A kill chain construction method based on simulation and deduction combined with virtual and real

Abstract

The present invention provides a method for constructing a kill chain based on a simulation deduction based on the combination of virtual and real, which relates to the field of simulation technology, and comprises: S1: obtaining a combat mission, identifying and defining a target according to the requirements of the combat mission, and determining the type of the kill chain based on the combat mission; S2: collecting corresponding scene data for different combat domains, and constructing a three-dimensional digital virtual environment based on the scene data; S3: establishing a multi-dimensional physical model for various types of combat platforms, and rendering the multi-dimensional physical model in real time into a virtual environment to form a virtual combat unit, wherein the combat platform comprises a plurality of sensors; S4: collecting sensor data of each combat platform, fusing the sensor data, and generating a unified target situation map; S5: making a plan decision by using a decision algorithm according to the target situation map and a combat knowledge base, and generating a combat plan; S6: deducing and verifying the combat plan in a virtual environment, and evaluating and optimizing the combat plan by using an evaluation method, and obtaining an optimized combat plan; S7: converting the optimized combat plan into an executable instruction parameter set, and transmitting the instruction parameter set to each combat platform through a heterogeneous network to coordinate and dispatch combat; S8: forming a kill chain according to the process of steps S1-S7. The present invention provides a reliable kill chain construction method for different combat domains, which not only improves the flexibility and efficiency of operations, but also lays a solid foundation for the development of future combat fields.

Description

Technical Field

[0001] The present invention relates to the field of simulation technology, and in particular to a kill chain construction method based on simulation deduction combining virtuality and reality. Background Art

[0002] The kill chain is a product of information warfare, usually referring to the closed-loop process of detecting targets, aiming at targets, engaging the enemy, and evaluating the results of the engagement. As a combat concept and theory, the kill chain plays an important role in combat operations and the entire combat process. In combat operations, whoever can complete the construction and implementation of the kill chain more quickly can seize the initiative and make decisions and take actions before the opponent. This can not only change the situation previously observed by the opponent, forcing it into a decision-making dilemma, thereby seizing the initiative on the battlefield, but also effectively interfere with the opponent's decision-making process, thereby forming a greater combat advantage.

[0003] With the development of human civilization, large-scale wars are increasingly restricted and constrained, military operations are increasingly developing in the form of local wars and military conflicts, and weapons and equipment are moving towards unmanned and intelligent. The kill chain built around large missiles, advanced fighters, aircraft carriers, etc. in the past is often bound and restricted by multiple factors. Therefore, it is particularly important to accelerate the construction of a kill chain based on combat platforms, which can not only effectively improve the flexibility and efficiency of combat, but also take the lead in future competition and provide strong support for winning the war. Summary of the invention

[0004] In view of this, the present invention provides a simulation deduction kill chain construction method based on the combination of virtual and real. By combining the combat elements of the real world with the simulation technology of the virtual environment, the combination of virtual and real not only improves the accuracy and efficiency of the deduction, but also demonstrates its unique value in constructing the kill chain in the simulation deduction.

[0005] The technical purpose of the present invention is achieved in this way:

[0006] The present invention provides a method for constructing a kill chain based on a simulation of combining virtuality and reality, comprising the following steps:

[0007] S1: Obtain combat missions, identify and define targets according to combat mission requirements, and determine the type of kill chain based on combat missions;

[0008] S2: Collect corresponding scenario data for different combat domains and build a three-dimensional digital virtual environment based on the scenario data;

[0009] S3: For each type of combat platform, a multi-dimensional physical model is established, and the multi-dimensional physical model is rendered in real time into a virtual environment to form a virtual combat unit, where the combat platform includes multiple sensors;

[0010] S4: Collect sensor data from each combat platform, fuse the sensor data, and generate a unified target situation map;

[0011] S5: Based on the target situation map and combat knowledge base, use the decision-making algorithm to make a plan decision and generate a combat plan;

[0012] S6: Perform simulation and verification of the combat plan in a virtual environment, evaluate and optimize the combat plan using evaluation methods, and obtain an optimized combat plan;

[0013] S7: Convert the optimized combat plan into an executable instruction parameter set, and transmit the instruction parameter set to each combat platform through a heterogeneous network for coordinated dispatch of operations;

[0014] S8: According to the process of steps S1-S7, a kill chain is formed.

[0015] Based on the above technical solution, preferably, step S4 includes:

[0016] S41: Collect data from multiple sensors and perform spatiotemporal calibration on the data from the multiple sensors;

[0017] S42: using a data association algorithm to match data of the same target observed by different sensors to form a target data set;

[0018] S43: using a fusion algorithm to perform intra-set fusion on each target data set to obtain fused data of each target;

[0019] S44: Convert the fused data into target information, and generate a target situation map according to the target information.

[0020] Based on the above technical solution, preferably, step S43 includes:

[0021] S431: Each sensor is based on the observed data and the nonlinear observation model h i (x) Update its local information matrix and information vector:

[0022]

[0023]

[0024] In the formula, I i is the local information matrix, y i is the information vector, R i is the observation noise covariance, H i is the sensitivity matrix of the observation model, Q i is the process noise covariance, z i is the observed value, x iis the state estimation, i represents the i-th sensor, h i (x i ) refers to the nonlinear observation model of the i-th sensor;

[0025] S432: Each sensor exchanges local information matrix and information vector through the network and performs fusion:

[0026]

[0027]

[0028]

[0029] In the formula, I fused represents the fused local information matrix, y fused represents the fused information vector, ω i represents the information weight of the i-th sensor, and SNR represents the signal-to-noise ratio;

[0030] S433: Solve the equation according to the fused local information matrix and the fused information vector to obtain the global state estimation:

[0031]

[0032] S434: Perform consistency check on the fused data. If any inconsistency is found, adjust or re-fuse the data to obtain fused data for each target.

