Methods and devices for constructing, computing, and sensing spatiotemporal knowledge graphs of battlefield situation

Abstract

This invention discloses a method and device for constructing, calculating, and perceiving a spatiotemporal knowledge graph of battlefield situation, belonging to the field of battlefield situation assessment. This invention avoids traditional target-type elements with complex behaviors and obscure patterns, choosing spatiotemporal nodes as the learning subject. It abstracts the uncertain set of entities in the existing knowledge graph into a definite set of spatiotemporal nodes, reducing the difficulty of summarizing and learning the patterns of battlefield situation and strengthening the grasp of the battlefield situation. When the entity and relation sets in the graph are uncertain, and the truth or falsehood of all newly added knowledge is unknown, this invention estimates the confidence level of the newly added knowledge at different levels, and finally integrates multiple calculation results to judge the truth or falsehood of the newly added knowledge. This supports the updating and iteration of the spatiotemporal situation knowledge graph, thereby ensuring data purity, strengthening the summarization, deduction, and prediction of battlefield situation, and greatly simplifying computational complexity while ensuring judgment accuracy.

Description

Technical Field

[0001] This invention belongs to the field of military information fusion and battlefield situation assessment. More specifically, it relates to methods and devices for constructing, calculating, and sensing a spatiotemporal knowledge graph of the battlefield situation. Background Technology

[0002] Situation assessment refers to perceiving elements in the environment within a certain time and space, understanding their meaning, and predicting their state in the near future. Battlefield situation assessment is the process of dynamically evaluating the situational elements of both sides on the battlefield. By comprehensively considering situational elements such as the distribution and activities of enemy and friendly forces, weapon systems, mobility, battlefield geographical environment, and weather, it identifies and analyzes past events and plans, determines estimates of the enemy's force structure, deployment, direction of action, and routes, identifies the enemy's combat style, infers the enemy's combat intentions, provides a reasonable interpretation of the current battlefield situation, and predicts future changes in the situation, ultimately establishing a comprehensive battlefield situation map organized by combat activities, events, time, location, and force elements.

[0003] In current mainstream research, in order to understand, analyze, and predict the situation, it is necessary to express the situation in the form of a situation knowledge graph. Existing situation knowledge graphs are essentially databases that store situation knowledge in the form of triples. For example, in the war scenario "fighter A attacks fighter B", fighter A, fighter B, and the attack constitute a situation knowledge, represented as a triple [target A, target B, and target A's intention towards target B].

[0004] To discover errors in situational knowledge graphs, improve their quality, and ultimately enhance the performance of knowledge-driven learning tasks, the concept of situational knowledge graph triple trustworthiness has been introduced. Situational knowledge graph triple trustworthiness measures the degree of truthfulness of the knowledge expressed by the triple. The value of situational knowledge graph triple trustworthiness ranges from [0,1]. The closer the value is to 0, the higher the probability that the triple is incorrect; conversely, the closer the value is to 1, the higher the probability that the triple is true.

[0005] Based on the applicable stage of knowledge graph triple confidence evaluation methods, existing methods can be summarized into three categories: The first category is used for the process of "extracting triples from text data," with a typical example being the KnowLife knowledge base from the Max Planck Center for Information Research in Germany. The second category is used for the embedding process, which aims to encode all entities and relations into a continuous vector space. Optimizing confidence and removing data noise during the embedding process has become a research hotspot in recent years, with typical methods including SCEF, TransE, and TransT. The third category directly evaluates the triples, measuring the reliability of triples obtained through knowledge reasoning. It is also applicable to the confidence evaluation of dynamic knowledge bases, with typical methods including KGTTM and CTransE.

