on the simulation element tuples to construct a multidimensional temporal collaborative graph containing multi-element collaborative edges;

[0021] For the multi-element collaborative edges, collaborative trigger sparsity attributes are extracted, and a temporal validity function is established to calculate the mean trigger density and transient trigger concentration of the multi-element collaborative edges, thereby obtaining a sparsity determination value.

[0022] The constraint-driven relationship fusion and dynamic update are performed, combining the element weights, the mean trigger density and the transient trigger concentration into a collaborative trigger vector, and combining the update gain to calculate the rule offset amplitude.

[0023] Based on the static field ratio and the collaborative triggering vector, the multidimensional time-series collaborative graph is recombined in two channels to calculate the initial state score and action chain score to determine the action chain allocation ratio.

[0024] Based on the action chain allocation ratio, a simulation configuration file is generated, and a weighted loading coverage rate is calculated based on the pattern mapping completeness to control the scene loading of the simulation engine.

[0025] Preferably, the collected military simulation data undergoes transient constraint perception preprocessing to extract semantic units, and the data quality score and static field ratio of the semantic units are calculated, including:

[0026] The military simulation data is segmented into semantic units based on a preset data source template;

[0027] The data quality score and the proportion of the static field are calculated according to the following formula:

[0028]

[0029]

[0030]

[0031]

[0032] In the formula, This indicates the data quality score. This indicates the percentage of the static field. This represents the field completeness ratio. Indicates the number of valid fields retained. Indicates the number of template baseline fields. Indicates the time decay exponent. This represents an exponential function with the natural constant as its base. This indicates the relative time aging amount of the scene anchor point. Indicates the reference scenario period. This indicates the reliability value of the source. Indicates the first preset weight. This indicates the second preset weight. This represents the third preset weight, and the sum of the first preset weight, the second preset weight, and the third preset weight is one. Indicates the number of static fields. Indicates the number of dynamic fields.

[0033] Military simulation data is the raw data set used to construct a military simulation knowledge graph. It includes entity state data, mission script data, engagement rule data, environmental data, and simulation log data. It can be collected through simulation engine log acquisition interfaces, database extraction interfaces, message bus subscription interfaces, situational data exchange files, and rule base export interfaces.

[0034] The data source template is a structured template used to specify the field names, field types, field order, required attributes, static and dynamic attributes, and time-location mapping relationships of military simulation data from different sources. It is divided into five categories: entity templates, task templates, rule templates, environment templates, and log templates, with each category containing 10 to 30 standard fields. This facilitates coverage of common military simulation input sources.

[0035] A semantic unit is the smallest semantic encapsulation fragment formed after transient constraint perception preprocessing, used to preserve local complete information about entities, tasks, rules, environment, and time constraints in a unified context.

[0036] The number of valid fields retained is the total number of valid fields that are still retained after semantic units have undergone missing data removal, format validation, conflict resolution, and field deduplication.

[0037] The number of baseline fields in a template refers to the total number of fields defined as baseline fields in the data source template. It represents the field size that a certain semantic unit should have in its complete state. The preferred number of fields for a task template is 10 to 18, for a rule template it is 8 to 15, for an environment template it is 10 to 16, and for a log template it is 12 to 30. This range covers the typical set of military simulation fields while avoiding distortion of the field completeness ratio due to overly broad templates.

[0038] The field completeness ratio is a normalized ratio of the number of valid fields retained relative to the number of baseline fields in the template, used to indicate the degree of completeness of semantic unit fields.

[0039] The relative scene anchor time aging is a measure of the time difference between the event time corresponding to the semantic unit and the scene anchor time. When the event occurs before the anchor, the absolute value of the time difference is taken. When the event occurs after the anchor, its value is directly taken as zero. It is used to indicate the newness of the semantic unit relative to the current scene context.

[0040] The reference scenario period is a standard period used to uniformly measure the time scale of a scenario, and is used to normalize the time aging process. A value of 600 seconds is preferred. This range can cover tactical-level transient trigger windows and short-to-medium-range action chain cycles, while 600 seconds balances the sensitivity of local events with the overall stability of the scenario.

[0041] The time decay index is a timeliness reduction factor calculated based on the ratio of the relative time aging of the scene anchor point to the reference scene period, and is used to reflect the time freshness of semantic units.

[0042] The source reliability value is a credibility indicator that normalizes the authority of the semantic unit source system, the stability of the data acquisition link, the data verification pass rate, and the historical error level. It can be obtained through data source registration ledgers, trusted source scoring tables, historical quality inspection records, link stability monitoring logs, and manual verification results.

[0043] The first preset weight is a weight parameter used to measure the impact of the field completeness ratio on the data quality score. A value of 0.4 is preferred. Field completeness directly determines whether subsequent feature extraction can be performed, therefore it is given the highest priority.

[0044] The second preset weight is a weighting parameter used to measure the degree of influence of the source reliability value on the data quality score. A value of 0.3 is preferred. Source reliability directly affects the credibility of facts, but is usually slightly less significant than the impact of field completeness on extraction stability.

[0045] The third preset weight is a weighting parameter used to measure the influence of the time decay index on the data quality score. A value of 0.3 is preferred. While time freshness significantly affects the effectiveness of tactical scenarios, its influence is roughly equivalent to that of source reliability in most simulation scenarios.

[0046] The data quality score is a comprehensive quality evaluation value obtained by weighting and fusing the field completeness ratio, source reliability value, and time decay index, and is used to measure the overall usability of semantic units.

[0047] The number of static fields is the total number of fields in a semantic unit that remain unchanged or change very little within a reference scenario period, such as entity type, platform category, deployment region, and fixed rule number.

[0048] The number of dynamic fields is the total number of fields in a semantic unit that changes over time, due to events, or changes in state, such as location, speed, threat level, fire control status, and available windows.

[0049] The static field ratio is the ratio of the number of static fields to the total number of static fields and dynamic fields, used to indicate the strength of the static structural features of a semantic unit.

[0050] In detail, the data source templates are constructed as follows: First, the input data is divided into five categories according to the data source and business function: entity templates, task templates, rule templates, environment templates, and log templates. Then, for each type of template, the following fields are defined in sequence: field name, field type, whether the field is required, default value, static / dynamic labels, time anchor field, location field, and source identifier field. For example, entity templates can set entity number, platform type, group, current location, speed, and status fields, while task templates can set task number, execution object, target effect, start time, and end time fields. After the field definitions are completed, each type of template is stored in the template library and a template version number is generated to ensure that subsequent semantic unit segmentation and field completeness statistics have a unified basis.

[0051] In detail, the method for determining the segmentation boundary of semantic units is as follows: first, the task number, event number, or rule trigger number is used as the first aggregation key; then, the entity identifier, temporal proximity window, and spatial proximity window are used as the second aggregation key; finally, the consistency of environmental conditions and target effect is used as the verification key. When multiple records simultaneously meet the set threshold on the above keys, they are merged into the same semantic unit; otherwise, they are split into different semantic units. For example, when the same formation performs interception against the same target within a 300-second window and shares the same rule number, they can be merged into one semantic unit, while two actions initiated against different targets should be split.

[0052] In detail, the method for determining the validity of retained fields is as follows: First, remove null fields, illegal format fields, duplicate fields, and fields that do not match the template type; second, perform legal value validation on enumerated fields, perform time format and time zone consistency validation on time fields, and perform coordinate system consistency validation on coordinate fields; third, retain conflicting fields in the following order: priority of higher source reliability value, priority of latest time, and priority of manual verification; for example, fields whose task end time is earlier than their start time should be directly judged as invalid, and when two sources provide different target numbers, the one with higher reliability and more recent time update should be retained first.