[0033] Based on the above technical solution, preferably, step S5 includes:

[0034] S51: Build a combat knowledge base, which includes historical combat data, historical tactical models, historical tactical rules and historical action guidelines;

[0035] S52: Analyze the target situation map in real time to extract information about both sides of the enemy and our own side and battlefield environment and status changes. The information about both sides of the enemy and our own side includes positions, movement information, and resource distribution information;

[0036] S53: Integrate and configure the decision network, including the rule-based network and the neural network model, wherein the neural network model is pre-trained using the operational knowledge base;

[0037] S54: The information of the enemy and our side in the target situation map and the battlefield environment and status changes are integrated with the combat knowledge base to obtain real-time combat data;

[0038] S55: Input real-time combat data into the decision network, use the rule-based network for rapid analysis, match tactical rules and action guidelines, and use the neural network model to predict battlefield actions;

[0039] S56: Generate combat plans based on the output of the decision network, including tactical actions, force deployment, and resource allocation.

[0040] Based on the above technical solution, preferably, the neural network model includes an input module, a feature attention encoding module, a reinforcement learning module and an output module, and the process of the neural network model processing real-time combat data is:

[0041] The input module performs standardization processing on the real-time combat data to obtain standardized data;

[0042] The feature attention encoding module includes a feature extraction layer and an attention layer. The feature extraction layer includes three hidden layers. The first hidden layer is a convolutional layer, the second hidden layer is a recurrent layer, and the third hidden layer is a fully connected layer. The convolutional layer is used to extract the spatial features of the standardized data. The recurrent layer is used to learn the time series of the standardized data. The fully connected layer is used to integrate the spatial features and the time series to form global features. The global features are input into the attention layer. The self-attention mechanism is used to process and encode the global features to obtain the encoded features.

[0043] The reinforcement learning module includes the reinforcement learning algorithm and the uncertainty evaluation layer. It parses the coded features and battlefield environment into the state representation of the reinforcement learning algorithm. According to the current state, the policy network is used to output the action strategy, including all possible actions and their probabilities. The expected return of each action is evaluated according to the value network, and the action with the highest expected return is selected. The action is executed, and the uncertainty evaluation layer is used to quantify the uncertainty of the action to form an uncertainty evaluation result. The reward value is calculated based on the uncertainty evaluation result. The strategy is iteratively updated based on the reward value until the strategy converges and the optimized action strategy is output.

[0044] The optimized action strategy is input into the output module, and the battlefield actions are predicted based on the probability of each action.

[0045] Based on the above technical solution, preferably, the reinforcement learning module adopts the Q learning update formula to update the strategy, and the Q learning update formula is:

[0046]

[0047] Where Q1 and Q2 are independent estimates of the action value function, represented by two different Qs, respectively. t is the state at time t, a t is in state s t The action taken is, α is the learning rate, r t+1 is the reward received at time t+1, γ is the discount factor, is the next state s t+1 Next, use Q1 to select the best action, represents the value of using Q2 to evaluate the best action chosen in Q1.

[0048] Based on the above technical solution, preferably, the uncertainty assessment layer uses an uncertainty estimation formula to perform uncertainty analysis, and the uncertainty formula is:

[0049]

[0050] Where U(a,s) is the uncertainty of action a in state s, P(d|a,s) is the probability distribution of various results d caused by action a in state s, E(s) represents the uncertainty of state s, ξ is an adjustment parameter to balance the uncertainty of the action result and the state, CIWidth(d) is the confidence interval width of the result d, and μ is a weight parameter to adjust the influence of the confidence interval width.

[0051] Based on the above technical solution, preferably, the calculation formula of the reward value is:

[0052]

[0053] Where r(s,a,s') is the reward value, r'(s,a,s') refers to the basic reward of action a from the current state s to the result state s' without considering uncertainty and transfer cost, U(a,s) is the uncertainty of action a in state s, C(s,a,s') is the transfer cost of action a from the current state s to the result state s', and φ=f2 are weight parameters, where f1 and f2 are functions based on time t, learning progress and environmental feedback.

[0054] Based on the above technical solution, preferably, step S6 includes:

[0055] S61: Set evaluation indicators, including combat effectiveness indicators, resource efficiency indicators, and operational risk indicators;

[0056] S62: Execute operational plans in a virtual environment and collect execution data on tactical actions, force deployment, and resource allocation;

[0057] S63: Statistically analyze the execution data, extract the enemy's unit time loss, maximum expected loss, total resource consumption, enemy loss value, own total force and own loss according to the evaluation indicators, calculate the combat effect score according to the enemy's unit time loss and maximum expected loss, calculate the resource efficiency score according to the total resource consumption and enemy loss value, and calculate the operation risk score according to the own total force and own loss;

[0058] S64: Quantitative score is calculated based on the combat effectiveness score, resource efficiency score and operational risk score:

[0059]

[0060] In the formula, S is the quantitative score of the combat plan, D is the combat effect score, Z is the resource efficiency score, W is the operational risk score, and λ D , Z , W is the dynamic factor;

[0061] S65: Based on the expert's review of the operational plan, a qualitative score is formed;

[0062] S66: Combine the quantitative and qualitative scores to form an evaluation report of the operation plan, which includes the advantages and disadvantages of the operation plan;

[0063] S67: Set the defects of the combat plan as the optimization target, and use a multi-objective optimization algorithm to optimize the optimization target. After obtaining the combat plan after each iteration, return to step S62 until the evaluation report meets the requirements and the optimized combat plan is obtained.

[0064] Based on the above technical solution, preferably, step S7 includes:

[0065] S71: According to the optimized combat plan, define the operation parameters, convert the operation parameters into command symbols, and organize all the command symbols into an instruction set;

[0066] S72: Encrypt the instruction set through the military communication network and transmit it to the corresponding combat platform;

[0067] S73: The combat platform generates distributed control logic according to the received instructions, coordinates and controls the actions of each combat platform, and enables each combat platform to be dispatched according to the optimized combat plan.

