A method and system for adaptive scoring and association recommendation of joint strike targets
Through the knowledge graph hypernetwork construction and reasoning technology, the problem of unconsidered correlation factors of the target system value in joint firepower strikes was solved, adaptive scoring and intelligent target recommendations were achieved, and the intelligence level of commander decision-making was improved.
Patent Information
- Application Number
- CN202111079754.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2021-09-15
- Publication Date
- 2024-11-19
- Estimated Expiration
- 2041-09-15
AI Technical Summary
The existing technology fails to effectively consider the system value correlation factors between goals, resulting in high subjectivity and arbitrary problems in joint firefighting, and the inability to achieve adaptive scoring and intelligent target recommendations.
Using the knowledge graph hypernetwork construction method, the value score of entities in the hypernetwork is calculated based on the hypernetwork through the construction of the target knowledge graph, the generation and optimization of the initial reasoning vector, the completion of the target knowledge graph, and the correlation recommendation of the firepower strike target. Network relationships determine the associated recommendation targets between entities.
The adaptive scoring and related recommendation of joint targets have been achieved, the intelligence level of commander decision-making has been improved, the problems of subjectivity and arbitrary have been overcome, and the accuracy and timeliness of target recommendations have been improved.
Smart Images

Figure CN113887888B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of joint combat command, and in particular to target value assessment and precise recommendation in joint firepower strikes. Background Art
[0002] With the continuous deepening and expansion of intelligent technology, the face of joint combat command is undergoing profound changes, and the combat command mode is gradually changing from informatization with links as the way to intelligentization with auxiliary decision-making as the goal. The target system evaluation method based on system integration, association recommendation, and breaking links and networks will become the mainstream of future joint strike target optimization and recommendation. For commanders, finding the preferred target in the shortest time and sorting it with other strike targets according to the correlation to obtain the recommendation results will have an important impact on the accuracy and timeliness of the commander's decision. Based on this, the present invention focuses on the construction of the strike target super network system and the target association recommendation algorithm, and attempts to change the existing tabular target recommendation method with knowledge graphs, and study the intrinsic attributes and mutual correlation mechanisms of the targets in the graph super network.
[0003] Prior art 1: Patent No. CN108416441A, which implements a ship-to-shore strike firepower allocation method based on genetic algorithm, takes a weapon on a ship as a firepower unit, takes the number of firepower allocations of the firepower unit to the target on the shore as the coding bit, and uses the random distribution generation method to encode each firepower unit to obtain the initial population; uses the fitness function to calculate the population fitness of the initial population, selects the initial population according to the population fitness combined with roulette, obtains the selected population, performs chromosome crossover and mutation on the selected population, obtains a new population; uses the new population to update the initial population and repeats the previous step, and after repeated multiple times, obtains the ship-to-shore strike firepower allocation plan. However, the system value association factors between targets are not considered, and the target system value analysis and system-based target firepower allocation cannot be realized.
[0004] Prior art 2: Patent No. CN110119580A implements a ground engineering target damage assessment system and method, which can call a fast damage effect algorithm to calculate the damage result data of the engineering target under the action of ammunition based on the characteristic parameter data of the received weapons and ammunition and the target characteristic parameters, and upload the calculation results to the damage assessment module and the simulation demonstration module. The present invention can carry out rapid prediction and evaluation of damage effects under different ammunition usage. However, the patent relies too much on the subjective scoring factors of experts and cannot achieve adaptive scoring and intelligent target recommendation of people outside the loop.
[0005] Therefore, the above-mentioned solutions of the prior art fail to take into account the system value correlation factors between the targets, and cannot achieve adaptive scoring and intelligent target recommendation; there are problems of high subjectivity and arbitrariness. Summary of the invention
[0006] In order to overcome the above-mentioned deficiencies of the prior art, the present invention provides a method, system and storage medium for adaptively scoring and associating recommendations for joint strike targets.
[0007] This paper focuses on the study of the complex hypernetwork system construction method for attack targets, attempts to use knowledge graph reasoning technology to complete and improve the target hypernetwork structure, and on this basis constructs emergent evaluation indicators to measure the system value of the target, and recommends related targets that meet the commander's decision-making intentions.
[0008] Aiming at the problem of high subjective arbitrariness in target value assessment and accurate recommendation in joint fire strike, the present invention designs a joint strike target adaptive scoring and associated recommendation method, which includes the following steps:
[0009] Step 1: Construction of target knowledge graph hypernetwork: Based on the acquired target intelligence data, the relationships between targets are used to annotate subnetwork relationships and construct a target knowledge graph hypernetwork structure model;
[0010] Step 2: Generate the initial reasoning vector of the knowledge graph: map the entity relationship into a spatial vector, that is, assign a random vector array to each entity and its relationship;
[0011] Step 3, optimization of the initial reasoning vector of the knowledge graph: use the group evolution idea of the meta-heuristic algorithm to achieve iterative optimization of entity and relationship vectors; through a finite number of iterations based on stochastic gradient descent, the loss function containing the scoring function converges to the minimum value;
[0012] Step 4: Completion of the target knowledge graph: Use the optimized vector of the optimal agent to calculate the comprehensive relationship scores of all entities, and compare them with the average scores of existing relationships in the hypernetwork to mine the hidden relationships in the hypernetwork and complete the hypernetwork.