[0053] In detail, the calculation method for the relative scene anchor time aging is as follows: First, determine the scene anchor time, prioritizing one of the following: the scene start time of the current simulation round, the center time of the current batch log time window, or the task scheduling baseline time; then calculate the difference between the scene anchor time and the corresponding event time. If the difference is negative, it is forcibly truncated to zero; finally, all times are uniformly converted to seconds as the calculation unit. For example, if the scene anchor time is 12:00:00 and the event time of a certain semantic unit is 12:03:20, then the relative scene anchor time aging is 200 seconds.

[0054] In detail, the reference scenario period is determined as follows: it is determined jointly based on the simulation task level, the duration of typical actions, and the log sampling period; when it is a tactical-level short-term action simulation, the reference scenario period is preferably set to 300 to 900 seconds; when it is a medium-range continuous collaborative simulation, the reference scenario period is preferably set to 900 to 1800 seconds; for example, 600 seconds can be used when modeling air defense interception actions, and 1200 seconds can be used when modeling multi-stage reconnaissance and strike collaborative processes; once selected, it should remain unchanged in the same scenario batch.

[0055] In detail, the method for obtaining and normalizing the source reliability value is as follows: First, a reliability scoring table is established for each data source. The scoring items include the source authority level, historical error rate, link packet loss rate, verification pass rate, and manual confirmation status. Then, each scoring item is normalized to the range of 0 to 1 and summed according to weight to obtain the source reliability value. For example, the task script library can be set to 0.95 due to the stability of the source and sufficient manual review, while real-time external environment data can be set to 0.72 due to link fluctuations. If the same source experiences continuous anomalies, its reliability can be dynamically adjusted down according to the number of anomalies.

[0056] In detail, the setting method for the first, second, and third preset weights is as follows: First, based on the impact of the completeness of historical sample statistical fields, the reliability of the source, and the timeliness of the time on the final extraction success rate, then the weights are calibrated while ensuring that the sum of the three weights is 1; the preferred values ​​are 0.4, 0.3, and 0.3; for example, in scenarios with extremely high real-time requirements, the third preset weight can be increased to 0.4, and the second preset weight can be reduced to 0.2 accordingly; the adjusted weights should be verified through sample playback to confirm that they will not cause distortion of the quality score.

[0057] In detail, the classification method for static and dynamic fields is as follows: First, fields that do not change or whose changes do not affect the execution logic of the action chain within a reference scenario period are classified as static fields. Then, fields that change with the progress of time, state changes, or event triggers and affect the deduction execution are classified as dynamic fields. For boundary fields, whether they trigger subsequent rule changes should be used as the final criterion. For example, platform model and group number are static fields, while location, speed, authorization status, and weather level are dynamic fields. If a field is fixed in a single round of the scenario but may change across rounds, its performance within the current reference scenario period should still be used for classification.

[0058] Preferably, the semantic unit is input into the military large language model to extract simulation element tuples, and the element weights of the simulation element tuples are calculated based on the data quality scores, including:

[0059] The semantic units are input into the military large language model, and the simulation element tuples are extracted according to the following calculation formula, and the element weights are calculated:

[0060]

[0061]

[0062]

[0063] In the formula, This represents the simulation element tuple. Indicates the slot for combat entities. Indicates the task slot. Indicates a rule slot. Indicates the environmental slot. Indicates the target effect slot. Indicates the time window slot. Indicates the semantic phase slot. This represents the numerical value of element completeness. Indicates the slot number. Indicates the first Slot extraction confidence level for each slot. Indicates the weight of the element. This indicates the data quality score.

[0064] Simulation element tuples are structured combination of elements extracted from a single semantic unit in a military large language model. They are used to uniformly carry combat entities, tasks, rules, environment, target effects, time windows, and semantic temporal information.

[0065] Combat entity slots are structured slots used to store entity information such as force units, equipment platforms, target objects, or support nodes.

[0066] Task slots are structured slots used to store semantic information about tasks such as reconnaissance, interception, assault, suppression, maneuver, and support.

[0067] Rule slots are structured slots used to store rule information such as engagement rules, cooperation constraints, authorization conditions, trigger restrictions, and security boundaries.

[0068] An environmental slot is a structured slot used to store environmental condition information such as topography, weather, electromagnetic fields, sea conditions, and line-of-sight.

[0069] The target effect slot is a structured slot used to store target effect semantic information such as successful suppression, failed interception, enhanced interference, confirmed reconnaissance, and mission completion.

[0070] The time window slot is a structured slot used to store time range information such as task effective time, trigger time window, executable duration, and constraint expiration time.

[0071] Semantic temporal slots are structured slots used to mark semantic phase information such as the preparation phase, maneuver phase, engagement phase, evaluation phase, and convergence phase.

[0072] Slot extraction confidence score is a confidence rating given by the military big language model for each slot extraction result, used to measure the consistency between the content of the slot and the input semantic unit.

[0073] The slot number is a sequence parameter used to identify the different slot positions in the simulation element tuple. In this application, it corresponds to 7 predefined slots.

[0074] The element completeness value is an integrity evaluation value obtained by summarizing and averaging the confidence scores of all slots, and is used to represent the structural completeness of the simulation element tuple.

[0075] Feature weight is a comprehensive weight obtained by combining data quality score and feature completeness value, and is used to reflect the importance of simulation feature tuples in subsequent mapping and fusion.

[0076] In detail, the domain adaptation and invocation method of the military big language model is as follows: First, the basic model is further trained using military simulation corpus, equipment terminology list, task template corpus and rule text corpus. Then, a domain prompt template containing entity, task, rule, environment and temporal labels is constructed. When invoking, a single semantic unit is serialized into a unified input format and accompanied by template type and scene context identifier. For example, entity description, task requirements and environment fragments can be put into the input at the same time to ensure that the model extracts 7 slot contents in the same context.

[0077] In detail, the output structure of the simulation element tuple is as follows: define field names, field values, confidence levels, source fragment indexes, and missing flags for each of the 7 slots; when multiple candidate values ​​exist for a slot, sort them from highest to lowest confidence level, retain only the first candidate as the primary value, and record the remaining candidates as alternative values; for example, the combat entity slot can contain the primary entity number and alternative entity numbers. If there are no available results for a rule slot, the missing flag should be explicitly recorded instead of leaving it blank, so as to calculate the element completeness value later.

[0078] In detail, the extraction rules for the seven slots are as follows: First, establish trigger word lists, template sentences, and conflict resolution rules for the combat entity slot, mission slot, rule slot, environment slot, target effect slot, time window slot, and semantic phase slot respectively; then, perform extraction in the order of entity priority, mission association, rule constraint, environment supplementation, effect confirmation, time window positioning, and phase marking; for example, when both "target entry window" and "allow launch" descriptions appear in the same statement, they should be filled into the time window slot and rule slot respectively, and should not be mixed in the same slot.

[0079] In detail, the generation method of slot extraction confidence is as follows: first, read the original probability or score of the candidate results output by the military big data model for each slot, then use historical labeled samples to perform temperature calibration or segmented mapping calibration, and finally compress the results uniformly into the range of 0 to 1; for example, the original model output gives 0.91 for combat entity slots and 0.63 for rule slots, which can be adjusted to 0.88 and 0.67 respectively after calibration to improve the comparability between different slots.

[0080] In detail, the handling of multi-value slots and missing slots is as follows: If a slot has multiple candidate values, the result with the highest confidence is retained as the principal value and recorded in the candidate list; if a slot is completely missing, its confidence is set to 0 and the reason for the missing value is recorded; for example, when an environmental slot identifies both low clouds and strong winds, the principal environmental label can be retained according to the primary and secondary relationship and the other result can be written into the candidate list; if a semantic temporal slot cannot be identified, the position confidence of that slot is directly set to 0 and used in the calculation of the feature completeness value.