[0068] The method of the present invention has the following beneficial effects compared with the prior art:

[0069] (1) The present invention combines real-world combat elements with simulation technology in a virtual environment. The combination of the real and the virtual not only improves the accuracy and efficiency of the simulation, but also demonstrates its unique value in constructing the kill chain in the simulation. In the simulation, a battlefield environment closer to actual combat can be simulated through the combination of the real and the virtual, providing real and effective data support for the construction of the kill chain. At the same time, simulation in a virtual environment allows various kill chain construction schemes to be tested and optimized without investing actual resources, greatly reducing costs and risks;

[0070] (2) The present invention maps the real physical model of the combat platform into the digital virtual system, and uses the real equipment as the benchmark to formulate the timing of platform reconnaissance, missile flight and other actions, integrating multiple combat platform resources, aiming to achieve a comprehensive kill chain construction for air, land and submarine. The implementation of this method covers multiple key links such as reconnaissance, judgment, decision-making and attack, forming a complete and efficient combat process;

[0071] (3) The present invention effectively integrates data from multiple sensors through a data fusion algorithm, enhancing the accuracy and reliability of target detection and recognition. During fusion, the data of each sensor updates the local information matrix and information vector according to the nonlinear observation model, and then exchanges these local information through the network and performs weighted fusion, and finally solves the global state estimation. This process not only improves the consistency of the data and the overall signal-to-noise ratio, but also ensures the high accuracy and operability of the target data;

[0072] (4) The present invention improves the intelligence and accuracy of combat plans by constructing and utilizing decision networks. It also provides pre-training for neural network models by utilizing combat knowledge bases, including rich historical combat data and tactical models, thereby ensuring the effectiveness and adaptability of the models in real battlefield environments. By analyzing the target situation map in real time, key enemy and friendly information and battlefield environment changes can be extracted. These data are then deeply processed by neural networks, in which the feature attention encoding module enhances the data representation capability by integrating spatial and temporal features;

[0073] (5) The application of the reinforcement learning module of the present invention enables the system to effectively learn and optimize strategies in a complex battlefield environment, and to adapt to dynamically changing battlefield conditions by continuously iterating and updating strategies. This decision-making framework based on deep learning and reinforcement learning not only improves the speed of decision-making, but also improves the accuracy of action strategies and the innovation of tactics, making combat plans more scientific and systematic, and greatly improving the efficiency of tactical actions and the rational allocation of manpower resources;

[0074] (6) The present invention improves the execution efficiency and effectiveness of combat plans through a systematic evaluation, optimization and scheduling process. By setting detailed evaluation indicators (combat effectiveness, resource efficiency, operational risk), the combat plans are tested and analyzed in a virtual environment to ensure that each plan undergoes rigorous evaluation and necessary adjustments before implementation. The multi-objective optimization algorithm adopted enables the combat plan to be dynamically adjusted according to actual performance and expert review, thereby achieving continuous improvement of the plan. The combat platform generates distributed control logic based on the received optimized instruction set, which not only improves the operational efficiency of the combat platform, but also enhances the synergy between combat units, making the overall combat more accurate and efficient. BRIEF DESCRIPTION OF THE DRAWINGS

[0075] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0076] Figure 1 The figure is a flow chart of a method according to an embodiment of the present invention. DETAILED DESCRIPTION

[0077] The following will be combined with the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. 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 creative work are within the scope of protection of the present invention.

[0078] like Figure 1 As shown, the present invention provides a method for constructing a kill chain based on a simulation deduction combining virtuality and reality, comprising the following steps:

[0079] S1: Obtain combat missions, identify and define targets according to combat mission requirements, and determine the type of kill chain based on combat missions;

[0080] S2: Collect corresponding scenario data for different combat domains and build a three-dimensional digital virtual environment based on the scenario data;

[0081] S3: For each type of combat platform, a multi-dimensional physical model is established, and the multi-dimensional physical model is rendered in real time into a virtual environment to form a virtual combat unit, where the combat platform includes multiple sensors;

[0082] S4: Collect sensor data from each combat platform, fuse the sensor data, and generate a unified target situation map;

[0083] S5: Based on the target situation map and combat knowledge base, use the decision-making algorithm to make a plan decision and generate a combat plan;

[0084] S6: Perform simulation and verification of the combat plan in a virtual environment, evaluate and optimize the combat plan using evaluation methods, and obtain an optimized combat plan;

[0085] S7: Convert the optimized combat plan into an executable instruction parameter set, and transmit the instruction parameter set to each combat platform through a heterogeneous network for coordinated dispatch of operations;

[0086] S8: According to the process of steps S1-S7, a kill chain is formed.

[0087] Kill Chain is a military concept used to describe a series of steps or stages from detecting enemy targets to successfully destroying or neutralizing targets. The concept of kill chain is mainly used to improve combat efficiency by systematically analyzing and executing each step to ensure that enemy targets can be effectively struck on the battlefield. The effective execution of the kill chain depends on a high degree of coordination and rapid decision-making.

[0088] In the process of building the kill chain, the target is required to be used as a traction or guide. Therefore, the kill chain can be called a "target kill chain". At this time, the kill chain can be divided into air defense and anti-missile kill chain, surface target kill chain, and land time-sensitive target kill chain. In addition, based on the battlefield space and combat domain of the kill chain distribution, the kill chain can be divided into land kill chain, sea kill chain, air kill chain, and network kill chain. Different types of kill chains emphasize the attributes of joint operations to organize and reflect multi-domain coordinated operations. Based on the types of military services involved in the construction of the kill chain, the kill chain can be divided into the army kill chain, the navy kill chain, the air force kill chain, and the cyber army kill chain. This method is currently used to divide the kill chain. This is because the kill chain has been led by the military services for a long time.