[0006] However, the above methods have some problems that need to be overcome: (1) The complete set of entities in the knowledge graph is not known in advance, and the addition of new entities may also bring new relationships. Many unpredictable situations will make the previously trained models unreliable or even unusable; (2) When the relationship between entities is not fixed, but changes with time and space, the labeling of entities is temporary. That is, it is difficult to find a stable set of knowledge as a training set in the spatiotemporally related knowledge graph; (3) Most of the existing judgment, prediction, and completion models or algorithms assume that all the knowledge in the graph is real knowledge and ignore the noise influence of false knowledge. A few algorithms have a process of judging the authenticity of knowledge, but it is only compared with real data. Summary of the Invention

[0007] In view of the shortcomings of the prior art, the purpose of this invention is to provide a method and device for constructing, calculating and perceiving a spatiotemporal knowledge graph of battlefield situation, aiming to solve the problems that existing battlefield knowledge graphs cannot handle newly added entities and conflicts in the relationships between entities, and that existing battlefield situation perception methods are interfered with by noise and conflict information.

[0008] To achieve the above objectives, in a first aspect, the present invention provides a method for constructing a spatiotemporal knowledge graph of battlefield situation, comprising:

[0009] Acquire situational knowledge of all scenarios in the target battlefield. The situational knowledge is a triple [head target, relationship, tail target]. The target is the combat unit appearing in the scenario. The head target and tail target are of the same type. The relationship is the intention of the head target towards the tail target.

[0010] For each scenario of the target battlefield, the following processing is performed to obtain a spatiotemporal knowledge graph of the battlefield situation:

[0011] 1) Obtain the type, time slice, and spatial block of the head and tail targets in the situational knowledge of the scene respectively;

[0012] 2) Encode the type, time slice, and spatial block of the head and tail targets respectively, and encode the relationship. The encoding methods for the head and tail targets are the same.

[0013] 3) Construct the spatiotemporal situational knowledge of the scene. The spatiotemporal situational knowledge is a triple [head entity code, relation code, tail entity code]. The entity code consists of the time slice code, spatial block code, and type code of the node.

[0014] Preferably, after constructing the knowledge graph, the method further includes:

[0015] When the spatiotemporal situational knowledge of the target battlefield changes, the confidence level of the spatiotemporal situational knowledge of that scenario is calculated and used as the basis for whether to update it into the situational knowledge graph: spatiotemporal situational knowledge that is judged to be true is added to the spatiotemporal situational knowledge graph; spatiotemporal situational knowledge that is judged to be false is discarded; spatiotemporal situational knowledge that cannot be judged is temporarily stored for reference and marked for record.

[0016] Preferably, the complete set of the relationships is: attack, defense, support, group, reconnaissance, and the complete set of the combat units is: equipment, personnel, organization.

[0017] To achieve the above objectives, in a second aspect, the present invention provides a method for calculating the confidence level of spatiotemporal situational knowledge for battlefield situational awareness, comprising:

[0018] T1. Based on the spatiotemporal knowledge graph of the battlefield situation, construct a directed graph with the head entities in the spatiotemporal knowledge graph as the root nodes. The spatiotemporal knowledge graph of the battlefield situation is constructed using the method described in the first aspect.

[0019] T2. Different relationships are divided into different levels, and different weight parameters RW are assigned to them according to the level during the calculation process;

[0020] T3. Iterate through the resource flow in the directed graph until it converges, and calculate the resource retention value of the tail entity in the spatiotemporal situation knowledge by combining the weight parameter RW.

[0021] T4. Construct the feature vector of this spatiotemporal situational knowledge = [R(t|h), ID(h), OD(h), ID(t), OD(t), Dep(h,t), RW], where R(t|h) is the resource reservation value of the tail entity, ID(h) is the in-degree of the head entity, OD(h) is the out-degree of the head entity, ID(t) is the in-degree of the tail entity, OD(t) is the out-degree of the tail entity, and Dep(h,t) is the depth from the head entity to the tail entity;

[0022] T5. The feature vector V, after being processed by the activation function, is transformed into the situational knowledge confidence level of the target entity.

[0023] Preferably, the formula for calculating the resource retention value of the tail entity is as follows:

[0024]

[0025] Where θ represents the probability that the resource flow of each entity node will directly jump to a random node, and M represents the probability of M. t Let R(e) be the set of all nodes leading to the tail entity node t. i |h) is the entity node e i Resource retention value, For entity node e i The bandwidth to the tail node t, OD(e i ) is the entity node e i The out-degree, N is M t Number of nodes in the middle.