[0081] In detail, the default processing method for feature completeness values ​​is as follows: the average calculation is performed with 7 slots as the denominator, and the denominator is not reduced due to missing slots; when the confidence of individual slot positions is abnormally higher than 1 or lower than 0, it should be truncated to the 0 to 1 range before participating in the calculation; for example, if only 4 slots are successfully extracted and the confidence levels are 0.9, 0.8, 0.7 and 0.6 respectively, the confidence levels of the other 3 missing slot positions are treated as 0, and the total feature completeness value should be averaged based on 7 slots to avoid artificially inflated values ​​due to missing slots.

[0082] Preferably, cross-layer entity alignment is performed on the simulation element tuples using multi-dimensional semantic similarity to construct a multi-dimensional temporal collaboration graph containing multi-element collaboration edges, including:

[0083] The multidimensional semantic similarity is calculated according to the following formula, and the updated entity latent state vector of the multidimensional temporal co-operation graph is generated:

[0084]

[0085]

[0086] In the formula, This represents the multidimensional semantic similarity. Represents semantic weight, Represents the cosine similarity function. Represents the semantic vector of the first entity. Represents the semantic vector of the second entity. Represents topological weights. Represents the Jaccard similarity function. Represents the first entity topological descriptor. This represents the topological descriptor of the second entity. Indicates time weighting, This represents an exponential function with the natural constant as its base. Indicates the first entity's time anchor point. Indicates the time anchor point of the second entity. Indicates the reference scenario period, Indicates spatial weights, Represents the position vector of the first entity. Represents the position vector of the second entity. Indicates the reference spatial scale, This represents the updated entity's hidden state vector. This represents the hidden state vector of the first entity in the current layer. This represents the hidden state vector of the second entity in the current layer. This represents the activation function. This represents the state transformation matrix. Represents the cooperative state transformation matrix. This represents the multi-element collaborative edge. Indicates the order of the cooperative edge. Indicates the index of the first entity node. This indicates the index of the second entity node.

[0087] Multidimensional semantic similarity is a comprehensive similarity score obtained by jointly weighting the similarity of entities in four dimensions: semantics, topology, time, and space. It is used to support cross-layer entity alignment.

[0088] The semantic vector of a first entity is a fixed-dimensional numerical representation formed by encoding the name, attributes, context, and task-related information of the first entity.

[0089] The second entity semantic vector is a fixed-dimensional numerical representation formed by encoding the name, attributes, context, and task-related information of the second entity.

[0090] The first entity topological descriptor is a set of descriptions used to characterize the adjacency relationship type, neighbor role category, connection rules, and local structural features of the first entity.

[0091] The second entity topological descriptor is a set of descriptions used to characterize the second entity's adjacency relationship type, neighbor role category, connection rules, and local structural features.

[0092] The first entity's time anchor is a representative point in time when the corresponding event, state snapshot, or rule takes effect. It can be obtained through simulation log timestamps, event script start and end times, rule effective timestamps, and task scheduling timestamps.

[0093] The second entity's time anchor is a representative point in time when the corresponding event, state snapshot, or rule takes effect. It can be obtained through simulation log timestamps, event script start and end times, rule effective timestamps, and task scheduling timestamps.

[0094] The first entity location vector is a coordinate vector that numerically represents the spatial location of the first entity. It can be obtained through geographic information system coordinates, simulation entity status tables, navigation and positioning playback data, and battlefield situation messages.

[0095] The second entity location vector is a coordinate vector that numerically represents the spatial location of the second entity. It can be obtained through geographic information system coordinates, simulation entity status tables, navigation and positioning playback data, and battlefield situation messages.

[0096] Cosine similarity is a semantic similarity metric used to measure the degree of alignment between the semantic vectors of a first entity and the semantic vectors of a second entity.

[0097] Jaccard similarity is a topological similarity metric used to measure the strength of the intersection and union relationship between the topological descriptors of a first entity and the topological descriptors of a second entity.

[0098] Temporal similarity is the degree of temporal closeness calculated based on the temporal distance between the first entity's time anchor point and the second entity's time anchor point relative to the period of the reference scene.

[0099] The reference spatial scale is a baseline quantity used to normalize differences in entity positions. To strictly guarantee the dimensionless nature of the exponent term, when the numerator of the spatial similarity formula takes the form of the square of the position difference, the reference spatial scale is preferably 25 million square meters; when the numerator is changed to Euclidean distance, the reference spatial scale is preferably 5 kilometers. This value can cover the local spatial range commonly encountered in tactical platform collaboration and ensure that the exponent term maintains a stable decay.

[0100] Spatial similarity is the degree of spatial proximity calculated based on the square of the spatial distance between the first entity position vector and the second entity position vector relative to the size of the reference spatial scale, in order to ensure that the molecular form is consistent with the dimensions of the reference spatial scale used as the area benchmark.

[0101] Semantic weight is a weighting parameter used to measure the contribution of semantic similarity components to multidimensional semantic similarity. A value of 0.35 is preferred. Entity semantic consistency is usually the primary criterion for cross-layer entity alignment, therefore it is given a relatively high weight.

[0102] Topological weight is a weighting parameter used to measure the contribution of topological similarity components to multidimensional semantic similarity. A value of 0.25 is preferred. Local structural relationships are an important auxiliary basis beyond entity semantics, but should generally not exceed the entity semantics themselves.

[0103] The temporal weight is a weighting parameter used to measure the contribution of temporal similarity components to multidimensional semantic similarity. A value of 0.2 is preferred. Temporal proximity can enhance the accuracy of military event alignment, but it is typically used as reinforcing information in cross-layer entity alignment.

[0104] Spatial weight is a weighting parameter used to measure the contribution of spatial similarity components to multidimensional semantic similarity. A value of 0.2 is preferred. Spatial proximity is important in battlefield collaboration scenarios, but its role is usually at the same level as temporal similarity.

[0105] The updated entity latent state vector is the entity representation vector generated after the graph neural network completes one collaborative edge message aggregation, and is used to carry the fused entity latent state.

[0106] The hidden state vector of the first entity in the current layer is the hidden state vector of the graph neural network corresponding to the first entity node in the current layer.

[0107] The hidden state vector of the second entity in the current layer is the hidden state vector of the graph neural network corresponding to the second entity node in the current layer.

[0108] The current layer is the layer number at which the graph neural network performs message passing and state updates.

[0109] The activation function is a function parameter that performs a nonlinear mapping on the result of a linear transformation, used to enhance the expressive power of graph neural networks. A linear rectified activation function is preferred. This value is computationally stable, simple to implement, and exhibits good training performance in sparse graph aggregation scenarios.

[0110] The self-state transformation matrix is ​​a trainable matrix parameter used to map the entity's own latent state to the target feature space.

[0111] The cooperative state transformation matrix is ​​a trainable matrix parameter used to map the aggregated neighborhood information from multi-factor cooperative edges to the target feature space. In actual operation, this parameter implicitly integrates a scaling factor that is the inverse of the order of the cooperative edge minus one, in order to correct the node aggregation mean.

[0112] Multi-element collaborative edges are high-order edges that connect multiple collaborative elements or entity nodes simultaneously, used to preserve collaborative relationships that must be established together in tactical constraints.

[0113] The order of a collaborative edge is the number of entity nodes or collaborative elements contained in the same multi-element collaborative edge. When performing neighborhood information aggregation in a graph neural network to obtain the average value, the denominator of the division should be the number of actual adjacent nodes obtained by subtracting the value from the order of the collaborative edge.

[0114] The first entity node index is an index parameter used to identify the position of the first entity node in a multi-feature collaborative edge.

[0115] The second entity node index is an index parameter used to identify the position of the second entity node in a multi-feature collaborative edge.

[0116] In detail, the entity semantic vector is generated as follows: First, collect entity name, platform model, combat role, group, current mission and rule context text, and then use a domain text encoder to encode the above content into a fixed-dimensional vector; if the dimensions of vectors from different sources are inconsistent, they are projected to the same dimension through a unified linear mapping; for example, drone nodes in the mission script and reconnaissance platform entries in the log should be projected to the same dimension before comparing cosine similarity.