[0089] With the advent of the era of information warfare and joint operations, various military services are intertwined and coupled in different combat domains. For example, the navy can perform combat missions in all combat domains. At this time, the navy's kill chain includes land kill chain, sea kill chain, air kill chain and network kill chain, etc.; similarly, military services need to be coordinated during maritime operations. The maritime kill chain includes the army kill chain, navy kill chain, air force kill chain and cyber army kill chain, etc. The present invention comprehensively dispatches each combat platform in the simulation and deduction system to construct a complete kill chain in different combat domains.

[0090] Specifically, in one embodiment of the present invention, step S1 includes:

[0091] Obtain combat missions, which usually include a description of the battlefield environment, the enemy and our own situation, expected goals, and tactical objectives.

[0092] According to the combat mission received, key targets are identified, including important enemy infrastructure, military equipment, command centers, communication nodes, etc. The determination of targets is based on detailed intelligence analysis, including the enemy's defense capabilities, the security of the target area, and the possible impact on friendly operations.

[0093] The appropriate kill chain type is determined based on the specific requirements of the combat mission. The type of kill chain depends on many factors, such as the combat environment, the expected tactical effect, the available resources and technology, and the desired speed and secrecy of action.

[0094] Specifically, in one embodiment of the present invention, step S2 includes:

[0095] Detailed scenario data needs to be collected for different combat domains such as land, sea, air and network. These data include terrain, weather, enemy and friendly deployment, communication environment, and other factors that may affect combat. Data sources include satellite images, reconnaissance reports, historical combat data, network monitoring data, etc.

[0096] Based on the collected scene data, a three-dimensional digital virtual environment is constructed using computer simulation technology. This virtual environment should be able to simulate real combat scenarios, including the complexity of terrain, changes in weather conditions and other environmental factors. This virtual environment not only helps to better understand the battlefield environment, but can also be used for subsequent tactical exercises and decision support.

[0097] When building a virtual environment, the needs of multi-domain collaborative operations must also be considered.

[0098] Specifically, in one embodiment of the present invention, step S3 includes:

[0099] The combat platforms of the present invention include manned combat platforms and unmanned combat platforms, and each combat platform requires a detailed physical model, which includes its structure, power system, sensor layout, communication equipment, etc.

[0100] Combat platforms are equipped with a variety of sensors, such as radar, infrared, sonar, cameras, etc. When building a physical model, it is necessary to ensure that the functions, ranges, and limitations of these sensors are also accurately simulated.

[0101] Use rendering algorithms to render physical models into virtual environments in real time to simulate actual combat, including lighting effects, physical collision simulation, and dynamic environment interaction.

[0102] The model of each combat platform represents a combat unit in the virtual environment. These virtual combat units are able to perform specific tasks, such as reconnaissance, strike, or support, and can cooperate with other combat units. In the virtual environment, different combat units need to be able to interact and collaborate according to preset tactics and strategies. For example, a drone may need to link data with an unmanned ship to share reconnaissance information or coordinate strike targets.

[0103] Specifically, in one embodiment of the present invention, step S4 includes:

[0104] S41: Collect data from multiple sensors, and perform spatiotemporal calibration on the data from the multiple sensors.

[0105] First, collect the data of each sensor, use a unified time source, such as GPS clock, to synchronize all sensors; add timestamps to the data collected by each sensor; for data with different sampling frequencies, use linear interpolation for correction:

[0106] If t1 < t < t2, then where x(t) is the corrected data, x(t1) is the data at time t1, and x(t2) is the data at time t2.

[0107] Convert all sensor data to the same coordinate system. If the sensors use different coordinate systems, coordinate transformation is required. For example, the transformation from the geographic coordinate system to the Cartesian coordinate system is:

[0108] x = Rcos(lat)cos(lon), y = Rcos(lat)sin(lon), z = RSin(lat)

[0109] where R is the radius of the earth, and lat and lon are the longitude and latitude respectively.

[0110] S42: Use a data association algorithm to match the data of the same target observed by different sensors to form a target data set.

[0111] First, preprocess the data after spatio-temporal calibration, such as denoising. Extract the features representing the target from the data of each sensor, including position coordinates, speed, size, shape descriptors, spectral features, radar reflection characteristics, etc.

[0112] Adopt a global optimization method to construct a cost matrix for each target. Each element in the cost matrix represents the cost of associating the observed data of one sensor with the observed data of another sensor, and the calculation method is the Euclidean distance.

[0113] Adopt the Hungarian algorithm to find the pairing method with the minimum total cost on the cost matrix, as follows:

[0114] Subtract its minimum value from each row, and then subtract the minimum value in the new matrix from each column;

[0115] In the adjusted matrix, cover all zero elements with as few lines as possible;

[0116] If an independent association cannot be found for each sensor, adjust the matrix to expose more zeros and repeat the marking process;

[0117] Determine the lowest-cost pairing scheme through the marked zeros in the matrix.

[0118] S43: Adopt a fusion algorithm to perform intra-set fusion on each target data set to obtain the fusion data of each target.

[0119] Specifically, step S43 includes:

[0120] S431: Each sensor updates its local information matrix and information vector according to the observed data and the nonlinear observation model hi(x):

[0121]

[0122]

[0123] In the formula, I i is the local information matrix, y i is the information vector, R i is the observation noise covariance, H i is the sensitivity matrix of the observation model, which describes the sensitivity of the state vector to the observation value, Q i is the process noise covariance, which represents the internal uncertainty in the state estimation process, z i is the observed value, x i is the state estimation, i represents the i-th sensor, h i (x i ) refers to the nonlinear observation model of the i-th sensor; wherein the calculation formula of the local information matrix enhances the information content of the system by combining the observation noise and the process noise. The calculation formula of the information vector is used to update the information vector according to the observed data, providing a quantification of the difference between the observed data and the model prediction.