[0026] Preferably, the confidence score RR(h,t) of the situational knowledge at the target entity level is calculated using the following formula:

[0027]

[0028] Where α is the nonlinear activation function, u is the nonlinear activation result, and W i and b i Let be the parameter matrix that can be adjusted during the training phase, where ii∈{1,2}.

[0029] To achieve the above objectives, in a third aspect, the present invention provides a battlefield situational awareness method based on cross-neural networks and spatiotemporal knowledge graphs, comprising:

[0030] The system receives new spatiotemporal situational knowledge of the target battlefield scenario, inputs it into a trained situational knowledge judgment model, and obtains the confidence level of the new spatiotemporal situational knowledge.

[0031] The situational knowledge assessment model is trained in the following way:

[0032] S1. Establish a sample data set: Based on the situational knowledge obtained from the battlefield spatiotemporal knowledge graph, all scenario situations are positive samples, and corresponding negative samples are constructed according to certain principles. Each sample is saved as a sample set in the form of entity encoding, and the overall sample set is divided into a training set, a test set, and a verification set. The spatiotemporal knowledge graph of the battlefield situation is constructed using the method described in the first aspect.

[0033] S2. Split the encoding of sample data: Separate the spatiotemporal encoding and type encoding in the entity encoding of the training set and save them separately.

[0034] S3. Generate word vectors for each sample data;

[0035] S4. Calculate the confidence level of situational knowledge at the target entity level for each sample;

[0036] S5. Calculate the confidence level of situational knowledge at the situational knowledge graph level for each sample.

[0037] S6. Training the cross-neural network: Link the output results of step S4 and step S5 together as input and input them to the fusion unit based on multilayer perceptron. The output layer of the fusion unit is a binary classifier. Assign the label y=1 to the positive sample and y=0 to the negative sample. Calculate the probability that the tested sample is true under the current input conditions. The result is a number between (0,1).

[0038] S7. When the prediction accuracy of the situation knowledge judgment model reaches the preset threshold, stop the optimization process and obtain the trained situation knowledge judgment model.

[0039] Preferably, step S4 employs the method described in the second aspect.

[0040] To achieve the above objectives, in a fourth aspect, the present invention provides a computing device, comprising: a processor and a memory;

[0041] The processor is used to store computer execution instructions;

[0042] The processor is used to execute the computer execution instructions, causing the above method to be performed.

[0043] To achieve the above objectives, in a fifth aspect, the present invention provides a computer-readable storage medium for storing a computer program, wherein the above-described method is performed when the computer program is executed.

[0044] In summary, the technical solutions conceived by this invention have the following beneficial effects compared with the prior art:

[0045] (1) In view of the problems that existing knowledge graphs are based on the complete set of entities and the complete set of relations, which are difficult to update and iterate dynamically and cannot handle the addition of new entities and the conflict of previous and subsequent relations between entities, this invention proposes a method for constructing a spatiotemporal knowledge graph of battlefield situation. It avoids the traditional target-type elements with complex behavior and obscure rules, and selects spatiotemporal nodes as the learning subject. It abstracts the uncertain complete set of entities in the existing knowledge graph into a definite complete set of spatiotemporal nodes to support the real-time evolution of spatiotemporal related situation knowledge, reduces the difficulty of summarizing and learning the rules of battlefield situation, and strengthens the grasp of battlefield situation.

[0046] (2) This invention proposes a method for calculating the confidence of spatiotemporal situational knowledge for battlefield situational awareness. It simulates the process of fluid resources flowing through pipes of different thicknesses to multiple interconnected reservoirs of different depths and eventually being distributed stably. It proposes the concept of the strength of the relationship between entities, and uses different weights to represent relationships of different strengths, thereby increasing the credibility and feasibility of the confidence calculation results.

