CN-122021839-A - Military simulation knowledge graph generation method based on large language model

CN122021839ACN 122021839 ACN122021839 ACN 122021839ACN-122021839-A

Abstract

The invention relates to the technical field of knowledge maps and discloses a military simulation knowledge map generation method based on a large language model, which comprises the steps of performing transient constraint perception preprocessing on military simulation data, and extracting simulation element tuples by using the military large language model to calculate element weights; the method comprises the steps of executing cross-layer entity alignment by utilizing multi-dimensional semantic similarity to construct a multi-dimensional time sequence collaborative graph, extracting collaborative trigger sparse attributes aiming at multi-element collaborative edges, establishing a time sequence effectiveness function, calculating trigger density and concentration, further executing relation fusion and dynamic updating, reorganizing the multi-dimensional time sequence collaborative graph to determine action chain allocation proportion, and finally generating a deduction configuration file according to the proportion to control scene loading of a simulation engine. The method can accurately reserve transient burst attributes of tactical constraints and ensure efficient execution of the scene file.

Inventors

ZHU ZHENYU
SUN JING
LIU WEIWEI

Assignees

南京宇天智云仿真技术有限公司

Dates

Publication Date: 20260512
Application Date: 20260410

Claims (8)

1. The military simulation knowledge graph generation method based on the large language model is characterized by comprising the following steps of: performing transient constraint perception preprocessing on the collected military simulation data, extracting a semantic unit, and calculating the data quality score and the static field duty ratio of the semantic unit; Inputting the semantic unit into a military large language model to extract a simulation element tuple, and calculating element weights of the simulation element tuple based on the data quality scores; performing cross-layer entity alignment on the simulation element tuples by utilizing the multi-dimensional semantic similarity, and constructing a multi-dimensional time sequence collaborative graph containing multi-element collaborative edges; Extracting cooperative triggering sparse attributes aiming at the multi-element cooperative edges, and establishing a time sequence effectiveness function to calculate the average trigger density and the transient trigger concentration of the multi-element cooperative edges so as to obtain a sparse judgment value; performing constraint state driven relation fusion and dynamic update, combining the element weight, the mean trigger density and the transient trigger concentration degree into a cooperative trigger vector, and calculating rule offset amplitude by combining update gain; performing double-channel recombination on the multi-dimensional time sequence collaborative graph based on the static field duty ratio and the collaborative trigger vector, and calculating an initial state score and a action chain score to determine an action chain allocation proportion; And generating a deduction configuration file according to the action chain allocation proportion, and calculating weighted loading coverage rate based on the mode mapping completeness so as to control scene loading of the simulation engine.
2. The large language model based military simulation knowledge graph generation method of claim 1, wherein performing transient constraint perception preprocessing on collected military simulation data, extracting semantic units, and calculating a data quality score and a static field duty ratio of the semantic units comprises: dividing the military simulation data into semantic units according to a preset data source template; Aiming at the semantic unit, acquiring the number of reserved effective fields and the number of template reference fields, and taking the ratio of the number of reserved effective fields to the number of template reference fields as a field integrity ratio; Acquiring the relative scene anchor point time aging amount and the reference scene period of the semantic unit, taking a natural constant as a base number, taking the negative number of the quotient value of the relative scene anchor point time aging amount and the reference scene period as an index, and calculating to obtain a time attenuation index; acquiring a source reliability degree value of the semantic unit; Multiplying the field integrity ratio, the source reliability degree value and the time attenuation index by a first preset weight, a second preset weight and a third preset weight respectively, and then summing to obtain the data quality score; And obtaining the static field quantity and the dynamic field quantity of the semantic unit, and dividing the static field quantity by the sum of the static field quantity and the dynamic field quantity to obtain the static field duty ratio.
3. The large language model based military simulation knowledge graph generation method of claim 2, wherein inputting the semantic unit into a military large language model to extract a simulation element tuple, calculating element weights of the simulation element tuple based on the data quality score, comprises: Inputting the semantic unit into the military large language model, and extracting the simulation element tuples comprising a plurality of slots, wherein the plurality of slots comprise a combat entity slot, a task slot, a rule slot, an environment slot, a target effect slot, a time window slot and a semantic time phase slot; Obtaining the slot position extraction confidence of each slot position corresponding to the plurality of slot positions in the simulation element tuple output by the military large language model; summing the extracted confidence degrees of the slots, dividing the sum result by seven, and calculating to obtain an element integrity value; multiplying the data quality score by the element integrity value to obtain the element weight.
4. The large language model based military simulation knowledge graph generation method of claim 3, wherein performing cross-layer entity alignment on the simulation element tuples using multi-dimensional semantic similarity, constructing a multi-dimensional timing synergy graph comprising multi-element synergy edges, comprising: acquiring a first entity semantic vector and a second entity semantic vector of an entity to be aligned, a first entity topology descriptor and a second entity topology descriptor, a first entity time anchor point and a second entity time anchor point, and a first entity position vector and a second entity position vector; calculating cosine similarity of the first entity semantic vector and the second entity semantic vector, and Jacquard similarity of the first entity topology descriptor and the second entity topology descriptor; Calculating time similarity based on the first entity time anchor point, the second entity time anchor point and the reference scene period; Calculating spatial similarity based on the first entity position vector, the second entity position vector and a reference spatial scale; Carrying out weighted summation on the cosine similarity, the Jaccard similarity, the time similarity and the space similarity to obtain the multidimensional semantic similarity; And completing entity alignment based on the multi-dimensional semantic similarity, and generating the multi-dimensional time sequence collaborative graph by aggregating entity hidden state vectors along the multi-element collaborative edges based on a graph neural network comprising a self state transformation matrix and a collaborative state transformation matrix.