[0013] Step 5, related recommendation of fire strike targets: Use the target system value calculation formula to calculate the value scores of all entities in the knowledge graph hypernetwork, and determine the related recommended target names between entities based on the hypernetwork relationship.
[0014] Furthermore, in step 1, the specific implementation method is: based on the acquisition of target intelligence data, the subnet relationship is marked using the entity relationship, affiliation and functional relationship between the targets. In order to more accurately describe the nature, composition and associated attributes of each subnet in the target super network, the target is specifically divided into 9 types of relationships: electronic countermeasures, air defense counterattack, rear support, transportation hub, sub-target affiliation, energy facilities, communication facilities, psychological warfare, and command and control. 9 subnets are constructed in turn, and random cutting is used to convert the undirected network into a directed network. The target entities in each subnet are intertwined with each other, and the target super network architecture is entangled and superimposed.
[0015] The target subnet division is shown in Table 1:
[0016] Table 1: Division of each subnet of the attack target
[0017]
[0018]
[0019] Furthermore, step 2 can be divided into the following sub-steps:
[0020] Step 2-1, generate a multidimensional random vector; let the vector dimension be k, then the random number should be selected A real number between , all entities and relations are assigned random vectors and normalized: entity
[0021] Step 2-2, generate a combination of positive and negative samples; the positive samples include all the triplets with connections in the hypernetwork, and the negative samples are triplets without connections in the current network constructed by randomly adjusting the head or tail entity based on the positive sample set. The positive and negative samples are set to (h, r, t) and (h', r, t') respectively; where h and h' are the head entity vectors of the positive and negative samples (if h is an entity vector pointing to a certain target, then h' is an entity vector pointing to a certain target excluding the target), r is the entity relationship, and t and t' are the tail entity vectors of the positive and negative samples;
[0022] Furthermore, step 3 can be divided into the following sub-steps:
[0023] Step 3-1, calculate the hypernetwork loss function; the goal of the algorithm is: h+r=t; h'+r≠t', which can be accurately measured by constructing a scoring function f(h,r,t), and the calculation formula is:
[0024] f(h,r,t)=||h+rt|| (1)
[0025] The calculation formula of the loss function L(h,r,t,r',t') is:
[0026]
[0027] In the formula, Ensure that the function value satisfies max(0,...), and the γ representing the entity relationship is replaced by a real number greater than 0; S represents the existing entity relationship set of the hypernetwork; S' represents the unassociated entity relationship set of the hypernetwork, satisfying the following conditions:
[0028]
[0029] Step 3-2, construct the agent structure; each agent contains an entity vector set, an edge vector set and a negative sample edge vector set;
[0030] Step 3-3, construct the initial agent population; adopt the random mutation method of the meta-heuristic algorithm, take the original agent as the sample, randomly select 10 entities and edges for vector fine-tuning; r represents the vector relationship, δ represents the fine-tuning parameter, ω represents the original parameter in the vector, and the calculation formula of the fine-tuned parameter ω' is:
[0031]
[0032] Step 3-4, calculate the comprehensive score of all agents in the population; according to the updated loss function scoring formula, bring the entity vector set, edge vector set, and negative sample edge vector set contained in each agent into the scoring formula to calculate the minimum score L of the existing edges min and the maximum score of negative samples L max ; The specific scoring principle is: L min As the main score of the agent; if the two agents L min Same, L max As an agent to assist scoring;
[0033] Step 3-5, population elimination: according to the comprehensive scores of each agent in the population, 90% of the agents in the population are eliminated, and the best scoring agent structure and corresponding score are recorded;
[0034] Step 3-6, population expansion: call the random mutation method in the above steps to generate new agents using the remaining agents after elimination as samples, so that the population reaches the original size;
[0035] Step 3-7, repeat steps 3-3 to 3-5 until the exit condition is met: L min Minimize, and L max If it stops growing, the algorithm is considered to have converged, and it exits and feeds back the best agent in the population.
[0036] Furthermore, in step 4, after obtaining the optimal entity and edge vector matrix of the hypernetwork, a hypernetwork relationship completion operation is introduced: the triples (h, r, t) in the existing hypernetwork are collected, and their head entities or tail entities are replaced with other entities in turn, thereby generating a set of candidate relationship arrays, and their evaluation values f(h', r, t) or f(h, r, t') are calculated respectively, and then sorted in ascending order according to the evaluation values, and the sorting position of the original triples (h, r, t) is recorded; if the ranking exceeds 5, the first 5 triple relations are supplemented to the hypernetwork; otherwise, the relations whose ranking exceeds the original triples are supplemented to the hypernetwork; the attribute array of the newly added relationship satisfies the formula: r = ht.