[0047] (3) To enhance the understanding of the battlefield situation and reduce the interference of noise and conflict information on battlefield situation perception, this invention proposes a battlefield situation perception method based on cross-neural networks and spatiotemporal knowledge graphs. When the complete sets of entities and relations in the graph are uncertain, and the truth or falsehood of all newly added knowledge is unknown, the confidence level of the newly added knowledge is estimated at different levels. Finally, by integrating multiple calculation results, the authenticity of the newly added knowledge is judged to support the updating and iteration of the spatiotemporal situation knowledge graph, thereby ensuring data purity and strengthening the summarization, deduction, and prediction of the battlefield situation. While ensuring the accuracy of the judgment, the computational complexity is greatly simplified. Attached Figure Description

[0048] Figure 1 The flowchart of a battlefield situation awareness method based on cross-neural networks and spatiotemporal knowledge graphs provided by this invention is shown.

[0049] Figure 2 A flowchart illustrating a method for constructing a spatiotemporal knowledge graph of battlefield situation provided by this invention.

[0050] Figure 3 The flowchart illustrates a method for calculating the confidence level of spatiotemporal situational knowledge for battlefield situational awareness, as provided by this invention.

[0051] Figure 4 This is a schematic diagram of a cross-neural network structure provided in an embodiment of the present invention. Detailed Implementation

[0052] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0053] Figure 1 This invention provides a flowchart of a battlefield situation awareness method based on cross-neural networks and spatiotemporal knowledge graphs. Figure 1 As shown, the method includes:

[0054] The system receives new spatiotemporal situational knowledge of the target battlefield scenario, inputs it into a trained situational knowledge judgment model, and obtains the confidence level of the new spatiotemporal situational knowledge.

[0055] The situational knowledge assessment model is trained in the following way:

[0056] Step S1: Establish a sample data set.

[0057] Based on the situational knowledge obtained from the battlefield spatiotemporal knowledge graph, all scenario situations are treated as positive samples. Corresponding negative samples are constructed according to certain principles. Each sample is saved as a sample set in the form of entity encoding. The overall sample set is divided into training set, test set and validation set.

[0058] In this embodiment, the situational awareness of a classic naval battlefield scenario is used as the experimental data. The default situational awareness knowledge is considered real-world situational awareness, i.e., positive samples. Correct situational awareness knowledge generated by the situational assessment system is collected, labeled "correct," with a quantity of no less than 10,000, as the positive sample set. Then, negative samples are generated by randomly selecting one relation from the entire intent set according to a uniform distribution to replace a relation in the positive samples, or randomly selecting one tail node from the entire target set according to a uniform distribution to replace a tail node in the positive samples, labeled "incorrect," with a quantity of no less than 10,000, as the negative sample set. The positive and negative sample sets are merged, and the target A type code, target A spatiotemporal code, target B type code, target B spatiotemporal code, and relation code for each situational awareness are calculated using relational encoding, type encoding, and spatiotemporal encoding methods. These codes for each situational awareness are concatenated into an entity code in the order of target A type code, target A spatiotemporal code, target B type code, target B spatiotemporal code, and relation code, and saved with the labels to form a sample data set. This set is then randomly divided into a training set, a test set, and a validation set using a uniform distribution.

[0059] This invention selects time slices and spatial blocks as the main body for knowledge learning judgment, rather than traditional target-type elements. Figure 2 A flowchart illustrating a method for constructing a spatiotemporal knowledge graph of battlefield situation provided by this invention. (For example...) Figure 2 As shown, the method includes:

[0060] Acquire situational knowledge of all scenarios in the target battlefield. The situational knowledge is a triple [head target, relationship, tail target]. The target is the combat unit appearing in the scenario. The head target and tail target are of the same type. The relationship is the intention of the head target towards the tail target.

[0061] For each scenario of the target battlefield, the following processing is performed to obtain a spatiotemporal knowledge graph of the battlefield situation:

[0062] 1) Obtain the type, time slice, and spatial block of the head and tail targets in the situational knowledge of the scene respectively;

[0063] 2) Encode the type, time slice, and spatial block of the head and tail targets respectively, and encode the relationship. The encoding methods for the head and tail targets are the same.