[0117] In detail, the entity topology descriptor is constructed as follows: First, the entity's adjacency task type, adjacency rule type, adjacency environment label, adjacency target effect label, and edge direction attribute in the current local graph are counted. Then, these labels are deduplicated and combined into a descriptor set. For example, if an entity is connected to reconnaissance tasks, authorization rules, and low cloud environment labels at the same time, its topology descriptor can be composed of these three types of labels and their relational directions, which are used for subsequent Jaccard similarity calculation.

[0118] In detail, the selection method for time anchors is as follows: prioritize the effective time of the entity or event, then the status update time, and finally the log recording time; when both start and end times exist, prioritize the start time or the center time of the time window; for example, if a task record has both a planned start time and an actual update time, and it is used to represent task triggering, the planned start time should be used as the time anchor.

[0119] In detail, the method for unifying the coordinate system and units of the position vector is as follows: First, the position data from different data sources are uniformly converted to the same two-dimensional or three-dimensional coordinate system, and then the coordinate units are unified to meters; if the original data uses latitude and longitude, then projection conversion is performed first; for example, when the battlefield situation system outputs latitude and longitude coordinates while the simulation engine uses planar metric coordinates, the coordinate projection should be completed first, and then the position vector should be generated to avoid spatial similarity distortion.

[0120] In detail, the reference spatial scale is set as follows: if the numerator of the spatial similarity formula uses the square of the position difference, then the reference spatial scale should also adopt the area dimension, preferably the square of the side length of the local tactical area; if the numerator of the spatial similarity formula uses Euclidean distance, then the reference spatial scale should adopt the length dimension, preferably the typical cooperative distance threshold; for example, if the side length of the local air defense combat zone is 5000 meters, if the square of the position difference is used, then the reference spatial scale should be 25,000,000 square meters to ensure that the dimensions of the exponent are consistent.

[0121] In detail, the weighting method for multidimensional semantic similarity is as follows: First, use manually annotated entity alignment samples as the training set, and statistically analyze the contributions of semantic similarity, topological similarity, temporal similarity, and spatial similarity to the alignment accuracy. Then, calibrate the weights under the constraint that all four weights are not less than 0 and their sum is 1. The preferred weights are 0.35, 0.25, 0.2, and 0.2. For example, if there are many scenarios with a high risk of spatial mismerging, the spatial weight can be appropriately increased while the topological weight is decreased simultaneously.

[0122] In detail, the entity alignment threshold and conflict resolution method are as follows: First, set an alignment threshold for multidimensional semantic similarity, preferably 0.75; when an entity matches multiple candidate entities and all of them are higher than the threshold, the candidate with the highest similarity and the most consistent temporal and spatial constraints is retained first; for example, when a log entity has similarities of 0.82 and 0.8 with two task layer entities, the one that is closer to its temporal anchor point and spatial location should be retained first to avoid duplicate alignment.

[0123] In detail, the graph neural network structure and training method are as follows: a 2- to 3-layer high-order graph aggregation network is preferred, and the hidden state dimension is preferably 128 or 256. During training, manually labeled entity alignment results and relationship fusion results are used as supervision targets, and cross-entropy loss or mean squared error loss is used for optimization. For example, the first layer aggregates neighborhood collaboration information, the second layer outputs the updated entity hidden state vector, and the inference stage updates layer by layer in sequence until the final multi-dimensional temporal collaboration graph representation is obtained.

[0124] Preferably, the process involves extracting collaborative trigger sparsity attributes from the multi-element collaborative edges, establishing a temporal effectiveness function to calculate the mean trigger density and transient trigger concentration of the multi-element collaborative edges, and obtaining a sparsity determination value, including:

[0125] Obtain the single-element temporal availability and element dependency weight of the simulated collaborative elements contained within the multi-element collaborative edge;

[0126] The timing effectiveness function is constructed based on the following calculation formula, and the mean trigger density, the transient trigger concentration, and the sparsity determination value are calculated:

[0127]

[0128]

[0129]

[0130]

[0131] In the formula, This represents the timing validity function. This represents the multi-element collaborative edge. This represents the simulated collaborative element within the multi-element collaborative edge. Indicates the multiplication symbol. This indicates the temporal availability of the single element. Represents the time variable. This indicates the subordinate weight of the element. This indicates the mean trigger density. Indicates the reference scenario period, To represent the definite integral symbol, This indicates the transient trigger concentration. This represents a function that seeks the maximum or extreme value. Indicates a zero-valued constant. This represents the sparsity determination value.

[0132] Collaborative triggering sparse properties are a set of temporal characteristics used to characterize the rare but potentially concentrated bursts of collaborative edges of multiple elements on the time axis.

[0133] A simulation collaboration element is a single collaboration component contained within the same multi-element collaboration edge, and can be any of the following: entity, task, rule, environmental condition, or effect constraint.

[0134] Single-element temporal availability is a numerical value representing the degree to which a simulation collaborative element is in a usable, satisfyable, or triggerable state at a specific moment. It can be obtained through entity status timeline logs, task execution timelines, rule activation records, environmental monitoring sequences, and event replay data.

[0135] Element dependency weight is a weight parameter used to measure the constraint status and influence strength of a simulation collaborative element within its multi-element collaborative edge. A value of 1.0 is preferred. This range can enhance the influence of key elements while avoiding excessive amplification or compression of values ​​caused by exponential operations.

[0136] The temporal effectiveness function is an overall temporal effectiveness function formed by multiplying the individual temporal availability of each simulation collaborative element within a multi-element collaborative edge through exponential weighting. It is used to represent the degree to which collaborative constraints are jointly valid at any given time.

[0137] A time-time variable is a continuous or discrete time variable that represents the specific time position within the domain of a time-series validity function.

[0138] Mean trigger density is the average validity of the temporal validity function over the period from time zero to the reference scenario, and is used to reflect the average density of triggerable edges as a whole.

[0139] The maximum extremum is the maximum function value that the time series validity function can reach within the time interval under consideration.

[0140] The zero-prevention constant is a very small positive parameter introduced to prevent the denominator from being zero or close to zero. A value of 0.0001 is preferred. This value effectively suppresses zero denominators and numerical explosion problems without significantly altering the normalization result.

[0141] The denominator is the sum of the mean trigger density and the zero-prevention micro constant, used for calculating the concentration of stable transient triggering.

[0142] Transient trigger concentration is the degree of concentration of the maximum extremum of the temporal validity function relative to its average validity level. It is used to reflect the strength of the explosive triggering of cooperative edges within a short time window. After extraction, a lower limit constraint is applied to it to ensure that its value is always not less than the value of one.

[0143] The difference term is the difference obtained by subtracting the mean trigger density from the value, and is used to emphasize the sparse features when the average trigger density is low.

[0144] The sparsity judgment value is a comprehensive judgment quantity composed of transient trigger concentration and difference term, used to determine whether multi-element collaborative edges have obvious sparse trigger attributes.

[0145] In detail, the method for calculating the temporal availability of a single element is as follows: First, read the entity status, rule activation, environmental changes, and task window data according to a uniform sampling step size. Then, map whether the corresponding element conditions are met at each sampling time to an availability value between a lower limit benchmark and a value of one. The lower limit benchmark is positively correlated with the total number of elements within the collaborative edge to alleviate excessive numerical decay caused by multiple product steps. If the conditions are fully met, the value is taken as one; if they are partially met, the value is mapped to the interval according to the rules. For example, if the meteorological conditions reach level one suitability, the value can be taken as 1; if they reach level two suitability, the value can be taken as 0.7; and if they are unsuitable, the value can be taken as 0.