[0037] Furthermore, in step 5, the associated recommendation of the firepower strike target is realized; the specific implementation method is: the target system value includes the vector weight value and edge weight value of the target; let the entity vector contained in the i-th sub-target entity be h i , after completing the network, the set of edges associated with this sub-goal is S i , then the system value Z of the i-th goal i for:
[0038]
[0039] The present invention provides a joint strike target adaptive scoring and association recommendation system, which is characterized by comprising:
[0040] Target knowledge graph super network construction module: Based on the acquired target intelligence data, the sub-network relationship is annotated using the relationship between targets to construct the target knowledge graph super network structure model;
[0041] Knowledge graph initial reasoning vector generation module: maps entity relationships into spatial vectors, that is, assigns a random vector array to each entity and its relationship;
[0042] Knowledge graph initial reasoning vector optimization module: It uses the group evolution idea of the meta-heuristic algorithm to achieve iterative optimization of entity and relationship vectors; through a finite number of iterations based on stochastic gradient descent, the loss function containing the scoring function converges to the minimum value;
[0043] Target knowledge graph completion module: Use the optimized vector of the optimal agent to calculate the comprehensive relationship scores of all entities, compare them with the average scores of existing relationships in the hypernetwork, and explore the hidden relationships in the hypernetwork to complete the hypernetwork.
[0044] Fire strike target association recommendation module: calculates the value scores of all entities in the knowledge graph hypernetwork, and determines the association recommendation target names between entities based on the hypernetwork relationships.
[0045] The present invention provides a computer-readable storage medium, characterized in that:
[0046] The storage medium stores a plurality of instructions, and the plurality of instructions are used for the processor to load and execute the above-mentioned joint strike target adaptive scoring and association recommendation method.
[0047] Based on the above scheme, the present invention aims at the problems of joint strike target association recommendation and system value assessment, attempts to break out of the traditional expert scoring method framework, establishes a multi-dimensional attribute vector model of the target from the perspective of hypernetwork adaptive learning and knowledge graph reasoning, and designs a mathematical model for joint strike target association recommendation on this basis. The effectiveness of the algorithm of the present invention is proved through comparative analysis of similar algorithms.
[0048] The present invention focuses on solving three problems: first, the complex hypernetwork construction problem of joint strike targets; second, the hidden relationship completion problem of complex hypernetwork systems; and third, the adaptive optimization and emergent measurement problem of strike target attribute vectors. Based on solving the above problems, the algorithm's computational effectiveness and decision-making applicability are verified through simulation experiments.
[0049] The innovations of the present invention are: first, the knowledge graph reasoning technology is introduced into the construction of the joint firepower strike target system, realizing the automatic linking of target association relationships; second, an emergent evaluation index model in a complex hypernetwork system is established, realizing the emergent value measurement of the strike target hypernetwork; third, the commander's associated recommendation and optimal recommendation of the preferred strike targets are realized, improving the level of intelligent decision-making assistance. BRIEF DESCRIPTION OF THE DRAWINGS
[0051] In order to more clearly illustrate the specific implementation methods of the present invention or the technical solutions in the prior art, the drawings required for use in the specific implementation methods or the description of the prior art will be briefly introduced below. Obviously, the drawings in the following description are some implementation methods of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work. BRIEF DESCRIPTION OF THE DRAWINGS
[0052] Figure 1 This is the overall system architecture diagram.
[0053] Figure 2 Give examples for three types of entities.
[0054] Figure 3 This is an example of the initial hypernetwork structure.
[0055] Figure 4 Schematic diagram of randomly cropping an undirected graph into a directed graph.
[0056] Figure 5 Schematic diagram of the intelligent agent structure.
[0057] Figure 6 To complete the hypernetwork structure example.
[0058] Figure 7 An example of the algorithm running effect graph.
[0059] Figure 8 a is a schematic diagram of the comprehensive score comparison results; Figure 8 b is a schematic diagram of the comparison results of standard TransE scoring; Figure 8 c is a schematic diagram of the comparison results of the improved TransE score; Figure 8 d is a schematic diagram of the comparison results of algorithm sample difference scores.
[0060] Fig. 9 a is a schematic diagram showing the comparison of the completion effects of hypernetworks matching different numbers of edges; Fig. 9 b is a schematic diagram comparing the number of free entities in hypernetworks matching different numbers of edges.
[0061] Fig.10 Comparative analysis of multi-dimensional vector attributes for various types of targets.
[0062] Fig.11 Comparison of agent vector evolution.
[0063] Fig.12 This is the comparison result of the target system value of each method. DETAILED DESCRIPTION
[0064] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution of the present invention will be clearly and completely described below in conjunction with the specific embodiments of the present invention and the corresponding drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.
[0065] Aiming at the problem of high subjective arbitrariness in target value assessment and accurate recommendation in joint fire strike, the present invention designs a joint strike target adaptive scoring and associated recommendation method, which includes the following steps:
[0066] Step 1: Construction of target knowledge graph hypernetwork: Based on the acquired target intelligence data, the subnetwork relationships are annotated using the relationships between targets to construct the target knowledge graph hypernetwork structure model;
[0067] Step 2: Generate the initial reasoning vector of the knowledge graph: map the entity relationship into a spatial vector, that is, assign a random vector array to each entity and its relationship;
[0068] Step 3, optimization of the initial reasoning vector of the knowledge graph: use the group evolution idea of the meta-heuristic algorithm to achieve iterative optimization of entity and relationship vectors; through a finite number of iterations based on stochastic gradient descent, the loss function containing the scoring function converges to the minimum value;
[0069] Step 4: Completion of the target knowledge graph: Use the optimized vector of the optimal agent to calculate the comprehensive relationship scores of all entities, and compare them with the average scores of existing relationships in the hypernetwork to mine the hidden relationships in the hypernetwork and complete the hypernetwork.