[0064] 3) Construct the spatiotemporal situational knowledge of the scene. The spatiotemporal situational knowledge is a triple [head entity code, relation code, tail entity code]. The entity code consists of the time slice code, spatial block code, and type code of the node.

[0065] For example, in the traditional definition of a scenario where fighter jet A attacks fighter jet B, fighter jet A, fighter jet B, and the attack form a triple as the head node, tail node, and relation, respectively. However, in this invention, it is understood as: located at time slice T A Space block P A The fighter jet A belongs to the TP type A For time slice T B Space block P B The fighter jet B belongs to the TP type. B If an attack is launched, then at this time, "T" A P A TP A “T” B P B TP B "Attack", forming a triple [T] A P A TP A Attack, T B P B TP B The present invention can learn from this knowledge: [At T] A P A TP A With T B P B TP B Or, it may be located in another spacetime location, but at a relative distance from T. A P A and T B P B TPs that are relatively close to each other A and TP B An "attack" relationship can occur, and the two specific targets involved are not necessarily fighter jet A and fighter jet B; they could also be fighter jet C, which is of the same type as fighter jet A, and fighter jet D, which is of the same type as fighter jet B. A target refers to a combat unit appearing in the training set, such as equipment, personnel, or organization. The complete target set refers to the set of all targets appearing in the training set. A group refers to a series of targets performing the same task and forming a group relationship with each other.

[0066] First, spatiotemporal codes for each target are generated based on the situation in a typical scenario. Then, type codes are added, determined by the order in which the target's type first appears, to generate complete entity codes. Finally, codes for each relation are determined based on the order in which different relations first appear. The specific coding method is as follows:

[0067] Time encoding method: Based on the standard timecode format, specifically: YYYY-MM-DD hh:mm:ss:ff (year, month, day, hour, minute, second, and frame number). The first six data points represent the Chinese standard time, where the year is represented by a 13-bit binary code, the month by a 4-bit binary code, the day and hour each by a 5-bit binary code, and the minute and second each by a 6-bit binary code. The seventh data point is the video frame number, represented by a 17-bit binary code. In this embodiment, the frame number is not involved; it is filled with 011111111111111111 when generating the time code. During calculation, the year, month, day, hour, minute, and second codes are extracted and calculated separately, ignoring the last 16 bits.

[0068] Spatial coding method: The longitude, latitude, and altitude of the target are encoded into 32-bit codes according to the Earth Spatial Grid Coding Rules (GB / T40087-2021) standard. In this embodiment, the first 24 bits are taken as the spatial code. Each level of code is represented by a single binary digit. Therefore, each level of spatial code is represented by an octal number.

[0069] Type encoding method: Arrange all types of targets into a sequence, such as fighter jets, bombers, early warning aircraft, transport aircraft, other types of aircraft, destroyers, frigates, transport ships, support ships, aircraft carriers, amphibious assault ships, other types of ships, conventional submarines, attack nuclear submarines, strategic nuclear submarines, other types of submarines, tanks, armored vehicles, transport vehicles, missile launchers, other types of vehicles, and other types of targets, and encode them in decimal by starting from 0 and incrementing by 1.

[0070] Relational Encoding Method: In this embodiment, the relationships in the knowledge graph are intentions, including target-to-target intentions, target-to-group intentions, group-to-target intentions, and group-to-group intentions. The complete set of intentions in this embodiment is: attack, defense, support, group, and reconnaissance. The encoding method for intentions is the same as that for target types, using the order of their first appearance and encoding them in decimal, starting from 0 and incrementing by 1.

[0071] For example, at 10:50:28 on June 15, 2022, an enemy Arleigh Burke-class destroyer appeared within our detection range, with latitude and longitude coordinates of (120.38390468507936, 14.229961024930361). Its altitude is calculated based on the difference in elevation between its location and the equipotential surface of the earth. Therefore, the complete time code for this target is as follows:

[0072] 00111111001100110011110101011001001110001111111111111111.