[0146] In detail, the method for determining the subordinate weight of elements is as follows: First, based on the necessity of the elements for triggering overall coordination, they are divided into three levels: core constraint elements, important constraint elements, and auxiliary constraint elements. Then, they are assigned higher, medium, and lower weights respectively. Priority is given to making the average subordinate weight of each element within the same multi-element coordination edge 1. For example, the combat authorization rule and target entry window condition can be set to 1.3, the platform online status can be set to 1.1, and the auxiliary environment label can be set to 0.8 to reflect different constraint statuses.

[0147] In detail, the discrete implementation of the time-series availability function is as follows: the period from time zero to the reference scenario is divided into sampling points of equal step size, and then the exponent of the time-series availability of all single elements is calculated at each sampling point and multiplied together to form a discrete sequence; for example, when the sampling step size is 1 second, a 600-second reference scenario period will form 601 sampling points, and each sampling point can obtain a time-series availability function value, which is used for subsequent integral approximation and maximum extremum search.

[0148] In detail, the integral calculation method for the mean trigger density is as follows: Under discrete implementation conditions, the continuous definite integral is mapped by discrete summation. That is, the time-series effectiveness function value at each time step is weighted and accumulated, and then divided by the reference scene period to obtain the mean trigger density. For example, when the sampling step size is 1 second, the time-series effectiveness function value corresponding to each second can be summed, multiplied by 1, and then divided by the total duration of 600 seconds to obtain the average trigger density of the cooperative edge.

[0149] In detail, the method for finding the maximum extremum is as follows: first, perform a light smoothing on the discrete time series validity function sequence to suppress single-point noise, and then search for the maximum value in the entire time interval as the maximum extremum; if smoothing is not performed, at least a minimum peak width constraint should be set to avoid misjudgment of noise peaks; for example, if a certain cooperative edge has a sustained high value from the 210th second to the 220th second, then the peak value of this plateau interval should be taken as the maximum extremum, rather than an isolated outlier.

[0150] In detail, the method for determining the value of the zero-valued constant is as follows: In all dimensionless normalization formulas, a small positive number of the same order of magnitude is used, preferably 0.0001; when the input data magnitude is significantly smaller or larger, it can be adjusted within the range of 0.000001 to 0.001, but the batch in the same scenario must be consistent; for example, when the mean trigger density and the collaborative vector norm are both in the range of 0 to 1, using 0.0001 can stabilize the calculation without significantly distorting the results.

[0151] In detail, the threshold and usage of the sparsity judgment value are as follows: First, the distribution of sparse triggering and non-sparse triggering collaborative edges is statistically analyzed based on historical samples. Then, the judgment threshold is dynamically adjusted in combination with the magnitude of the maximum extreme value. Alternatively, after performing normalization mapping on the judgment result, 0.6 is preferentially taken as a fixed benchmark. When the sparsity judgment value is higher than the threshold, the edge is marked as a transient burst-type collaborative edge, and its priority is increased in subsequent dual-channel reorganization and action chain configuration. For example, when the average trigger density of an edge is only 0.15 but the transient trigger concentration is high, it is often classified as an object that needs to maintain the action chain order.

[0152] Preferably, the constraint-driven relationship fusion and dynamic update are performed, combining the element weights, the mean trigger density, and the transient trigger concentration into a collaborative trigger vector, and combining this vector with a preset update gain calculation rule to determine the offset magnitude, including:

[0153] The collaborative trigger vector is constructed according to the following calculation formula, and the fusion collaborative importance and the rule offset magnitude are calculated:

[0154]

[0155]

[0156]

[0157]

[0158]

[0159] In the formula, Indicates the normalized transient trigger concentration. This indicates the transient trigger concentration. Indicates the global maximum trigger concentration. This represents the zero-prevention micro constant. This represents the collaborative triggering vector. This represents the average weight of the factors. This indicates the mean trigger density. Indicates a measure of source diversity. Represents the matrix transpose symbol. Indicates the importance of integration and collaboration. Represents the collaborative fusion weight vector. This represents the updated collaborative vector. This indicates the update gain. Indicates the current collaborative vector. Represents the log incremental collaboration vector. This indicates the offset magnitude of the rule. The symbol for the 2-norm of a vector.

[0160] The global maximum trigger concentration is the maximum value among the transient trigger concentrations of all multi-factor collaborative edges within the current statistical scope.

[0161] Normalized transient trigger concentration is an independent metric obtained by scaling the original transient trigger concentration relative to the global maximum trigger concentration. This normalized independent metric is consistently used in all subsequent processing steps and does not overwrite the original concentration value.

[0162] The mean element weight is the average of the element weights corresponding to all simulated element tuples within the same multi-element collaborative edge.

[0163] Source diversity measure is a normalized index used to reflect the degree of dispersion and balance of data source types referenced by the same multi-factor collaborative edge.

[0164] The collaborative triggering vector is a multidimensional feature vector composed of the mean of element weights, mean triggering density, normalized transient triggering concentration, and source diversity measure.

[0165] The collaborative fusion weight vector is a weight vector used to measure the contribution of each component of the collaborative triggering vector to the importance of fusion collaboration. Preferably, the four components are 0.3, 0.25, 0.25, and 0.2 respectively. This value takes into account element quality, average triggering level, transient burst characteristics, and source dispersion, thus maintaining the stability of the fusion result.

[0166] The importance of fusion and collaboration is a comprehensive importance index obtained by linearly fusing the fusion weight vector and the collaboration trigger vector.

[0167] The current round number represents the number of the current iteration step or the current log update batch in the relationship fusion and dynamic update process.

[0168] The current collaborative vector is the collaborative trigger vector before the log increment is introduced in the current round. When a new relationship is established in the first inference round, this vector is initialized to a non-zero baseline unit state by default to avoid abnormal amplification of the offset magnitude in subsequent division operations. After that, it is updated normally according to the increment.

[0169] The log incremental co-vector is a co-feature vector extracted from newly arrived incremental log data. It can be obtained through the streaming log subscription interface, event incremental queue, time window difference comparison module, and incremental feature extraction module.

[0170] The update gain is an update coefficient used to control the fusion ratio of the current collaborative vector and the log incremental collaborative vector. A value of 0.2 is preferred. This range strikes a balance between stability and agility, avoiding excessively slow updates or oversensitivity to short-term noise.

[0171] The updated collaboration vector is the next round of collaboration trigger vector obtained by merging the current collaboration vector and the log incremental collaboration vector.

[0172] Rule offset magnitude is the relative change strength of the updated collaborative vector with respect to the current collaborative vector, used to measure whether the relation rule has drifted significantly.

[0173] In detail, the statistical range of the global maximum trigger concentration is as follows: the maximum value is first counted among all multi-element collaborative edges in the current simulation scenario, the current loading batch, or the current update window, without mixing calculations across different task batches; for example, if a scenario loading contains 300 collaborative edges, the global maximum trigger concentration should be selected from the transient trigger concentration of these 300 edges to ensure that the normalization result is consistent with the current scenario.

[0174] In detail, the aggregation range of the average element weights is as follows: within a single multi-element collaborative edge, the arithmetic mean of the weights of all simulation element tuples that participate in the edge is directly calculated; if there are elements of the same type from the same source, duplicates can be removed before averaging; for example, if an edge contains 3 entity tuples, 1 rule tuple and 1 environment tuple, then the weights of these 5 elements should be averaged to obtain the average element weight of the edge.

[0175] In detail, the calculation method for source diversity measurement is as follows: First, count the different source categories and their proportions referenced by the same multi-factor collaborative edge, and then prioritize the use of normalized information entropy to calculate the source diversity measurement; if the implementation is simplified, the ratio of the number of unique source values ​​to the total number of sources can also be used; for example, if an edge references four types of sources at the same time, namely task scripts, real-time logs, rule bases and environmental sensor data, and the proportions are close, then its source diversity measurement should be close to 1.