[0124] S432: Each sensor exchanges local information matrix and information vector through the network and performs fusion:

[0125]

[0126]

[0127]

[0128] In the formula, I fused represents the fused local information matrix, y fused represents the fused information vector, ω i represents the information weight of the i-th sensor, and SNR represents the signal-to-noise ratio. The weight is assigned based on the signal-to-noise ratio of each sensor, ensuring that sensors with higher signal quality have a greater impact on the fusion result.

[0129] S433: Solve the equation according to the fused local information matrix and the fused information vector to obtain the global state estimation:

[0130]

[0131] S434: Perform consistency check on the fused data. If any inconsistency is found, adjust or re-fuse the data to obtain fused data for each target.

[0132] The information fusion method used in this embodiment can effectively integrate data from multiple sensors, improve the estimation accuracy of target states in complex environments and the reliability of the system. In addition, the algorithm dynamically adjusts the influence of each sensor in the fusion process by considering its signal-to-noise ratio, thereby optimizing the fusion effect and enhancing the resistance to noise and interference.

[0133] S44: Convert the fused data into target information, and generate a target situation map according to the target information.

[0134] Specifically, based on the targets identified during data association, the targets are continuously tracked, the status changes of each target are monitored, and its location and other relevant information are updated.

[0135] According to the current status and position of all identified and tracked targets in an area, a target situation map is generated, as follows:

[0136] Convert target information into graphical representations such as dots, icons, or track lines, each of which represents a specific type of target and shows its location and direction of movement.

[0137] In this embodiment, the target situation map should show the deployment of both the enemy and our side, the movement trajectories of various dynamic targets, and the control status of key geographical locations.

[0138] Specifically, in one embodiment of the present invention, step S5 includes:

[0139] S51: Build a combat knowledge base, which includes historical combat data, historical tactical models, historical tactical rules and historical action guidelines.

[0140] In this embodiment, the combat knowledge base is used to store and manage historical data and knowledge related to combat. The historical combat data includes past combat records, combat results, performance of participating units, etc. The historical tactical model includes a model formed based on historical data and expert knowledge, which describes the tactical actions to be taken in specific situations. The historical tactical rules include guiding principles for tactical execution. The historical action guidelines include operational level guidelines to guide specific action steps, such as retreat, advance, reconnaissance, etc.

[0141] S52: Analyze the target situation map in real time to extract information about both sides of the enemy and our own side, as well as battlefield environment and status changes. The information about both sides of the enemy and our own side includes their positions, movement information, and resource distribution information.

[0142] Specifically, the situation map provides a comprehensive view of the current battlefield, including key information such as the location, movement, and resource distribution of both sides. Analyze the situation map to identify and record the current location, movement trajectory, and possible intentions of both sides. Determine the resource distribution of both sides, such as troops, weapon systems, and logistical support locations. Monitor and analyze changes in the battlefield environment, such as weather conditions, terrain changes, etc., and the potential impact of these changes on combat operations.

[0143] S53: Integrate and configure decision networks, including rule-based networks and neural network models, where the neural network model is pre-trained using a combat knowledge base.

[0144] In this embodiment, an expert system is used to apply the rules and criteria in the combat knowledge base. These rules can be condition-action pairs, for example, if enemy forces approach a key location, deploy specific defensive measures. By continuously analyzing battlefield data and combat results, the rules in the rule-based system are automatically adjusted or updated.

[0145] Integrate and configure decision networks, define and encode combat rules and action guidelines, and set condition-action correspondence. Build a neural network model, which includes an input module, a feature attention encoding module, a reinforcement learning module, and an output module. Use the combat knowledge base to pre-train the neural network model.

[0146] In one embodiment, the neural network model uses forward propagation and back propagation algorithms to iteratively optimize the model weights, and the loss function selects the mean square error loss. A cross-validation method is used to ensure the generalization ability of the model.

[0147] S54: The information of both sides of the target situation map, battlefield environment and status changes are integrated with the combat knowledge base to obtain real-time combat data.

[0148] Specifically, the information fusion process includes:

[0149] Data alignment: Ensure that the data format, time scale, and reference frame in the real-time data and knowledge base are consistent.

[0150] Data integration: Matching and integrating real-time data with corresponding historical data or models. For example, comparing the current movement patterns of enemy forces with data from similar situations in history.

[0151] Through the above fusion process, real-time combat data is generated.

[0152] S55: Input real-time combat data into the decision network, use the rule-based network for rapid analysis, match tactical rules and action guidelines, and use the neural network model to predict battlefield actions.

[0153] First, rule-based networks are used to quickly respond to clear tactical situations, such as automatically initiating defensive measures when enemy forces enter a preset warning area.

[0154] Use neural network models to process real-time combat data and predict battlefield actions.

[0155] The process of the neural network model processing real-time combat data is as follows:

[0156] The input module normalizes the real-time combat data to obtain standardized data; the normalization formula is as follows:

[0157]

[0158] Among them, μ is the mean and σ is the standard deviation.

[0159] The feature attention encoding module includes a feature extraction layer and an attention layer. The feature extraction layer includes three hidden layers. The first hidden layer is a convolutional layer, the second hidden layer is a recurrent layer, and the third hidden layer is a fully connected layer. The convolutional layer is used to extract the spatial features of the standardized data, the recurrent layer is used to learn the time series of the standardized data, and the fully connected layer is used to integrate the spatial features and the time series to form global features. The global features are input into the attention layer, and the self-attention mechanism is used to process and encode the global features to obtain the encoded features.

[0160] The purpose of this module is to extract and encode key features of the data in order to better understand and utilize the spatial and temporal attributes of battlefield data.

[0161] The convolution layer uses multiple convolution kernels to extract the spatial features of the input data:

[0162] f(x) = ReLU(W*x+b)

[0163] In the formula, ReLU is the activation function, W is the network weight, x is the input data, and b is the bias term;

[0164] The recurrent layer is LSTM, which is used to process the time series features of the data:

[0165] h t =LSTM(x t ,h t-1 )

[0166] In the formula, h represents the hidden state;

[0167] The fully connected layer integrates the outputs of the convolutional layer and the recurrent layer to form a global feature: y = W*h+b;

[0168] The attention layer uses the self-attention mechanism to strengthen important features and suppress unimportant information:

[0169]

[0170] Where Q is the query matrix, K is the key matrix, V is the value matrix, and dk represents the data dimension.