[0073] In the calculation, the first 40 bits of the code are used in this invention.

[0074] The complete spatial encoding of the target is as follows:

[0075] Longitude: 00111100001011100001000001110100;

[0076] Latitude: 00000111000110110111111011100000;

[0077] Height: 1111111111111111111111111010011.

[0078] The Earth Spatial Grid Coding Rules (GB / T 40087-2021) standard uses Morton cross-coding to achieve spatial continuity of the codes. In this invention, the first 24 bits of each code are used in the calculation.

[0079] In this training sample set, the target's type, "Arleigh Burke class," is coded as 4.

[0080] Linking the above three parts together, the entity code of the target is: 3f333d593892dbf692ef7e9b7db4 4.

[0081] Preferably, after constructing the knowledge graph, the method further includes:

[0082] When the spatiotemporal situational knowledge of the target battlefield changes, the confidence level of the spatiotemporal situational knowledge of that scenario is calculated and used as the basis for whether to update it into the situational knowledge graph: spatiotemporal situational knowledge that is judged to be true is added to the spatiotemporal situational knowledge graph; spatiotemporal situational knowledge that is judged to be false is discarded; spatiotemporal situational knowledge that cannot be judged is temporarily stored for reference and marked for record.

[0083] Preferably, the complete set of the relationships is: attack, defense, support, group, reconnaissance, and the complete set of the combat units is: equipment, personnel, organization.

[0084] Step S2: Split the encoding of the sample data.

[0085] The spatiotemporal encoding and type encoding in the entity encoding of the training set are separated and saved separately. In this embodiment, the 24-level spatial encoding and type encoding in the entity encoding of the training set are separated, and the relation encoding is copied 25 times. These copies are saved together with the corresponding copied relation encodings as separate training data.

[0086] Step S3: Generate word vectors for each sample data.

[0087] The spatial, type, and relational codes of all situational knowledge in the training set are calculated. The translation-based energy function algorithm (TEF) is used, based on the transE algorithm published by Bordes et al. at NIPS in 2013. Word vectors are generated for each level of spatial, type, and relational codes according to the split training set.

[0088] The formula for calculating the vector dimension is: n≥8*ln(N), where n is the vector dimension, N is the number of words, and in this embodiment, N is the sum of the number of entities and relations that may appear at each level of encoding.

[0089] From the above calculation formula, each level of spatial encoding should be represented by a 24-dimensional vector, calculated from the number of all possible target types in the scene. Type encoding should be represented by a 51-dimensional vector. Finally, based on the complete training set, the vectors from each segment are combined into a 627-dimensional entity vector or relation vector.

[0090] Step S4: Calculate the situational knowledge confidence level of each sample at the target entity level.

[0091] Figure 3 This invention provides a flowchart of a method for calculating the confidence level of spatiotemporal situational knowledge for battlefield situational awareness. (See flowchart for example.) Figure 3 As shown, the method includes the following steps:

[0092] T1. Based on the spatiotemporal knowledge graph of the battlefield situation, construct a directed graph with the head entities in the spatiotemporal knowledge graph as the root nodes. The spatiotemporal knowledge graph of the battlefield situation is constructed using the method described above.

[0093] T2. Different relationships are divided into different levels, and different weight parameters RW are assigned according to the level during the calculation process.

[0094] If the relationship between entity pairs (h, t) is strong, a significant amount of resources will flow from the head entity h through all associated paths to the tail entity t in the knowledge graph. Based on this, this invention categorizes different relationships into different levels, each representing a different weight parameter during the calculation process. For example, in this embodiment, attack and defense are the highest-level relationships, reconnaissance is the second level, grouping is the third level, and support is the fourth level. The higher the level, the larger the weight parameter value.

[0095] T3. Iterate through the resource flow in the directed graph until it converges, and calculate the resource retention value of the tail entity in the spatiotemporal situational knowledge by combining the weight parameter RW.