[0070] This step solves the problem of excessive network sparsity.
[0071] Step 5, related recommendation of fire strike targets: Use the target system value calculation formula to calculate the value scores of all entities in the knowledge graph hypernetwork, and determine the related recommended target names between entities based on the hypernetwork relationship.
[0072] This step achieves the technical effect of intelligent target recommendation in joint firepower strikes.
[0073] Furthermore, in step 1, the specific implementation method is as follows: based on the acquisition of the target intelligence data, 9 types of subnet relationships are marked using the entity relationships, affiliations and functional relationships between the targets. In order to more accurately describe the nature, composition and associated attributes of each subnet in the target super network, the target is specifically divided into 9 types of relationships: electronic countermeasures, air defense counterattack, rear support, transportation hub, sub-target affiliation, energy facilities, communication facilities, psychological warfare, and command and control. Random cutting is used to convert the undirected network into a directed network, and 9 subnets are constructed in turn. The target entities in each subnet are intertwined with each other, and the target super network architecture is entangled and superimposed.
[0074] The target subnet division is shown in Table 1:
[0075] Table 1: Division of each subnet of the attack target
[0076]
[0077] Furthermore, step 2 can be divided into the following sub-steps:
[0078] Step 2-1, generate a multidimensional random vector; let the vector dimension be k, then the random number should be selected A real number between , all entities and relations are assigned random vectors and normalized: entity
[0079] Step 2-2, generate a combination of positive and negative samples; the positive samples include all the triplets with connections in the hypernetwork, and the negative samples are triplets without connections in the current network constructed by randomly adjusting the head or tail entity based on the positive sample set. The positive and negative samples are set to (h, r, t) and (h', r, t') respectively; where h and h' are head entity vectors (if h is an entity vector pointing to a certain target, then h' is an entity vector pointing to a certain target excluding the target), r is the entity relationship, and t and t' are tail entity vectors;
[0080] Furthermore, in step 3, the idea of group evolution of the meta-heuristic algorithm is used to implement iterative optimization of entity and relationship vectors, so as to minimize the comprehensive score of existing relationships and maximize the comprehensive score of negative sample relationships after multiple generations; a limiting constraint is made: if there is a relationship between two entities, the head entity vector plus the relationship should be approximately equal to the tail entity vector; otherwise, it should be as unlikely as possible; this converts the graph network connection problem into an iterative convergence problem of multi-dimensional space vectors, that is, through a finite number of iterations based on stochastic gradient descent, the loss function containing the scoring function converges to the minimum value;
[0081] Furthermore, step 3 can be divided into the following sub-steps:
[0082] Step 3-1, calculate the hypernetwork loss function; the goal of the algorithm is: h+r=t; h'+r≠t', which can be accurately measured by constructing a scoring function f(h,r,t), and the calculation formula is:
[0083] f(h,r,t)=||h+rt|| (1)
[0084] The calculation formula of the loss function L(h,r,t,r',t') is:
[0085]
[0086] In the formula, Ensure that the function value satisfies max(0,...), and the γ representing the entity relationship is replaced by a real number greater than 0; S represents the existing entity relationship set of the hypernetwork; S' represents the unassociated entity relationship set of the hypernetwork, satisfying the conditions:
[0087]
[0088] Step 3-2, construct the agent structure; each agent contains an entity vector set, an edge vector set and a negative sample edge vector set;
[0089] Step 3-3, construct the initial agent population; adopt the random mutation method of the meta-heuristic algorithm, take the original agent as the sample, randomly select 10 entities and edges for vector fine-tuning; r represents the vector, δ represents the fine-tuning parameter, ω represents the original parameter in the vector, and the calculation formula of the fine-tuned parameter ω' is:
[0090]
[0091] Step 3-4, calculate the comprehensive score of all agents in the population; according to the updated loss function scoring formula, bring the entity vector set, edge vector set, and negative sample edge vector set contained in each agent into the scoring formula to calculate the minimum score L of the existing edges min and the maximum score of negative samples L max ; The specific scoring principle is: L min As the main score of the agent; if the two agents L min Same, L max As an agent to assist scoring;
[0092] Step 3-5, population elimination: according to the comprehensive scores of each agent in the population, 90% of the agents in the population are eliminated, and the best scoring agent structure and corresponding score are recorded;
[0093] Step 3-6, population expansion: call the random mutation method in the above steps to generate new agents using the remaining agents after elimination as samples, so that the population reaches the original size;
[0094] Step 3-7, repeat steps 3-3 to 3-5 until the exit condition is met: L min Minimize, and L max If it stops growing, the algorithm is considered to have converged, and it exits and feeds back the optimal agent structure in the population at this time.
