Training method of multi-step credibility evaluation model of simulation system
By building a hypernetwork for the simulation system and training a multi-step timing prediction neural network model, the problem of low reliability calculation efficiency in the simulation system is solved, fast and accurate confidence prediction is achieved, and computing efficiency is improved.
Patent Information
- Application Number
- CN202411093008.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-09
- Publication Date
- 2024-11-08
AI Technical Summary
The reliability calculation efficiency in existing simulation systems is low, especially in complex systems, which leads to inefficient efficiency.
By building a hypernetwork of the simulation system, extracting the historical simulation timing data and using a sliding window to build a training sample set, training a multi-step timing prediction neural network model, calculating the credibility of each historical simulation moment, and using the formula Loss=γ1Lmse+γ2Ld +γ3Lt performs loss calculations to improve the reliability prediction efficiency.
It realizes the confidence of the system in a fast and accurate manner, avoids complex computing processes, and improves the reliability computing efficiency.
Smart Images

Figure CN118917206A_ABST
Abstract
Description
Technical Field
[0001] The invention relates to the field of simulation technology, and in particular to a training method for a multi-step credibility evaluation model of a simulation system. Background Art
[0002] During the simulation process, the simulation system needs to constantly observe and collect data, respond to environmental changes, and add or delete nodes. Therefore, the system network is constantly changing, and the credibility of the system at each moment will also change. The training method of the multi-step credibility assessment model of the full-process network simulation system needs to calculate the credibility of the system at each simulation moment. When the simulation system is complex, the amount of calculation is large and the efficiency is low. Summary of the invention
[0003] In view of the above analysis, an embodiment of the present invention aims to provide a training method for a multi-step credibility evaluation model of a simulation system, so as to solve the problem of low efficiency of credibility calculation of the simulation system in the existing method.
[0004] On the one hand, an embodiment of the present invention provides a method for training a multi-step credibility assessment model of a simulation system, comprising the following steps:
[0005] Constructing a hypernetwork corresponding to each historical simulation moment of the simulation system; calculating the credibility of the simulation system at each historical simulation moment based on the hypernetwork;
[0006] Extracting historical simulation time series data of the simulation system, sliding a sliding window with a first length on the historical simulation time series data, obtaining a training sample in each sliding step to construct a training sample set, wherein the input data of the sample is the simulation time series data and the corresponding credibility time series data in the sliding window, and the label of the sample is the credibility of multiple moments after the sliding window;
[0007] A multi-step time series prediction neural network model is constructed, and the time series prediction neural network model is trained based on the training sample set to obtain a credibility prediction model.
[0008] Based on the further improvement of the above method, the following formula is used to calculate the training loss of the multi-step time series prediction neural network model:
[0009] Loss = γ1L mse +γ2L d +γ3L t
[0010] Among them, L mse represents the mean square error loss, L d represents the distance loss between the prediction credibility sequence and the label credibility sequence, L trepresents the time difference loss between the prediction credibility sequence and the label credibility sequence, and γ1, γ2 and γ3 represent weights.
[0011] Based on the further improvement of the above method, the following formula is used to calculate the distance loss between the prediction credibility sequence and the label credibility sequence:
[0012]
[0013] Among them, Y i represents the label credibility sequence of the i-th sample, represents the prediction credibility sequence of the i-th sample, represents the distance matrix between the label credibility sequence and the prediction credibility sequence, <·,·> represents the DTW distance, γ represents the parameter, represents a set of k×k matrices, where k is the length of the label credibility sequence.
[0014] Based on the further improvement of the above method, the following formula is used to calculate the time difference loss:
[0015]
[0016]
[0017] Among them, Y i represents the label credibility sequence of the i-th sample, represents the prediction credibility sequence of the i-th sample, represents the distance matrix between the label credibility sequence and the prediction credibility sequence, <·,·> represents the DTW distance, γ represents the parameter, represents a set of k×k matrices, k is the length of the label credibility sequence, and Ω is a k×k matrix.