[0176] In detail, the setting or training method of the collaborative fusion weight vector is as follows: when the sample is sufficient, the collaborative importance of manual review is used as the supervision target to train the weight vector; when the sample is insufficient, the initial values ​​of 0.3, 0.25, 0.25 and 0.2 set manually are preferred, and the values ​​are continuously fine-tuned through playback verification; for example, if the actual test shows that the diversity of sources has a weak impact on the loading success rate, its weight can be appropriately reduced and the weight of the average weight of the elements can be increased.

[0177] In detail, the generation method of log incremental collaborative vector is as follows: First, perform element extraction, trigger density and concentration calculation in the same way as the main process on the current incremental log. Then, extract four features from it: mean element weight, mean trigger density, normalized transient trigger concentration and source diversity measure. The diversity measure is re-evaluated based on the global data source set that merges the current incremental record and historical related records. The data is then concatenated in the same order as the current collaborative vector to form the log incremental collaborative vector. For example, after adding a new fire control rule change log, the 4-dimensional incremental features of that side should be recalculated instead of directly using the original text.

[0178] In detail, the strategy for determining the update gain is as follows: when log updates are frequent, short-term situational fluctuations are significant, and a rapid response is required, a value of 0.2 to 0.3 should be prioritized; when the scenario is relatively stable and noise suppression is required, a value of 0.1 to 0.2 should be prioritized; for example, 0.25 can be used in the real-time adversarial simulation stage to quickly absorb the impact of new logs, while 0.15 can be used in the debriefing and analysis stage to maintain the stability of the overall rules.

[0179] In detail, the current round and update cycle are defined as follows: each time a log batch with a fixed time window is received or each complete incremental processing is completed, it is recorded as a round; the round length is preferably 30 seconds, 60 seconds, or 120 seconds; for example, when the update cycle is 60 seconds, 12:00 to 12:01 is the first round, and 12:01 to 12:02 is the second round; all collaborative vector updates should be performed in this unified round numbering order.

[0180] In detail, the threshold and triggering strategy for rule offset magnitude are as follows: First, the stable range of the rule is statistically analyzed based on historical scenario replays. Then, a mild offset threshold and a severe offset threshold are set, preferably 0.1 and 0.25. When the rule offset magnitude exceeds the mild threshold, only the edge weight and importance are updated. When it exceeds the severe threshold, entity alignment review, relation renaming, and partial regeneration of the configuration file are triggered. For example, when the offset magnitude of a certain edge reaches 0.3, it should be considered that the relation semantics have changed significantly.

[0181] Preferably, the multi-dimensional temporal coordination graph is recombined in a dual-channel manner based on the static field proportion and the coordination trigger vector, and the initial state score and action chain score are calculated to determine the action chain allocation ratio, including:

[0182] Obtain the number of static collaborative elements and the order of the collaborative edge within the multi-element collaborative edge, and calculate the initial state score, the action chain score, and the action chain allocation ratio according to the following formula:

[0183]

[0184]

[0185]

[0186]

[0187]

[0188] In the formula, This indicates the percentage of the static field. This indicates the number of static collaborative elements. Indicates the order of the cooperative edge. Indicates the percentage of dynamic fields. This represents the initial state score. Indicates the first initial weight. Indicates the second initial weight. This represents the mean trigger density. Indicates the third initial weight. This represents the normalized transient trigger concentration. This represents the score of the action chain. Indicates the weight of the first action. Indicates the weight of the second action. Indicates the weight of the third action. Indicates the importance of the fusion and collaboration. This indicates the allocation ratio of the action chain. This represents the zero-prevention micro constant.

[0189] The number of static collaborative elements is the number of collaborative elements in the same multi-element collaborative edge that are determined to be dominated by static attributes.

[0190] The proportion of dynamic fields is the proportion of dynamic features obtained by subtracting the proportion of static fields from the value, and is used to represent the strength of the dynamics of collaborative edges.

[0191] The first initial weight is a weight parameter used to measure the influence of the proportion of static fields on the initial state score. A value of 0.45 is preferred. The initial state channel primarily serves the construction of the scene's static base; therefore, the role of the proportion of static attributes should be emphasized.

[0192] The second initial weight is a weighting parameter used to measure the influence of the mean trigger density on the initial state score. It is preferably 0.3. The mean trigger density reflects the persistence of effective constraints and plays a secondary important role in initial state stability modeling.

[0193] The third initial weight is a weight parameter used to measure the influence of the difference between the numerical value and the normalized transient trigger concentration on the initial state score. A value of 0.25 is preferred. This parameter reflects the non-explosive characteristic and has an auxiliary enhancing effect on the initial state channel; therefore, its value is preferably slightly lower than the first two values.

[0194] The initial state score is a comprehensive score used to characterize whether a certain multi-factor collaborative edge is more suitable to enter the initial state channel.

[0195] The first action weight is a weighting parameter used to measure the influence of the proportion of dynamic fields on the action chain score. A value of 0.25 is preferred. While the proportion of dynamic fields is important for action chain identification, it still needs to be considered in conjunction with the importance of transient triggering and coordination.

[0196] The second action weight is a weighting parameter used to measure the influence of normalized transient trigger concentration on the action chain score. A value of 0.35 is preferred. Transient concentrated burst characteristics are an important basis for identifying action chain channels, and therefore should be given a higher weight.

[0197] The third action weight is a weighted parameter used to measure the degree of influence of the importance of integration and synergy on the action chain score. A value of 0.4 is preferred. The importance of integration and synergy comprehensively reflects the quality of elements, trigger density, and source diversity, and best represents the value of the action chain; therefore, it should be given the highest priority.

[0198] The action chain score is a comprehensive score used to characterize whether a multi-element collaborative edge is more suitable to enter the action chain channel.

[0199] The action chain allocation ratio is the proportion of the action chain score in the total initial state score and the action chain score. It is used to control the intensity of knowledge allocation to the action chain profile.

[0200] In detail, the method for determining the number of static collaborative elements is as follows: First, check whether each element in the multi-element collaborative edge has undergone significant state changes within the reference scenario period, and then determine whether the change will affect the execution of the action chain; if no significant changes occur or the changes do not affect the execution logic, they are recorded as static collaborative elements; for example, platform model, fixed affiliation and long-term effective rules should usually be included in static collaborative elements, while temporary windows and real-time fire control status are not included.

[0201] In detail, the specific graph reorganization method of dual-channel reorganization is as follows: First, calculate the initial state score and action chain score for each multi-element collaborative edge. Then, according to the action chain allocation ratio, the knowledge within the edge is divided into three categories: initial state node attributes, static relationships, and dynamic triggering relationships. Static attributes and static relationships are written into the initial state channel, while dynamic triggering relationships and sequential dependencies are written into the action chain channel. For example, platform deployment within the same edge is written into the initial state channel, while the task triggering sequence relationship is written into the action chain channel.

[0202] In detail, the setting method for the first, second, and third initial weights is as follows: the weights are calibrated based on the influence of static attributes on the initialization success rate in historical scenarios, and the sum of the three weights is kept to 1; they are preferably set to 0.45, 0.3, and 0.25; for example, if a certain type of simulation scenario is more dependent on the continuously effective rules, the second initial weight can be appropriately increased, but the third initial weight should be reduced at the same time to avoid the total weight being greater than 1.

[0203] In detail, the setting method for the weights of the first action, the second action, and the third action is as follows: First, calculate the impact of the proportion of dynamic fields, the concentration of normalized transient triggers, and the importance of fusion and collaboration on the execution accuracy of the action chain, and then calibrate them under the premise that the sum of the three weights is 1; the preferred settings are 0.25, 0.35, and 0.4; for example, if the temporal burst features are more critical in a certain scenario, the weight of the second action can be appropriately increased, but it should not exceed the weight of the third action by too much, so as not to ignore the overall importance.