[0095] Furthermore, in step 4, after obtaining the optimal entity and edge vector matrix of the hypernetwork, a hypernetwork relationship completion operation is introduced: the triples (h, r, t) in the existing hypernetwork are collected, and their head entities or tail entities are replaced with other entities in turn, thereby generating a set of candidate relationship arrays, and their evaluation values f(h', r, t) or f9h, r, t') are calculated respectively, and then sorted in ascending order according to the evaluation values, and the sorting position of the original triples (h, r, t) is recorded; if the ranking exceeds 5, the first 5 triple relations are supplemented to the hypernetwork; otherwise, the relations whose ranking exceeds the original triples are supplemented to the hypernetwork; the attribute array of the newly added relationship satisfies the formula: r = ht.
[0096] Furthermore, in step 5, the associated recommendation of the firepower strike target is realized; the specific implementation method is: the target system value includes the vector weight value and edge weight value of the target; let the entity vector contained in the i-th sub-target entity be h i , after completing the network, the set of edges associated with this sub-goal is S i , then the system value Z of the i-th goal i for:
[0097]
[0098] Furthermore, the above weight values are calculated using the Euclidean space distance method.
[0099] The present invention provides a joint strike target adaptive scoring and association recommendation system, which is characterized by comprising:
[0100] Target knowledge graph super network construction module: Based on the acquired target intelligence data, the sub-network relationship is annotated using the relationship between targets to construct the target knowledge graph super network structure model;
[0101] Knowledge graph initial reasoning vector generation module: maps entity relationships into spatial vectors, that is, assigns a random vector array to each entity and its relationship;
[0102] Knowledge graph initial reasoning vector optimization module: It uses the group evolution idea of the meta-heuristic algorithm to achieve iterative optimization of entity and relationship vectors; through a finite number of iterations based on stochastic gradient descent, the loss function containing the scoring function converges to the minimum value;
[0103] Target knowledge graph completion module: Use the optimized vector of the optimal agent to calculate the comprehensive relationship scores of all entities, compare them with the average scores of existing relationships in the hypernetwork, and explore the hidden relationships in the hypernetwork to complete the hypernetwork.
[0104] Fire strike target association recommendation module: calculates the value scores of all entities in the knowledge graph hypernetwork, and determines the association recommendation target names between entities based on the hypernetwork relationships.
[0105] The present invention provides a computer-readable storage medium, characterized in that:
[0106] The storage medium stores a plurality of instructions, and the plurality of instructions are used for the processor to load and execute the above-mentioned joint strike target adaptive scoring and association recommendation method.
[0107] In the method, system and storage medium of another embodiment, the method steps or functional modules are the same as those in the above embodiment. In step 4, the system value of the target is generated by other value scoring methods, and the target strike list is pushed in order of value.
[0108] Figure 1 The overall architecture diagram of the present invention is divided into 5 functional modules, namely, the target knowledge graph super network construction module, the knowledge graph initial reasoning vector generation module, the knowledge graph initial reasoning vector optimization module, the target knowledge graph completion module, and the firepower strike target association recommendation module.
[0109] Figure 2 The following is an example of three types of target entities. The target attack intelligence data format includes the target entity name and the sub-target name. For example, in the sub-target name "DZZ-2 Squadron.CS_PRD-10A DZZC Machine", "DZZ-2 Squadron" is the target entity name and "CS_PRD-10A DZZC Machine" is the sub-target name; the perception strength calibrates the current integrity of the target; the detection source calibrates the reliability of the target intelligence acquisition; the target type calibrates the major category to which the target belongs, which is divided into intelligence support, command and control, firepower strike, electronic interference, equipment support, transportation hub, energy facilities, and communication facilities; the target sub-category calibrates the sub-category to which the target belongs; and the unit to which it belongs calibrates the target entity name. In order to more scientifically divide the entity relationship, the concept of process entity is introduced on the basis of sub-target entity and target entity. The process entity is to accurately describe the affiliation between targets. The artificially introduced virtual target node aims to make the command affiliation relationship between the upper and lower levels clearer.
[0110] Figure 3 This is an example of the initial super network structure constructed after dividing the 9 types of subnetwork relationships. A super network refers to a network that transcends the network, that is, a complex network built on the basis of intelligence, command, firepower, reactance, equipment, transportation, energy, and communication subnetworks. However, the super network is still in an incomplete structure at this time, and many hidden relationships within it are not clearly expressed. The relationship matrix is still relatively sparse. For example, in the command affiliation relationship, there is an association between "A command post" and "A-1 battalion command vehicle", and there is also an association between "A-1 battalion command vehicle" and "A-1-1 air defense company", but there is no association between "A command post" and "A-1-1 air defense company". Therefore, it is necessary to use a knowledge reasoning algorithm to automatically calibrate such hidden associations and complete the super network system. In the initial network construction stage, the initial network of the target system is manually constructed through entity internal relationship calibration, troop affiliation calibration, and information flow relationship calibration.
[0111] Figure 4It is a schematic diagram of the algorithm converting an undirected graph into a directed graph through random cropping operations. During the entire execution process, it was found that the super network system composed of the attack targets cannot be directly applied to the TransE algorithm, mainly manifested in: First, the attack target relationship is an undirected relationship, and the constructed super network belongs to an undirected graph network, and vector addition cannot be used directly; second, the super network matrix is too sparse, and the association relationship between each entity is quite different, which makes it difficult for TransE's loss function to converge to the equivalent effect of the left and right vectors, thereby greatly reducing the subsequent completion effect. In order to solve the above problems, the present invention compresses the undirected graph matrix by random cropping to form a directed graph structure that only contains half of the valid edges.