[0018] Based on the further improvement of the above method, the multi-step time series prediction neural network model includes:
[0019] A first encoding module, used for extracting features from input simulation time series data to obtain a first feature;
[0020] A second encoding module is used to extract features from the input credibility time series data to obtain a second feature;
[0021] A feature fusion module, used for fusing the first feature and the second feature to obtain a fusion feature;
[0022] The prediction unit is used to predict the multi-step credibility values of the complex simulation system based on the fusion features.
[0023] Based on the further improvement of the above method, the first encoding module includes a multi-layer gated recurrent unit and a weighting module;
[0024] Each layer of gated recurrent units is used to extract hidden features based on input data; the input of the first layer of gated recurrent units is the simulation time series data; the hidden features extracted by each layer of gated recurrent units except the last layer are the input of the next layer of gated recurrent units; the hidden features output by the last layer of gated recurrent units are transmitted to the weighting module;
[0025] The weighted module performs weighted fusion on the hidden features output by the last layer of gated recurrent units at each time step to obtain the first feature.
[0026] Based on the further improvement of the above method, the credibility of the simulation system at each historical simulation moment is calculated based on the hypernetwork, including:
[0027] Calculate the importance of each node in the simulation system at each historical simulation moment;
[0028] The credibility of the simulation system at each historical simulation moment is calculated based on the importance of each node.
[0029] Based on the further improvement of the above method, the importance of each node in the simulation system at each historical simulation moment is calculated in the following way:
[0030] Based on the system's hyper-network search for OODA loops at each historical simulation moment, the initial comprehensive confrontation capability of each OODA loop is calculated;
[0031] For each node, treat it as a failed node and update the super network based on the network cascade failure analysis; calculate the updated comprehensive confrontation capability of each OODA loop based on the updated super network;
[0032] Based on the initial comprehensive confrontation capability and the updated comprehensive confrontation capability of each OODA loop, the importance of the failed node is obtained.
[0033] Based on the further improvement of the above method, the importance of the failed node is calculated using the following formula:
[0034]
[0035]
[0036] in, represents the initial comprehensive confrontation capability of the jth OODA loop, The updated comprehensive confrontation capability of the jth OODA loop, represents the weight of the jth OODA loop, Value(v) represents the importance evaluation result of node v, and N T represents the number of target nodes in the simulation system, Indicates the number of OODA loops where the i-th target node is located.
[0037] Based on the further improvement of the above method, the initial comprehensive confrontation capability of each OODA loop is calculated in the following way:
[0038]
[0039] Among them, P op represents the target damage probability of the OODA loop, Q op represents the stability of the topological structure of the OODA loop, w1 and w2 represent the weights of the target damage probability and topological structure stability of the OODA loop respectively.
[0040] Compared with the prior art, the present invention constructs a training sample set by calculating the credibility of the system at each historical simulation moment and extracting the corresponding historical simulation time series data, and trains a multi-step time series prediction neural network model based on the constructed sample set, thereby obtaining a credibility prediction model that can predict the credibility of the system at multiple moments in the future. Based on the credibility prediction model, the credibility of the system can be quickly obtained, avoiding complex calculation processes, thereby improving the credibility calculation efficiency.
[0041] In the present invention, the above-mentioned technical solutions can also be combined with each other to achieve more preferred combination solutions. Other features and advantages of the present invention will be described in the subsequent description, and some advantages can become obvious from the description, or can be understood by practicing the present invention. The purpose and other advantages of the present invention can be realized and obtained through the contents particularly pointed out in the description and the drawings. BRIEF DESCRIPTION OF THE DRAWINGS
[0042] The accompanying drawings are only used for the purpose of illustrating specific embodiments and are not to be considered as limiting the present invention. In the entire drawings, the same reference symbols represent the same components;
[0043] Figure 1 The present invention is a flowchart of a method for training a multi-step credibility assessment model of a simulation system according to an embodiment of the present invention. DETAILED DESCRIPTION
[0044] The preferred embodiments of the present invention are described in detail below in conjunction with the accompanying drawings, wherein the accompanying drawings constitute a part of this application and are used together with the embodiments of the present invention to illustrate the principles of the present invention, but are not used to limit the scope of the present invention.