5. The large language model based military simulation knowledge graph generation method of claim 4, wherein extracting cooperative triggering sparse attributes for the multi-element cooperative edges, establishing a time sequence validity function to calculate average triggering density and transient triggering concentration of the multi-element cooperative edges, and obtaining a sparse decision value comprises: acquiring single-element time sequence availability and element dependent weight of simulation collaborative elements contained in the multi-element collaborative edge; Performing exponentiation operation with the element dependent weight as an exponent on the single-element time sequence availability, and performing continuous multiplication on exponentiation operation results of all simulation cooperative elements in the multi-element cooperative edge to obtain the time sequence effectiveness function; performing fixed integration operation on the time sequence effectiveness function in a time interval from zero time to the reference scene period, and dividing a fixed integration operation result by the reference scene period to obtain the average trigger density; acquiring the maximum extremum of the time sequence effectiveness function in the time interval; Summing the average trigger density and the zero-prevention tiny constant to obtain a denominator term, and dividing the maximum extremum by the denominator term to obtain the transient trigger concentration; And subtracting the mean trigger density from the first value to obtain a difference term, and multiplying the transient trigger concentration by the difference term to obtain the sparse judgment value.
6. The method for generating a large language model based military simulation knowledge graph as set forth in claim 5, wherein performing constraint state driven relationship fusion and dynamic update, combining the element weight, the mean trigger density and the transient trigger concentration into a collaborative trigger vector, and calculating a rule offset amplitude in combination with updating gain, comprises: Obtaining a global maximum trigger concentration, dividing the difference value of the transient trigger concentration minus a value I by the sum value of the global maximum trigger concentration minus the value I and the zero-proof tiny constant, and obtaining a normalized transient trigger concentration; Acquiring element weight average values and corresponding source diversity metrics based on the element weights, and combining the element weight average values, the average trigger densities and the normalized transient trigger concentrations into the collaborative trigger vectors; acquiring a collaborative fusion weight vector and performing inner product calculation with the collaborative trigger vector to obtain a fusion collaborative importance degree; recording the collaborative trigger vector of the current round as a current collaborative vector, and acquiring a log increment collaborative vector; multiplying the difference value of the first value minus the update gain by the current cooperative vector, and adding the product of the update gain and the log increment cooperative vector to obtain an updated cooperative vector; And calculating the sum of the two norms of the updated cooperative vector and the current cooperative vector difference value and dividing the sum of the two norms of the current cooperative vector and the zero-prevention tiny constant to obtain the rule offset amplitude.
7. The large language model based military simulation knowledge graph generation method of claim 6, wherein performing a two-channel reorganization on the multi-dimensional time sequence collaborative graph based on the static field duty ratio and the collaborative trigger vector, calculating an initial state score and an action chain score to determine an action chain allocation ratio, comprises: The number of the static cooperative elements and the cooperative side orders in the multi-element cooperative side are obtained, the static field duty ratio is obtained by dividing the number of the static cooperative elements by the cooperative side orders, and the dynamic field duty ratio is obtained by subtracting the static field duty ratio from the numerical value one; Multiplying the static field duty ratio, the average trigger density, the value I and the difference value of the normalized transient trigger concentration degree by a first initial weight, a second initial weight and a third initial weight respectively, and then summing to obtain the initial state score; multiplying the dynamic field duty ratio, the normalized transient trigger concentration degree and the fusion cooperative importance degree by a first action weight, a second action weight and a third action weight respectively, and then summing to obtain the action chain value; dividing the action chain score by the initial state score, and adding the action chain score and the zero prevention tiny constant to obtain the action chain allocation proportion.
8. The large language model based military simulation knowledge graph generation method of claim 7, wherein generating a deduction configuration file based on the operation chain allocation proportion, calculating weighted loading coverage based on pattern mapping completeness, to control scene loading of a simulation engine, comprises: generating the deduction configuration file comprising an initial state configuration file, a action chain configuration file and an incremental repair configuration file according to the action chain allocation proportion; Obtaining the number of mapped standard fields and the number of standard reference fields of the multi-element cooperative edge, dividing the number of mapped standard fields by the number of standard reference fields, and obtaining the pattern mapping completeness; acquiring a global collaborative edge set, multiplying the fused collaborative importance corresponding to all the multi-element collaborative edges in the global collaborative edge set by the pattern mapping completeness, and then summing to obtain a loading molecule item; Summing all the fusion cooperative importance levels in the global cooperative edge set, and adding the zero-proof tiny constant to obtain a loading denominator item; Dividing the loading numerator item by the loading denominator item to obtain the weighted loading coverage rate, and controlling the simulation engine to sequentially load the deduction configuration file based on the weighted loading coverage rate.