[0112] Figure 5 It is a schematic diagram of the agent structure. Each agent contains an entity vector set, an edge vector set, and a negative sample edge vector set. Figure 6 It is a schematic diagram of the target knowledge graph hypernetwork structure after completion. Figure 7 This is a diagram showing the effect of the algorithm running.
[0113] In order to verify the effectiveness of the algorithm, the target data in a certain battle scenario exercise was used as the initial input data. Relevant experiments such as complex hypernetwork construction, target association recommendation, and comparison of similar methods were designed to test the recommendation effect and feasibility of the algorithm.
[0114] 1. Complex hypernetwork construction experiment
[0115] The purpose of the experiment is to test the comprehensive performance of the algorithm of this invention and to verify the effectiveness and application advantages of the algorithm by comparing the scores with those of the standard TransE algorithm.
[0116] 1. Knowledge Reasoning Convergence Experiment
[0117] During the algorithm's adaptive iteration process, the best agent of each generation is selected and its comprehensive score, existing edge score, and negative sample score are recorded. The difference between different scores is calculated as follows:
[0118]
[0119] The scoring convergence of the three feedback scoring functions is as follows: Figure 8 shown.
[0120] Through the analysis of the comprehensive scoring formula, the purpose of iteration is to reduce the comprehensive score. The greater the reduction, the better the learning effect. Through experimental comparison and analysis, it can be seen that the comprehensive score in (a) shows that the score reduction of the improved algorithm is significantly weaker than that of the standard algorithm. After 120 generations of iterations, the standard algorithm reduces the comprehensive score by 10.64%, while the improved algorithm can only reduce the comprehensive score by 6.51%. However, after a detailed analysis of the existing edge and negative sample scores of the two algorithms, the existing edge score and negative sample score of the standard algorithm in (b) both increase during the learning process, which is an undesirable result of the algorithm. The ideal result should be: as the number of iterations increases, the existing edge score gradually decreases while the negative sample score gradually increases. The improved algorithm in (c) meets the algorithm concept. As the number of iterations increases, the score gap between the existing edge score and the negative sample score becomes larger and larger. By calculating the difference between the two algorithms, it can be found in (d) that both algorithms have achieved algorithm convergence after multiple generations of iterations. The convergence jitter of the improved algorithm is more obvious, while the standard algorithm gradually tends to 0, proving that the algorithm has achieved multi-generation convergence.
[0121] 2. Hypernetwork Sparsity Experiment
[0122] In order to test the completion effect of the hypernetwork, the number of network entities, the number of edges, and the number of free entities before and after knowledge reasoning are statistically analyzed, and the concept of the minimum number of matching edges is introduced. The number of matching edges is 5, which means that the algorithm calculates the shortest matching edge threshold for each edge as 5. The comparison results of entities and edges before and after incompleteness are shown in Figure 2. Fig. 9 shown.
[0123] Through comparative analysis, it can be seen that when the initial hypernetwork is constructed, the network contains a total of 3368 entity nodes and 7796 relationship edges; through the hypernetwork completion operation (matching 5 edges), 77957 hidden edges are added, which is 11 times the existing relationship edges; from the hypernetwork sparsity analysis, the maximum number of relationship edges that the hypernetwork can carry is 5670028, the sparsity of the initial hypernetwork is 0.14%, the sparsity of the hypernetwork after matching 5 edges is 1.51%, and the sparsity of the hypernetwork after matching 10 edges is 2.89%; through the analysis of the number of free entities in (b), it can be seen that after the hypernetwork completion, the number of free entities in the network is greatly reduced. Through entity analysis, the hidden relationship edges after completion promote the establishment of indirect associations between free entities and entities inside the hypernetwork, thereby greatly improving the association of entities, and the number of free entities gradually decreases with the increase of the number of matching edges.
[0124] 2. Agent Evolution Experiment
[0125] After adaptive learning optimization, the vector parameter changes of various target entities are analyzed, and 10 target types are distinguished for trend radar chart analysis. The multi-dimensional vector attribute comparison analysis of each type of target is as follows: Fig.10shown.
[0126] Through comparative analysis, it can be seen that although the initial vector adjustment parameters of various types of targets are randomly generated, as the number of iterations increases, the mean of the vector adjustment parameters tends to change, such as the adjustment parameters of sensor targets of 6 and the adjustment parameters of facility targets of 2, the change range is obvious, indicating that each target entity has undergone self-adaptive transformation in the process of hypernetwork adaptive iteration, so that the global vector of the hypernetwork is balanced, that is, it develops in the direction of satisfying the comprehensive scoring formula. In general, the radar adjustment parameters of various types of targets gradually change with the increase of the number of iterations, and the adjustment direction of the parameters of different types of targets changes with the network structure. The mean changes of the entity node vector and the relationship edge vector of the intelligent agent are shown in Figure 2. Fig.11 shown.