[0045] A specific embodiment of the present invention discloses a training method for a multi-step credibility evaluation model of a simulation system, such as Figure 1 As shown, the following steps are included:
[0046] S1, constructing a super network corresponding to each historical simulation moment of the simulation system; calculating the credibility of the simulation system at each historical simulation moment based on the super network;
[0047] S2, extracting the historical simulation time series data of the complex simulation system, sliding a sliding window with a first length on the historical simulation time series data, obtaining a training sample in each sliding step to construct a training sample set, wherein the input data of the sample is the simulation time series data and the corresponding credibility time series data in the sliding window, and the label of the sample is the credibility of the sliding window at multiple moments thereafter;
[0048] S3. Construct a multi-step time series prediction neural network model, and train the time series prediction neural network model based on the training sample set to obtain a credibility prediction model.
[0049] It should be noted that the simulation system may be, for example, a system confrontation simulation system.
[0050] Compared with the prior art, the training method of the multi-step credibility assessment model of the simulation system provided in this embodiment calculates the credibility of the system at each historical simulation moment and extracts the corresponding historical simulation time series data to construct a training sample set, and trains a multi-step time series prediction neural network model based on the constructed sample set, thereby obtaining a credibility prediction model that can predict the credibility of the system at multiple moments in the future. The credibility of the system can be quickly obtained based on the credibility prediction model, avoiding complex calculation processes, thereby improving the credibility calculation efficiency.
[0051] Construct a hypernetwork corresponding to each historical simulation moment of the simulation system. As a complex network organization structure, the hypernetwork can connect and interact multiple subnetworks to form a larger and more complex network system. This form of network organization can provide more powerful, flexible and collaborative functions, and is suitable for application scenarios that require scale expansion and functional integration. Modeling the simulation system based on the hypernetwork theory can overcome the limitations of general complex network models in modeling and characterizing the network structure and functional characteristics of the simulation system. During implementation, the hypernetwork modeling idea is used to construct a hypernetwork of a complex simulation system in accordance with the principle of "units are condensed into points, relationships are abstracted into edges, and systems are transformed into networks."
[0052] During implementation, each entity in the simulation system is a node in the hypernetwork. The relationship between entities is abstracted as edges in the network. The entities in the complex simulation system are of different types, and nodes of the same type are divided into a network, and the networks are coupled to form a hypernetwork.
[0053] Specifically, the nodes of the complex simulation system include four types: sensor nodes S, which are used to collect raw situation information, process situation information to generate intelligence information, such as radars and reconnaissance satellites; command and control nodes D, which make decisions based on intelligence information, specify confrontation plans, and command and control confrontation activities. The command and control center is the command and control node; action nodes, which receive command and control information and intelligence information, execute decisions, and implement confrontation on targets; target nodes, which are the confrontation targets in the simulation system.
[0054] In a simulation system, a simulation entity may have multiple node types. For example, an airborne early warning aircraft can not only conduct reconnaissance and surveillance activities in a confrontation environment, but also serve as a command and control center to command and control related confrontation activities. This type of node is called a composite node. The multiple functions of a node may lead to multiple associations between the same node pair, which is not conducive to a clear description of the specific functions of the node in the confrontation process. In view of this situation, based on the idea of super-network, the composite node is abstracted into multiple network nodes according to its function for description during implementation. For example, the airborne early warning aircraft can be abstracted into a sensor node S and a command and control node D to describe its reconnaissance and surveillance capabilities and command and control capabilities respectively.
[0055] During implementation, the interaction edges in the simulation system include: the reconnaissance and surveillance link T→S for the sensor node to monitor the target node and collect target intelligence, the target intelligence sharing link S→S between the sensor nodes, the intelligence support link S→D for the sensor node to send the target intelligence information to the command node, the intelligence support link S→A for the sensor node to send the target intelligence information to the action node, the collaborative decision-making link D→D for sharing intelligence and decision-making collaboration between the same-layer command nodes, the command and control link D→D for the upper-level command node to send the command information to its lower-level command node, the command and control link D→A for the command and control node to send the command information to the action node, and the damage link A→T for the action node to carry out fire strikes or electronic interference on the target node.