[0171] The reinforcement learning module includes a reinforcement learning algorithm and an uncertainty assessment layer. It parses the encoded features and battlefield environment into a state representation of the reinforcement learning algorithm. According to the current state, the policy network is used to output the action strategy, including all possible actions and their probabilities. The expected return of each action is evaluated according to the value network, and the action with the highest expected return is selected. The action is executed, and the uncertainty of the action is quantified using the uncertainty assessment layer to form an uncertainty assessment result. The reward value is calculated based on the uncertainty assessment result, and the strategy is iteratively updated based on the reward value until the strategy converges and the optimized action strategy is output.

[0172] Specifically, the processing of the reinforcement learning module involves multiple steps from the representation of the environment state to action selection and strategy optimization.

[0173] In the reinforcement learning algorithm, the concepts involved include:

[0174] The state representation describes all relevant information about the current environment. t The state at time t includes information such as the position of the battlefield, the status of the units, the position of the enemy, etc. This information is encoded into a vector or matrix.

[0175] The action space defines all possible actions that the agent can perform. These actions can be simple discrete actions, such as "go forward", "back up", "shoot", or continuous actions, such as changing the degree of speed or direction.

[0176] The policy network uses a simple three-layer network model that accepts the current state as input and outputs the probability of each possible action. The goal of the policy network is to learn which actions are most likely to lead to the highest long-term reward in a given state.

[0177] The value network is used to evaluate the possible future rewards for a given state and action. The value network helps the policy network optimize action selection by providing estimates of future rewards.

[0178] While processing, at each decision point, the agent uses the policy network to select an action. After the agent performs an action, it updates its policy based on the outcome of the action (including the new state and the reward received).

[0179] This embodiment introduces an uncertainty assessment layer in the reinforcement learning module, performs uncertainty assessment on each executed action, and calculates the reward value based on the uncertainty assessment result to help the model better understand and adapt to the complex and uncertain battlefield environment.

[0180] Specifically, the reinforcement learning module uses the Q learning update formula to update the strategy. The Q learning update formula is:

[0181]

[0182] Where Q1 and Q2 are independent estimates of the action value function, represented by two different Qs, respectively. t is the state at time t, a t is in state s t The action taken is, α is the learning rate, r t+1 is the reward received at time t+1, γ is the discount factor, is the next state s t+1 Next, use Q1 to select the best action, represents the value of using Q2 to evaluate the best action chosen in Q1.

[0183] In this embodiment, the overall framework of the reinforcement learning algorithm adopts double Q learning, that is, two independent Q functions are used for estimation to reduce the problem of over-estimation. During learning, two Q functions are randomly used alternately to update the other one. This method improves the learning stability, and the double Q learning can converge faster to a solution closer to the optimal strategy.

[0184] The uncertainty assessment layer uses the uncertainty estimation formula to perform uncertainty analysis. The uncertainty formula is:

[0185]

[0186] Where U(a,s) is the uncertainty of action a in state s, P(d|a,s) is the probability distribution of various results d caused by action a in state s, E(s) represents the uncertainty of state s, ξ is an adjustment parameter to balance the uncertainty of the action result and the state, CIWidth(d) is the confidence interval width of the result d, and μ is a weight parameter to adjust the influence of the confidence interval width.

[0187] In this embodiment, when calculating uncertainty, a comprehensive uncertainty assessment is provided by combining the probability distribution of action results, the inherent uncertainty of the state, and the width of the confidence interval of the result. This helps to consider and balance various risk factors in the decision-making process. By quantifying uncertainty, it can help to choose actions more wisely, especially in high-risk or high-uncertainty environments. This helps to optimize resource allocation and risk management. By dynamically adjusting the ξ and μ parameters, the focus of uncertainty assessment can be adjusted according to different situations and needs, making the decision-making process more adaptable and flexible.

[0188] The reward value is calculated as follows:

[0189]

[0190] Where r(s,a,s') is the reward value, r'(s,a,s') refers to the basic reward of action a from the current state s to the result state s' without considering uncertainty and transfer cost, U(a,s) is the uncertainty of action a in state s, C(s,a,s') is the transfer cost of action a from the current state s to the result state s', and φ=f2 are weight parameters, where f1 and f2 are functions based on time t, learning progress and environmental feedback.

[0191] In this embodiment, an adjusted reward value is calculated by comprehensively considering the basic reward of the action, the uncertainty of the action, and the cost of executing the action. Among them, the penalty of uncertainty reflects the impact of the uncertainty of the action result on the decision. High uncertainty usually means high risk. The dynamics of and φ indicate that the calculation of the reward value can be adjusted according to environmental changes, learning progress or other feedback mechanisms. This makes the model more flexible and adaptable, and can better cope with complex and dynamic decision-making environments.

[0192] The optimized action strategy is input into the output module, and the battlefield actions are predicted based on the probability of each action.

[0193] S56: Generate combat plans based on the output of the decision network, including tactical actions, force deployment, and resource allocation.

[0194] Specifically, battlefield operations are analyzed to identify key variables that affect decision-making, such as enemy strength, location, potential actions, etc.

[0195] Design specific tactical actions based on the enemy situation and your own advantages, including offensive strategies, defensive measures, special missions, retreat and retreat, etc. Plan the specific details of the action, such as the time, place and method of the action.

[0196] Determine the specific configuration and deployment location of each unit based on tactical needs. Allocate necessary resources such as weapons, ammunition, food, medical equipment, etc. according to the combat plan. Ensure adequate logistical support, including transportation, communications, maintenance, etc.