[0127] As the number of iterations increases, the proportion of the internal vector composition of the agent is gradually adjusted. By comparing the optimal agent of the 120th generation with the first-generation agent, it can be seen that in the first-generation agent, the average score of the node vector parameter adjustment is -2.24 points, and after iteration, the score becomes -12.93 points; the average score of the relationship edge is gradually adjusted from 22.53 points to 0.37 points, and the changes of the two are obvious. It shows that through the evolution of comprehensive scoring constraints, the parameter adjustment score of the relationship edge gradually transfers to the node score, and the importance of the node score gradually increases; after the completion of the hypernetwork and the expansion of the hidden relationship edge, the edge score is gradually compressed and weakened to enhance the influence of the entity node in the hypernetwork. The above phenomenon is also consistent with the actual effect of the joint strike target association recommendation.
[0128] 3. Comparative Experiments on Similar Methods
[0129] In order to verify the effectiveness and rationality of the target system value evaluation algorithm proposed in this paper, the system value evaluation algorithm and the correlation analysis algorithm are selected as comparison methods to comprehensively compare the comprehensive scores of various targets. The comparison results are as follows: Fig.12 shown.
[0130] Through comparative analysis, it can be seen that the target system value score of the algorithm in this paper gives a relatively high score in mobile reconnaissance stations and missile battalion positions. This result is the same as other similar algorithms, but the focus is more concentrated. Correspondingly, the weight distribution in information transmission facilities such as microwave communication hubs, communication headquarters, and data analysis stations is relatively weakened; considering that the algorithm in this paper does not use the expert subjective scoring method, it has obtained a system value scoring result similar to that of other similar algorithms, which proves the correctness and wide application of the algorithm in this paper.
[0131] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.
Claims
1. A joint strike target adaptive scoring and association recommendation method, characterized in that: Step 1: Construction of target knowledge graph hypernetwork: Based on the acquired target intelligence data, the relationships between targets are used to annotate subnetwork relationships and construct a target knowledge graph hypernetwork structure model; Step 2: Generate the initial reasoning vector of the knowledge graph: map the entity relationship into a spatial vector, that is, assign a random vector array to each entity and its relationship; Step 3, optimization of the initial reasoning vector of the knowledge graph: use the group evolution idea of the meta-heuristic algorithm to achieve iterative optimization of entity and relationship vectors; through a finite number of iterations based on stochastic gradient descent, the loss function containing the scoring function converges to the minimum value; Step 4: Completion of the target knowledge graph: Use the optimized vector of the optimal agent to calculate the comprehensive relationship scores of all entities, and compare them with the average scores of existing relationships in the hypernetwork to mine the hidden relationships in the hypernetwork and complete the hypernetwork. Step 5: Recommendation of associated firepower strike targets: Calculate the value scores of all entities in the knowledge graph hypernetwork, and determine the names of associated recommended targets between entities based on the hypernetwork relationships; Among them, step 3 is divided into the following sub-steps: Step 3-1, calculate the hypernetwork loss function; the goal of the algorithm is: h+r=t; h'+r≠t', which can be accurately measured by constructing a scoring function f(h,r,t), and the calculation formula is: f(h,r,t)=||h+rt|| The calculation formula of the loss function L(h,r,t,r',t') is: Where S represents the existing entity relationship set of the hypernetwork; S' represents the unassociated entity relationship set of the hypernetwork, satisfying the following conditions: Step 3-2, construct the agent structure; each agent contains an entity vector set, an edge vector set and a negative sample edge vector set; Step 3-3, construct the initial agent population; adopt the random mutation method of the meta-heuristic algorithm, take the original agent as the sample, randomly select 10 entities and edges for vector fine-tuning; r represents the vector relationship, δ represents the fine-tuning parameter, ω represents the original parameter in the vector, and the calculation formula of the fine-tuned parameter ω' is: Step 3-4, calculate the comprehensive score of all agents in the population; according to the updated loss function scoring formula, bring the entity vector set, edge vector set, and negative sample edge vector set contained in each agent into the scoring formula to calculate the minimum score L of the existing edges min and the maximum score of negative samples L max ; The specific scoring principle is: L min As the main score of the agent; if the two agents L min Same, L max As an agent to assist scoring; Step 3-5, population elimination: according to the comprehensive scores of each agent in the population, 90% of the agents in the population are eliminated, and the best scoring agent structure and corresponding score are recorded; Step 3-6, population expansion: call the random mutation method in the above steps to generate new agents using the remaining agents after elimination as samples, so that the population reaches the original size; Step 3-7, repeat steps 3-3 to 3-5 until the exit condition is met: L min Minimize, and L max If it stops growing, the algorithm is considered to have converged, and it exits and feeds back the best agent in the population at this time; Among them, in step 4, after obtaining the optimal entity and edge vector matrix of the hypernetwork, the hypernetwork relationship completion operation is introduced: collect the triples (h, r, t) in the existing hypernetwork, replace their head entities or tail entities with other entities in turn, and then generate a set of candidate relationship arrays, calculate their evaluation values f(h', r, t) or f(h, r, t'), and then sort them in ascending order according to the evaluation values, and record the sorting position of the original triples (h, r, t); if the sorting exceeds 5, the first 5 triple relations are added to the hypernetwork; otherwise, the relations whose sorting exceeds the original triples are added to the hypernetwork; the attribute array of the newly added relationship satisfies the formula: r = ht.