[0204] In detail, the mapping method from action chain allocation ratio to knowledge distribution action is as follows: when the action chain allocation ratio is greater than or equal to 0.7, all core relationships of the multi-element collaborative edge are written into the action chain configuration file, and only necessary static anchors are retained in the initial state configuration file; when the ratio is between 0.4 and 0.7, static attributes and dynamic relationships are split proportionally; when the ratio is less than 0.4, the initial state configuration file takes precedence; for example, for an edge with a ratio of 0.85, its temporal dependency and triggering order should be maintained first.

[0205] Preferably, a simulation configuration file is generated based on the action chain allocation ratio, and a weighted loading coverage rate is calculated based on a preset pattern mapping completeness to control the scene loading of the simulation engine, including:

[0206] Based on the action chain allocation ratio, the deduction configuration file, which includes the initial state configuration file, the action chain configuration file, and the incremental patching configuration file, is generated;

[0207] Obtain the number of mapped standard fields and the number of standard baseline fields of the multi-element collaborative edge, and calculate the pattern mapping completeness and the weighted loading coverage according to the following formulas:

[0208]

[0209]

[0210] In the formula, This indicates the completeness of the pattern mapping. This indicates the number of mapped standard fields. Indicates the number of the standard benchmark fields. This represents the weighted load coverage. Represents the set of globally collaborative edges. This represents the multi-element collaborative edge. The summation symbol is used to represent the summation symbol. Indicates the importance of the fusion and collaboration. This represents the zero-prevention micro constant;

[0211] Based on the weighted loading coverage, the simulation engine is controlled to load the simulation configuration file sequentially.

[0212] The simulation configuration file is a collection of structured configuration files used to drive the simulation engine to load and execute battlefield scenarios.

[0213] The initial state configuration file is used to describe the initial deployment of an entity, basic attributes, start rules, and static environment conditions.

[0214] The action chain configuration file is used to describe the event triggering order, task execution dependencies, dynamic rule switching, and action sequence relationships.

[0215] Incremental patch configuration files are used to describe additional corrections, partial replacements, and runtime patch content added to the initial state configuration file and action chain configuration file.

[0216] The number of mapped standard fields is the total number of fields in a multi-feature collaborative edge that have been successfully mapped to the target standard pattern field set.

[0217] The number of standard baseline fields is the total number of baseline fields required for scene loading in the target standard pattern. Preferably, it is 32. This range covers the entity, task, rule, environment, and trigger information sets of common simulation engine loading patterns, and 32 fields balance completeness and operability.

[0218] Schema mapping completeness is a normalized coverage metric of the number of mapped standard fields relative to the number of standard baseline fields.

[0219] The global collaborative edge set is the set of all multi-element collaborative edges in the current scene to be loaded that participate in the coverage calculation.

[0220] The loading molecule is a weighted loading molecule formed by summing the product of the fusion importance of each multi-element collaborative edge in the global collaborative edge set and the pattern mapping completeness.

[0221] The loading denominator is a weighted loading denominator formed by summing the importance of all fused collaboration edges within the global collaborative edge set and superimposing a zero-prevention micro constant.

[0222] Weighted load coverage is an overall load coverage index obtained by comprehensively considering the importance of collaborative edges of various factors and the completeness of pattern mapping.

[0223] In detail, the file structure of the simulation configuration files is as follows: They are uniformly saved using a key-value pair structured text format or a tag-based structured text format, and each scene object is assigned a unique primary key, object type, source edge number, and version number. The initial state configuration file must contain at least an entity definition area, an environment definition area, and a rule base area; the action chain configuration file must contain at least a trigger condition area, an execution order area, and a dependency relationship area; and the incremental patch configuration file must contain at least a patch target area, a patch content area, and an effective condition area. For example, each combat unit object should have a unique number to facilitate mutual referencing between the three files.

[0224] In detail, the way the action chain allocation ratio is mapped to the three types of configuration files is as follows: First, the strength of writing core knowledge into the action chain configuration file is determined based on the action chain allocation ratio. Then, the remaining stable attributes are written into the initial state configuration file. At the same time, local corrections that are not yet fully determined but may affect execution are written into the incremental patch configuration file. For example, when the action chain allocation ratio is 0.8, 80% of the triggering relationships and timing dependencies can be written into the action chain configuration file, 20% of the stable anchors can be written into the initial state configuration file, and parameters that may continue to change with the logs can be put into the incremental patch configuration file.

[0225] In detail, the statistical method for the number of mapped standard fields is as follows: only standard fields that have successfully completed semantic matching, type matching, and value range validation are counted; fields that are partially mapped but still require manual correction are not counted; when multiple source fields are mapped to the same standard field, they are counted as one standard field; for example, if a location field has only completed name matching but the coordinate units are not unified, it cannot be counted in the number of mapped standard fields.

[0226] In detail, the timing and method for updating the global collaborative edge set are as follows: after each round of relationship fusion and dynamic update, all valid multi-element collaborative edges in the current scenario are recalculated, and the latest result is used as the input for the next round of weighted loading coverage calculation. For example, if a new log causes an edge to become invalid or a new collaborative edge is generated, the global collaborative edge set should be updated before the next coverage calculation to avoid using an expired set.

[0227] In detail, the control threshold method for weighted loading coverage is as follows: the full loading threshold is preferentially set to 0.85, and the partial loading threshold is 0.65; when the weighted loading coverage is greater than or equal to 0.85, the three types of configuration files are loaded in normal order; when it is between 0.65 and 0.85, partial loading is allowed while retaining key collaborative edges; when it is below 0.65, mode patching or manual review should be triggered; for example, if the coverage in a critical combat scenario is only 0.6, it is not advisable to directly put it into automatic loading.

[0228] In detail, the loading order and failure rollback method of the simulation engine are as follows: First, load the initial state configuration file to verify whether the entity and environment objects have been successfully instantiated; then load the action chain configuration file to verify whether the task dependencies and trigger relationships have been successfully bound; finally, load the incremental patch configuration file to verify whether the patch hits the target object; if any stage fails, roll back to the version that has been confirmed to be successful in the previous step and record the error log; for example, if the action chain configuration file fails to load, the current action chain update should be canceled, and the initial state object that has been successfully loaded should be retained.

[0229] The embodiments of this example have been described above. However, this example is not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms based on the guidance of this example, and all of them are within the protection scope of this example.

Claims

1. A method for generating military simulation knowledge graphs based on large language models, characterized in that, include: The collected military simulation data undergoes transient constraint perception preprocessing to extract semantic units and calculate the data quality score and static field ratio of the semantic units. The military simulation data includes entity state data, task script data, combat rule data, environmental data, and simulation log data, which are collected through the simulation engine log collection interface, database extraction interface, message bus subscription interface, situation data exchange file, and rule base export interface. The semantic unit is input into the military big language model to extract simulation element tuples, and the element weights of the simulation element tuples are calculated based on the data quality scores. Multidimensional semantic similarity is used to perform cross-layer entity alignment on the simulation element tuples to construct a multidimensional temporal collaborative graph containing multi-element collaborative edges; For the multi-element collaborative edges, collaborative trigger sparsity attributes are extracted, and a temporal validity function is established to calculate the mean trigger density and transient trigger concentration of the multi-element collaborative edges, obtaining a sparsity determination value, including: Obtain the single-element temporal availability and element dependency weight of the simulated collaborative elements contained within the multi-element collaborative edge; The temporal availability of the single element is exponentially operated on with the element's subordinate weight as the exponent, and the exponential operation results of all the simulated collaborative elements within the multi-element collaborative edge are multiplied together to obtain the temporal effectiveness function. The time series validity function is subjected to a definite integral operation over the time interval from time zero to the reference scene period, and the result of the definite integral operation is divided by the reference scene period to obtain the mean trigger density. Obtain the maximum extreme value of the time series validity function within the time interval; The mean trigger density is summed with the zero-prevention micro constant to obtain the denominator term, and the maximum extreme value is divided by the denominator term to obtain the transient trigger concentration. Subtract the mean trigger density from the value to obtain the difference term, and multiply the transient trigger concentration by the difference term to obtain the sparsity determination value. The constraint-driven relationship fusion and dynamic update are performed, combining the element weights, the mean trigger density and the transient trigger concentration into a collaborative trigger vector, and combining the update gain to calculate the rule offset amplitude. Based on the static field ratio and the collaborative triggering vector, the multidimensional time-series collaborative graph is recombined in two channels to calculate the initial state score and action chain score to determine the action chain allocation ratio. A simulation configuration file is generated based on the action chain allocation ratio. A weighted loading coverage rate is calculated based on the pattern mapping completeness to control the scene loading of the simulation engine. The simulation configuration file includes an initial state configuration file, an action chain configuration file, and an incremental patch configuration file. The loading order and failure rollback method of the simulation engine are controlled as follows: first, the initial state configuration file is loaded to verify whether the entity and environment objects are successfully instantiated; then, the action chain configuration file is loaded to verify whether the task dependencies and trigger relationships are successfully bound; finally, the incremental patch configuration file is loaded to verify whether the patch hits the target object. If any stage fails, the system rolls back to the version that was confirmed to be successful in the previous step and records the error log.