[0197] Generate combat plans based on tactical actions, force deployment, and resource allocation.

[0198] Specifically, in one embodiment of the present invention, step S6 includes:

[0199] S61: Set evaluation indicators, including combat effectiveness indicators, resource efficiency indicators, and operational risk indicators;

[0200] S62: Execute operational plans in a virtual environment and collect execution data on tactical actions, force deployment, and resource allocation;

[0201] S63: Statistically analyze the execution data, extract the enemy's unit time loss, maximum expected loss, total resource consumption, enemy loss value, own total force and own loss according to the evaluation indicators, calculate the combat effect score according to the enemy's unit time loss and maximum expected loss, calculate the resource efficiency score according to the total resource consumption and enemy loss value, and calculate the operation risk score according to the own total force and own loss;

[0202] S64: Quantitative score is calculated based on the combat effectiveness score, resource efficiency score and operational risk score:

[0203]

[0204] In the formula, S is the quantitative score of the combat plan, D is the combat effect score, Z is the resource efficiency score, W is the operational risk score, and λ D , Z , W is a dynamic factor; where D = enemy unit time loss / maximum expected loss*100; Z = (total resource consumption / enemy loss value) -1 *100; W = (1-own losses / own total forces) * 100. D , Z , W It can be adjusted according to the specific needs of the combat environment. This calculation method is not only sensitive to low scores, but also increases the adaptability of the calculation results, and can automatically adjust the evaluation focus according to changes in the actual combat situation.

[0205] S65: Based on the expert's review of the operational plan, a qualitative score is formed;

[0206] S66: Combine the quantitative and qualitative scores to form an evaluation report of the operation plan, which includes the advantages and disadvantages of the operation plan;

[0207] S67: Set the defects of the combat plan as the optimization target, and use a multi-objective optimization algorithm to optimize the optimization target. After obtaining the combat plan after each iteration, return to step S62 until the evaluation report meets the requirements and the optimized combat plan is obtained.

[0208] Specifically, in one embodiment of the present invention, step S7 includes:

[0209] S71: Based on the optimized combat plan, define the operation parameters, convert the operation parameters into command symbols, and organize all the command symbols into an instruction set.

[0210] According to the requirements of the combat plan, the necessary operating parameters are defined, such as movement routes, target identification parameters, weapon use rules, communication protocols, etc. These operating parameters are converted into command symbols that the combat platform can understand and execute. Before actual deployment, these command parameter sets are tested in a simulated environment to ensure that they can be executed correctly in various situations and complete the scheduled tasks without errors.

[0211] S72: The instruction set is encrypted and transmitted to the corresponding combat platform through the military communication network; to prevent the data from being intercepted or tampered with during the transmission process, all transmitted data is encrypted. During the data transmission process, the network status and data integrity are monitored in real time to ensure that all combat platforms can accurately receive the instruction parameter set.

[0212] S73: The combat platform generates distributed control logic based on the received instructions, coordinates and controls the actions of each combat platform, and dispatches each combat platform according to the optimized combat plan. According to the real-time battlefield situation, the action instructions of the combat platform are dynamically adjusted to respond to emergencies and changing battlefield environments.

[0213] In this embodiment, the optimized combat plan is converted into actual combat actions. The combat platform effectively executes tactical tasks through highly coordinated actions, thereby enhancing combat effectiveness and battlefield adaptability.

[0214] According to the above steps performed in the virtual environment, a complete kill chain is formed. Specifically, if it is necessary to build an air kill chain, a land kill chain and a sea kill chain, the corresponding combat missions are obtained and can be obtained by executing steps S1-S7 in the present invention.