[0096] Preferably, the formula for calculating the resource retention value of the tail entity is as follows:

[0097]

[0098] Where θ represents the probability that the resource flow of each entity node will directly jump to a random node, and M represents the probability of M. t Let R(e) be the set of all nodes leading to the tail entity node t. i |h) is the entity node e i Resource retention value, For entity node e i The bandwidth to the tail node t, OD(e i ) is the entity node e i The out-degree, N is M t Number of nodes in the middle.

[0099] T4. Construct the feature vector of this spatiotemporal situational knowledge:

[0100] V=[R(t|h), ID(h), OD(h), ID(t), OD(t), Dep(h, t), RW]

[0101] Where R(t|h) is the resource reservation value of the tail entity, ID(h) is the in-degree of the head entity, OD(h) is the out-degree of the head entity, ID(t) is the in-degree of the tail entity, OD(t) is the out-degree of the tail entity, and Dep(h,t) is the depth from the head entity to the tail entity.

[0102] T5. The feature vector V, after being processed by the activation function, is transformed into the situational knowledge confidence level of the target entity.

[0103] Preferably, the confidence score RR(h,t) of the situational knowledge at the target entity level is calculated using the following formula:

[0104]

[0105] Where α is the nonlinear activation function, u is the nonlinear activation result, and Wi and b i Let be the parameter matrix that can be adjusted during the training phase, i∈{1,2}.

[0106] The range of RR(h,t) values ​​is between [0,1]. The closer the value is to 1, the more likely there is a relationship between h and t.

[0107] Step S5: Calculate the confidence level of situational knowledge at the situational knowledge graph level for each sample.

[0108] Using the reachable paths inference algorithm, in a directed graph, the semantic distance between the target triple and different paths from the target head node to the target tail node is calculated using the entity relation vector calculated in step S3. The three paths with the largest semantic distances are selected, and the output h of the top three paths is used. t The vectors are linked together and then subjected to nonlinear processing to obtain the value RP((h, r, t)), which represents the confidence level of the target decision related to the triples in the knowledge graph.

[0109] The specific steps are as follows:

[0110] 1) Given a triple (h, r, t).

[0111] 2) Find all reachable paths P(h, r, t) = {p1, p2...p...} from the head node h to the tail node t in the knowledge graph. n}

[0112] 3) For each p i ∈P(h, r, t), calculate p respectively i The average semantic distance between the head node of all triples and h, the average semantic distance between the tail node of all triples and t, and the average semantic distance between the relation node of all triples and r are given. Here, the semantic distance can be referred to as the L1 norm or L2 norm.

[0113] 4) Take the average of the three semantic distances from step 3) as the current path p. i Semantic distance.

[0114] 5) Extract the three paths with the largest semantic distance in P.

[0115] 6) Output h of the top three paths t Link them together to form a vector S, and then perform nonlinear processing on vector S to obtain the value RP((h, r, t)):

[0116]

[0117] Where β is a nonlinear activation function, and Wj and b j This is a parameter matrix that can be adjusted during model training, j∈{3,4}. The value of RP((h,r,t)) ranges between [0,1]. The closer the value is to 1, the higher the confidence that the triple (h,r,t) is a correct triple.

[0118] Step S6: Train the cross neural network.

[0119] The outputs of step S4 and step S5 are linked together to obtain the feature vector f(s) = [RR(h,t), RP((h,r,t))], which is then input into a multilayer perceptron-based fusion processor. The output layer of the fusion processor is a binary classifier that assigns the label y = 1 to positive samples and y = 0 to negative samples. The probability that the tested sample is true under the current input conditions is calculated, and the result is a number in the interval (0,1).

[0120] f(s) is input into the fusion processor and transformed through multiple hidden layers. A binary classification output layer assigns true and false labels y to each triple. Finally, the sigmoid function is used to calculate p(y = 1 | f(s)).

[0121]

[0122] Among them, h i It is the i-th hidden layer. and W is the parameter matrix that the i-th hidden layer will be learned and trained. o and b o It is the parameter matrix of the output layer.