2. The method for generating a military simulation knowledge graph based on a large language model according to claim 1, characterized in that, The collected military simulation data undergoes transient constraint perception preprocessing to extract semantic units, and the data quality score and static field ratio of the semantic units are calculated, including: The military simulation data is segmented into semantic units based on a preset data source template; For the semantic unit, the number of retained valid fields and the number of template baseline fields are obtained, and the ratio of the number of retained valid fields to the number of template baseline fields is used as the field completeness ratio. Obtain the relative scene anchor time aging amount and reference scene period of the semantic unit, and use the negative of the quotient of the relative scene anchor time aging amount and the reference scene period as the exponent to calculate the time decay index. Obtain the source reliability value of the semantic unit; The data quality score is obtained by multiplying the field completeness ratio, the source reliability value and the time decay index by the first preset weight, the second preset weight and the third preset weight respectively, and then summing them. Obtain the number of static fields and the number of dynamic fields of the semantic unit, and divide the number of static fields by the sum of the number of static fields and the number of dynamic fields to obtain the proportion of static fields.

3. The method for generating a military simulation knowledge graph based on a large language model according to claim 2, characterized in that, The semantic units are input into the military large language model to extract simulation element tuples. Based on the data quality scores, the element weights of the simulation element tuples are calculated, including: The semantic unit is input into the military big language model, and the simulation element tuple containing multiple slots is extracted. The multiple slots include: combat entity slot, mission slot, rule slot, environment slot, target effect slot, time window slot and semantic temporal phase slot. Obtain the slot extraction confidence of each slot in the simulation element tuple output by the military big language model. The confidence scores of each slot are summed, and the summation result is divided by seven to calculate the element completeness value. The data quality score is multiplied by the element completeness value to obtain the element weight.

4. The method for generating a military simulation knowledge graph based on a large language model according to claim 3, characterized in that, Cross-layer entity alignment is performed on the simulation element tuples using multi-dimensional semantic similarity to construct a multi-dimensional temporal collaboration graph containing multi-element collaboration edges, including: Obtain the first entity semantic vector and the second entity semantic vector, the first entity topological descriptor and the second entity topological descriptor, the first entity temporal anchor point and the second entity temporal anchor point, and the first entity position vector and the second entity position vector of the entity to be aligned; Calculate the cosine similarity between the semantic vector of the first entity and the semantic vector of the second entity, and the Jaccard similarity between the topological descriptor of the first entity and the topological descriptor of the second entity; The time similarity is calculated based on the first entity time anchor point, the second entity time anchor point, and the reference scene period. Spatial similarity is calculated based on the first entity position vector, the second entity position vector, and the reference spatial scale. The multidimensional semantic similarity is obtained by weighted summation of the cosine similarity, the Jaccard similarity, the temporal similarity, and the spatial similarity. Entity alignment is completed based on the multidimensional semantic similarity, and the entity latent state vectors are aggregated along the multi-element collaborative edges based on the graph neural network containing its own state transformation matrix and collaborative state transformation matrix to generate the multidimensional temporal collaborative graph.

5. The method for generating a military simulation knowledge graph based on a large language model according to claim 4, characterized in that, The constraint-driven relationship fusion and dynamic update process combines the element weights, the mean trigger density, and the transient trigger concentration into a collaborative trigger vector, and calculates the rule offset magnitude based on the update gain, including: Obtain the global maximum trigger concentration, and divide the difference between the transient trigger concentration and the value one by the sum of the global maximum trigger concentration minus the value one and the zero-prevention micro constant to obtain the normalized transient trigger concentration. Based on the element weights, the mean element weights and the corresponding source diversity measure are obtained, and these are combined with the mean trigger density and the normalized transient trigger concentration to form the collaborative trigger vector. Obtain the collaborative fusion weight vector and perform an inner product calculation with the collaborative trigger vector to obtain the fusion collaborative importance; Record the collaborative trigger vector of the current round as the current collaborative vector, and obtain the log incremental collaborative vector; Multiply the difference between the value one and the update gain by the current collaborative vector, and add the product of the update gain and the log increment collaborative vector to obtain the updated collaborative vector. The rule offset magnitude is obtained by calculating the L2 norm of the difference between the updated cooperative vector and the current cooperative vector, dividing it by the sum of the L2 norm of the current cooperative vector and the zero-prevention micro constant.

6. The method for generating a military simulation knowledge graph based on a large language model according to claim 5, characterized in that, Based on the static field proportions and the collaborative triggering vector, the multidimensional time-series collaborative graph is reorganized in a dual-channel manner, and the initial state score and action chain score are calculated to determine the action chain allocation ratio, including: Obtain the number of static collaborative elements and the order of the collaborative edge within the multi-element collaborative edge. Divide the number of static collaborative elements by the order of the collaborative edge to obtain the proportion of static fields. Subtract the proportion of static fields from the value to obtain the proportion of dynamic fields. The initial state score is obtained by multiplying the static field ratio, the mean trigger density, the difference between the value one and the normalized transient trigger concentration by the first initial weight, the second initial weight and the third initial weight, respectively. The action chain score is obtained by multiplying the dynamic field ratio, the normalized transient trigger concentration, and the fusion and collaboration importance by the first action weight, the second action weight, and the third action weight, respectively, and then summing them. The action chain allocation ratio is obtained by dividing the action chain score by the initial state score and the sum of the action chain score and the zero-prevention constant.

7. The method for generating a military simulation knowledge graph based on a large language model according to claim 6, characterized in that, Based on the action chain allocation ratio, a simulation configuration file is generated, and a weighted loading coverage rate is calculated based on the pattern mapping completeness to control the scene loading of the simulation engine, including: Based on the action chain allocation ratio, the deduction configuration file, which includes the initial state configuration file, the action chain configuration file, and the incremental patching configuration file, is generated; Obtain the number of mapped standard fields and the number of standard reference fields of the multi-element collaborative edge, and divide the number of mapped standard fields by the number of standard reference fields to obtain the pattern mapping completeness. Obtain the global collaborative edge set, multiply the fusion collaborative importance corresponding to all the multi-element collaborative edges in the global collaborative edge set by the pattern mapping completeness, and then sum them to obtain the loading numerator; Summing up the importance of all the fusion collaborations within the global collaborative edge set and adding the zero-prevention micro constant, yields the loading denominator. The loading numerator is divided by the loading denominator to obtain the weighted loading coverage, and the simulation engine is controlled to load the simulation configuration file sequentially based on the weighted loading coverage.

Images