2. The method according to claim 1, characterized in that: In step 1, the specific implementation method is as follows: based on the acquisition of target intelligence data, the entity relationship, affiliation and functional relationship between the targets are used to mark the subnet relationships, and the targets are specifically divided into 9 types of relationships: electronic countermeasures, air defense counterattack, rear-end support, transportation hubs, sub-target affiliation, energy facilities, communication facilities, psychological warfare, and command and control. Nine subnets are constructed in turn, and random cutting is used to convert the undirected network into a directed network. The target entities in each subnet are intertwined with each other, and the target super-network architecture is entangled and superimposed.
3. The method according to claim 1, characterized in that Step 2 can be divided into the following sub-steps: Step 2-1, generate a multidimensional random vector; let the vector dimension be k, then the random number should be selected A real number between , all entities and relations are assigned random vectors and normalized: entity Step 2-2, generate a combination of positive and negative samples; the positive samples include all the connected triplets in the hypernetwork, and the negative samples are triplets without connections in the current network constructed by randomly adjusting the head or tail entity based on the positive sample set. The positive and negative samples are set to (h, r, t) and (h', r, t'), respectively; where h and h' are the head entity vectors of the positive and negative samples, r is the entity relationship, and t and t' are the tail entity vectors of the positive and negative samples.
4. The method according to claim 1, characterized in that: In step 5, the associated recommendation of the firepower strike target is realized; the specific implementation method is: the target system value includes the vector weight value and edge weight value of the target; let the entity vector contained in the i-th sub-target entity be h i , after completing the network, the set of edges associated with this sub-goal is S i , then the system value Z of the i-th goal i for:
5. A joint strike target adaptive scoring and association recommendation system, characterized in that: include: Target knowledge graph super network construction module: Based on the acquisition of target intelligence data, the sub-network relationship is annotated using the relationship between targets to build a target knowledge graph super network structure model; Knowledge graph initial reasoning vector generation module: maps entity relationships into spatial vectors, that is, assigns a random vector array to each entity and its relationship; Knowledge graph initial reasoning vector optimization module: Use the group evolution idea of the meta-heuristic algorithm to achieve iterative optimization of entity and relationship vectors; through a finite number of iterations based on stochastic gradient descent, the loss function containing the scoring function converges to the minimum value; specifically includes: Calculate the hypernetwork loss function; the goal of the algorithm is: h+r=t; h'+r≠t', which can be accurately measured by constructing a scoring function f(h,r,t), and the calculation formula is: f(h,r,t)=||h+rt|| The calculation formula of the loss function L(h,r,t,r',t') is: Where S represents the existing entity relationship set of the hypernetwork; S' represents the unassociated entity relationship set of the hypernetwork, satisfying the following conditions: Construct an agent structure; each agent contains an entity vector set, an edge vector set, and a negative sample edge vector set; Construct the initial agent population; adopt the random mutation method of the meta-heuristic algorithm, take the original agent as the sample, and randomly select 10 entities and edges for vector fine-tuning; r represents the vector relationship, δ represents the fine-tuning parameter, ω represents the original parameter in the vector, and the calculation formula of the fine-tuned parameter ω' is: Calculate the comprehensive score of all agents in the population; according to the updated loss function scoring formula, bring the entity vector set, edge vector set, and negative sample edge vector set contained in each agent into the scoring formula to calculate the minimum score L of the existing edges min and the maximum score of negative samples L max ; The specific scoring principle is: L min As the main score of the agent; if the two agents L min Same, L max As an agent to assist scoring; Population elimination: According to the comprehensive scores of each agent in the population, 90% of the agents in the population are eliminated, and the best scoring agent structure and corresponding scores are recorded; Population expansion: calling the random mutation method in the above steps to generate new agents using the remaining agents after elimination as samples, so that the population reaches the original size; Repeat the above steps until the exit condition is met: L min Minimize, and L max If it stops growing, the algorithm is considered to have converged, and it exits and feeds back the best agent in the population at this time; Target knowledge graph completion module: Use the optimized vector of the optimal agent to count the comprehensive scores of all entity relationships, and compare them with the average scores of existing relationships in the hypernetwork to discover hidden relationships in the hypernetwork and complete the hypernetwork; After obtaining the optimal entity and edge vector matrix of the hypernetwork, the hypernetwork relationship completion operation is introduced: collect the triples (h, r, t) in the existing hypernetwork, replace their head entities or tail entities with other entities in turn, and then generate a set of candidate relationship arrays, calculate their evaluation values f(h', r, t) or f(h, r, t'), and then sort them in ascending order according to the evaluation values, and record the sorting position of the original triples (h, r, t); if the sorting exceeds 5, the first 5 triple relations are added to the hypernetwork; otherwise, the relations whose sorting exceeds the original triples are added to the hypernetwork; the attribute array of the newly added relationship satisfies the formula: r = ht; Fire strike target association recommendation module: calculates the value scores of all entities in the knowledge graph hypernetwork, and determines the association recommendation target names between entities based on the hypernetwork relationships.
6. A computer-readable storage medium, characterized in that: The storage medium stores a plurality of instructions; the plurality of instructions are used by a processor to load and execute the method of any one of claims 1 to 4.