Description

Military simulation knowledge graph generation method based on large language model Technical Field The invention relates to the technical field of knowledge maps, in particular to a military simulation knowledge map generation method based on a large language model. Background Along with the evolution of artificial intelligence technology, knowledge maps have been widely applied to battlefield environmental cognition and deduction planning. In early battlefield knowledge modeling work, a triplet representation was typically employed to concatenate various types of intelligence with entities. However, this representation can severely fragment battlefield environmental knowledge. Later, though a dynamic evolution model introducing a time dimension appears, high-order interactive modeling is generally lacking. In the actual military simulation study, the simulation scene is not a simple entity and relation set, but is formed by interweaving various elements such as an agent, a task, a rule, a start-stop state of action, time sequence association, a trigger rule and the like. The requirements of different simulation hierarchies and simulation modes on the scene script are not consistent. In the generation process, whether the simulation action has the executable performance is really determined, and a high-order constraint beam is often adopted, for example, a certain fight unit, a certain task, a certain fight rule, a certain terrain shielding condition, a certain weather window, a certain electromagnetic state and a certain time slice. The current common atlas generation flow is still used to forcibly disassemble the high-order constraint into a plurality of binary edges, and then the local alignment and relationship fusion are carried out by using a graph neural network. This increases connectivity on the atlas surface but results in severe splitting of the execution trigger bundles that would otherwise have to be held together within a short window. When the dynamic update of the subsequent superposition deduction log is carried out, the local side weight and the relation naming can change, and then a hidden and troublesome trigger drift phenomenon is caused, wherein the generated simulation scene configuration file can be normally loaded by a simulation engine, but the starting time, the trigger sequence and the environment dependence of an action chain are greatly deviated from the original tactical constraint. Disclosure of Invention The invention provides a military simulation knowledge graph generation method based on a large language model, which solves the technical problem proposed by the background technology. The invention provides a military simulation knowledge graph generation method based on a large language model, which comprises the following steps: performing transient constraint perception preprocessing on the collected military simulation data, extracting a semantic unit, and calculating the data quality score and the static field duty ratio of the semantic unit; Inputting the semantic unit into a military large language model to extract a simulation element tuple, and calculating element weights of the simulation element tuple based on the data quality scores; performing cross-layer entity alignment on the simulation element tuples by utilizing the multi-dimensional semantic similarity, and constructing a multi-dimensional time sequence collaborative graph containing multi-element collaborative edges; Extracting cooperative triggering sparse attributes aiming at the multi-element cooperative edges, and establishing a time sequence effectiveness function to calculate the average trigger density and the transient trigger concentration of the multi-element cooperative edges so as to obtain a sparse judgment value; performing constraint state driven relation fusion and dynamic update, combining the element weight, the mean trigger density and the transient trigger concentration degree into a cooperative trigger vector, and calculating rule offset amplitude by combining update gain; performing double-channel recombination on the multi-dimensional time sequence collaborative graph based on the static field duty ratio and the collaborative trigger vector, and calculating an initial state score and a action chain score to determine an action chain allocation proportion; And generating a deduction configuration file according to the action chain allocation proportion, and calculating weighted loading coverage rate based on the mode mapping completeness so as to control scene loading of the simulation engine. The method has the advantages that the simulation element tuples are extracted through transient constraint perception preprocessing and large language model joint, high-order cooperative association in a complex battlefield environment is completely reserved, a multi-dimensional time sequence cooperative graph and cooperative triggering sparse attribute extraction are further utilized, tran