[0215] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A kill chain construction method based on virtual-real simulation, characterized in that: The following steps are involved: S1: Obtain combat missions, identify and define targets according to combat mission requirements, and determine the type of kill chain based on combat missions; S2: Collect corresponding scenario data for different combat domains and build a three-dimensional digital virtual environment based on the scenario data; S3: For each type of combat platform, a multi-dimensional physical model is established, and the multi-dimensional physical model is rendered in real time into a virtual environment to form a virtual combat unit, where the combat platform includes multiple sensors; S4: Collect sensor data from each combat platform, fuse the sensor data, and generate a unified target situation map; S5: Based on the target situation map and combat knowledge base, use the decision-making algorithm to make a plan decision and generate a combat plan; S6: Perform simulation and verification of the combat plan in a virtual environment, evaluate and optimize the combat plan using evaluation methods, and obtain an optimized combat plan; S7: Convert the optimized combat plan into an executable instruction parameter set, and transmit the instruction parameter set to each combat platform through a heterogeneous network for coordinated dispatch of operations; S8: According to the process of steps S1-S7, a kill chain is formed; Step S4 includes: S41: Collect data from multiple sensors and perform spatiotemporal calibration on the data from the multiple sensors; S42: using a data association algorithm to match data of the same target observed by different sensors to form a target data set; S43: using a fusion algorithm to perform intra-set fusion on each target data set to obtain fused data of each target; S44: converting the fused data into target information, and generating a target situation map according to the target information; Step S43 includes: S431: Each sensor is based on the observed data and the nonlinear observation model h i (x) Update its local information matrix and information vector: In the formula, I i is the local information matrix, y i is the information vector, R i is the observation noise covariance, H i is the sensitivity matrix of the observation model, Q i is the process noise covariance, z i is the observed value, x i is the state estimation, i represents the i-th sensor, h i (x i ) refers to the nonlinear observation model of the i-th sensor; S432: Each sensor exchanges local information matrix and information vector through the network and performs fusion: In the formula, I fused represents the fused local information matrix, y fused represents the fused information vector, ω i represents the information weight of the i-th sensor, and SNR represents the signal-to-noise ratio; S433: Solve the equation according to the fused local information matrix and the fused information vector to obtain the global state estimation: S434: Check the consistency of the fused data. If any inconsistency is found, adjust or re-fuse the data to obtain fused data for each target. Step S5 includes: S51: Build a combat knowledge base, which includes historical combat data, historical tactical models, historical tactical rules and historical action guidelines; S52: Analyze the target situation map in real time to extract information about both sides of the enemy and our own side and battlefield environment and status changes. The information about both sides of the enemy and our own side includes positions, movement information, and resource distribution information; S53: Integrate and configure the decision network, including the rule-based network and the neural network model, wherein the neural network model is pre-trained using the operational knowledge base; S54: The information of the enemy and our side in the target situation map and the battlefield environment and status changes are integrated with the combat knowledge base to obtain real-time combat data; S55: Input real-time combat data into the decision network, use the rule-based network for rapid analysis, match tactical rules and action guidelines, and use the neural network model to predict battlefield actions; S56: Generate combat plans based on the output of the decision network, including tactical actions, force deployment, and resource allocation; The neural network model includes an input module, a feature attention encoding module, a reinforcement learning module and an output module. The process of the neural network model processing real-time combat data is as follows: The input module performs standardization processing on the real-time combat data to obtain standardized data; The feature attention encoding module includes a feature extraction layer and an attention layer. The feature extraction layer includes three hidden layers. The first hidden layer is a convolutional layer, the second hidden layer is a recurrent layer, and the third hidden layer is a fully connected layer. The convolutional layer is used to extract the spatial features of the standardized data. The recurrent layer is used to learn the time series of the standardized data. The fully connected layer is used to integrate the spatial features and the time series to form global features. The global features are input into the attention layer. The self-attention mechanism is used to process and encode the global features to obtain the encoded features. The reinforcement learning module includes the reinforcement learning algorithm and the uncertainty evaluation layer. It parses the coded features and battlefield environment into the state representation of the reinforcement learning algorithm. According to the current state, the policy network is used to output the action strategy, including all possible actions and their probabilities. The expected return of each action is evaluated according to the value network, and the action with the highest expected return is selected. The action is executed, and the uncertainty evaluation layer is used to quantify the uncertainty of the action to form an uncertainty evaluation result. The reward value is calculated based on the uncertainty evaluation result. The strategy is iteratively updated based on the reward value until the strategy converges and the optimized action strategy is output. The optimized action strategy is input into the output module, and battlefield actions are predicted based on the probability of each action; The reinforcement learning module uses the Q learning update formula to update the strategy. The Q learning update formula is: Where Q1 and Q2 are independent estimates of the action value function, represented by two different Qs, respectively. t is the state at time t, a t is in state s t The action taken is, α is the learning rate, r t+1 is the reward received at time t+1, γ is the discount factor, argmax a Q1(s t+1 , a) is the next state s t+1 Next, use Q1 to select the best action, Q2(s t+1 , argmax a Q1(s t+1 , a)) means using Q2 to evaluate the value of the best action selected in Q1.

2. A method for constructing a kill chain based on virtual-real simulation as claimed in claim 1, characterized in that: The uncertainty assessment layer uses the uncertainty estimation formula to perform uncertainty analysis. The uncertainty formula is: Where U(a, s) is the uncertainty of action a in state s, P(d|a, s) is the probability distribution of various results d caused by action a in state s, E(s) represents the uncertainty of state s, ξ is an adjustment parameter to balance the uncertainty of the action result and the state, CIWidth(d) is the confidence interval width of the result d, and μ is a weight parameter to adjust the influence of the confidence interval width.

3. A method for constructing a kill chain based on virtual-real simulation as claimed in claim 2, characterized in that: The reward value is calculated as follows: In the formula, r(s, a, s′) is the reward value, r′(s, a, s′) refers to the basic reward of action a from the current state s to the result state s′ without considering uncertainty and transfer cost, U(a, s) is the uncertainty of action a in state s, C(s, a, s′) is the transfer cost of action a from the current state s to the result state s′, and φ=f2 are weight parameters, where f1 and f2 are functions based on time t, learning progress and environmental feedback.

4. The method for constructing a kill chain based on virtual-real simulation as claimed in claim 3, characterized in that: Step S6 includes: S61: Set evaluation indicators, including combat effectiveness indicators, resource efficiency indicators, and operational risk indicators; S62: Execute operational plans in a virtual environment and collect execution data on tactical actions, force deployment, and resource allocation; S63: Statistically analyze the execution data, extract the enemy's unit time loss, maximum expected loss, total resource consumption, enemy loss value, own total force and own loss according to the evaluation indicators, calculate the combat effect score according to the enemy's unit time loss and maximum expected loss, calculate the resource efficiency score according to the total resource consumption and enemy loss value, and calculate the operation risk score according to the own total force and own loss; S64: Quantitative score is calculated based on the combat effectiveness score, resource efficiency score and operational risk score: In the formula, S is the quantitative score of the combat plan, D is the combat effect score, Z is the resource efficiency score, W is the operational risk score, and λ D , Z , W is the dynamic factor; S65: Based on the expert's review of the operational plan, a qualitative score is formed; S66: Combine the quantitative and qualitative scores to form an evaluation report of the operation plan, which includes the advantages and disadvantages of the operation plan; S67: Set the defects of the combat plan as the optimization target, and use a multi-objective optimization algorithm to optimize the optimization target. After obtaining the combat plan after each iteration, return to step S62 until the evaluation report meets the requirements and the optimized combat plan is obtained.

5. The method for constructing a kill chain based on virtual-real simulation as claimed in claim 1, characterized in that: Step S7 includes: S71: According to the optimized combat plan, define the operation parameters, convert the operation parameters into command symbols, and organize all the command symbols into an instruction set; S72: Encrypt the instruction set through the military communication network and transmit it to the corresponding combat platform; S73: The combat platform generates distributed control logic based on the received instructions, coordinates and controls the actions of each combat platform, and enables each combat platform to be dispatched according to the optimized combat plan.

Images