[0123] Figure 4 This is a schematic diagram of a cross-neural network structure provided in an embodiment of the present invention. Figure 4 The training process of the cross-neural network shown is as follows: First, the sample set is randomly divided into a training set, a test set, and a validation set in a ratio of 8:1:1. Then, the cross-neural network is trained using the Adam optimization method with 128 samples per batch. Here, the default parameters of the Adam method are used for training.

[0124] The Adam method is a first-order optimization algorithm that can replace the traditional stochastic gradient descent (SGD) process. Its default learning rate is set to 0.001.

[0125] Step S7: When the prediction accuracy of the situation knowledge judgment model reaches the preset threshold, stop the optimization process and obtain the trained situation knowledge judgment model.

[0126] In this embodiment, the preset threshold is 99%.

[0127] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for calculating the confidence level of spatiotemporal situational knowledge for battlefield situational awareness, characterized in that, include: T1. Based on the spatiotemporal knowledge graph of the battlefield situation, construct a directed graph with the head entities in the spatiotemporal situation knowledge as the root nodes; T2. Divide different relationships into different levels, and assign different weight parameters during the calculation process according to the level. ; T3. Iterate through the resource flows in the directed graph until they converge, and then combine the weight parameters. Calculate the resource retention value of the tail entity in this spatiotemporal situational knowledge; T4. Construct the feature vector of this spatiotemporal situational knowledge. ,in, The resource reservation value for the tail entity. The in-degree of the head entity. For the out-degree of the head entity, The in-degree of the tail entity. For the out-degree of the tail entity, The depth from the head entity to the tail entity; T5. Eigenvectors After processing by the activation function, it is transformed into the situational knowledge confidence level at the target entity level; The formula for calculating the resource retention value of tail entities is as follows: in, The probability that the resource flow of each entity node will directly jump to a random node. For all nodes leading to the tail entity The set of nodes, For entity nodes Resource retention value, To the entity node Ending entity node bandwidth, For entity nodes The degree of exit, for Number of nodes; Confidence of situational knowledge at the target entity level The calculation formula is as follows: in, It is a non-linear activation function. This is the result of nonlinear activation. and The parameter matrix is ​​adjustable during the training phase. .

2. A battlefield situation awareness method based on cross-neural networks and spatiotemporal knowledge graphs, characterized in that, include: The system receives new spatiotemporal situational knowledge of the target battlefield scenario, inputs it into a trained situational knowledge judgment model, and obtains the confidence level of the new spatiotemporal situational knowledge. The situational knowledge assessment model is trained in the following way: S1. Establish a sample data set: Based on the situational knowledge obtained from the battlefield spatiotemporal knowledge graph, all scenario situations are positive samples. Corresponding negative samples are constructed according to certain principles. Each sample is saved as a sample set in the form of entity encoding. The overall sample set is divided into a training set, a test set, and a validation set. S2. Split the encoding of sample data: Separate the spatiotemporal encoding and type encoding in the entity encoding of the training set and save them separately; S3. Generate word vectors for each sample data; S4. Using the calculation method described in claim 1, calculate the situational knowledge confidence level of each sample at the target entity level; S5. Calculate the confidence level of situational knowledge at the situational knowledge graph level for each sample. S6. Training the cross-neural network: Link the output results of step S4 and step S5 together as input and input them to the fusion unit based on multilayer perceptron. The output layer of the fusion unit is a binary classifier. Assign the label y=1 to the positive sample and y=0 to the negative sample. Calculate the probability that the tested sample is true under the current input conditions. The result is a number between (0,1). S7. When the prediction accuracy of the situation knowledge judgment model reaches the preset threshold, stop the optimization process and obtain the trained situation knowledge judgment model.

3. A computing device, characterized in that, include: Including processor and memory; The processor is used to store computer execution instructions; The processor is used to execute the computer execution instructions, causing the method of claim 1 or 2 to be executed.

4. A computer-readable storage medium, characterized in that, The computer-readable storage medium is used to store a computer program, which, when executed, performs the method of claim 1 or 2.

Images