[0056] Nodes in different networks are connected through these links to form a super network. The node set V in the constructed super network is expressed as: V = {S, D, A, T}.
[0057] in, represents the set of sensor nodes in the simulation system, N s Indicates the number of sensor nodes; Represents the collection of accusation nodes, N D Indicates the number of accusation nodes; Represents the set of action nodes, N A Indicates the number of action nodes; Represents the target class node set in the simulation system, N T Indicates the number of target class nodes.
[0058] According to the connection relationship between each node, the edges in the hypernetwork are constructed to obtain the adjacency matrix X of the hypernetwork.
[0059] Calculating the credibility of the simulation system at each historical simulation moment based on the hypernetwork includes:
[0060] S11, calculating the importance of each node in the simulation system at each historical simulation moment;
[0061] S12. The credibility of the simulation system at each historical simulation moment is obtained based on the importance of each node.
[0062] Specifically, in step S12, the credibility of the simulation system is calculated based on the importance of each node in the following manner:
[0063]
[0064] Among them, Y represents the credibility of the system, ω j represents the importance of the jth node in the system, y j represents the credibility of the jth node in the system, N G Indicates the number of nodes in the system.
[0065] During implementation, the credibility of each node can be evaluated based on the simulation data of each node using an existing simulation credibility evaluation method. For example, the credibility of the node can be obtained by performing a comprehensive uncertainty measurement on the node.
[0066] As the simulation system continues to change, the importance of each node will change, and its contribution to the credibility of the entire system will also change. Therefore, it is necessary to calculate the importance of each node in the system at each simulation moment, so as to evaluate the credibility of the system more objectively.
[0067] Specifically, the importance of each node in the system at each historical simulation moment is obtained in the following way:
[0068] S111, searching the OODA loops based on the hyper-network of the system at each historical simulation moment, and calculating the initial comprehensive confrontation capability of each OODA loop;
[0069] S112, for each node, treat it as a failed node, and update the super network based on the network cascade failure analysis; calculate the updated comprehensive confrontation capability of each OODA loop based on the updated super network;
[0070] S113. Based on the initial comprehensive confrontation capability and the updated comprehensive confrontation capability of each OODA loop, the importance of the failed node is obtained.
[0071] At each historical simulation moment, the OODA loop in the corresponding hypernetwork is searched. OODA refers to a cyclic model composed of observation (Obeserve) - judgment (Orient) - decision (Decide) - action (Act), which is used to describe the entire process of the entire system's confrontation action. In the OODA loop, information flow plays a vital role, and information flow is constantly circulating in the OODA loop. Each cycle collects information through observation, forms new actions through positioning and decision-making, and then observes the results again, and then adjusts and updates the collection and analysis of information to form new decisions and actions. The rapidity and flexibility of this cycle make the OODA loop an effective decision-making model for dealing with complex and dynamic environments.
[0072] According to the OODA concept diagram, the process of completing a confrontation mission can be regarded as a cycle, and the process can be briefly described as follows: the reconnaissance and surveillance simulation entity S obtains information about the enemy target, and transmits this information to our command center D through the communication relationship between entities. The command center D processes and analyzes the information obtained, and makes a decision by comprehensively considering the current battlefield situation and the strength of both sides, and issues an action order to the action entity A. After receiving the order, A immediately takes corresponding enemy control measures against the enemy target T. In order to complete a specific confrontation mission, it is necessary to build one or more closed loops formed by the reconnaissance and surveillance entity S, command and control entity D, action entity A, and enemy target entity T in our system, and ensure that these loops can complete their respective confrontation tasks with the highest possible stability.
[0073] The OODA loop in the present invention is an OODA loop in which the starting node and the ending node are the same node and the node types are all target nodes. Taking the typical OODA loop in the confrontation process: T→S→D→A→T as an example, the OODA loop indicates that the red party target node is discovered by the blue party reconnaissance node, and the blue party reconnaissance node transmits effective information to the accusation node through information transmission, and the accusation node processes and analyzes the information and issues an attack command to the action node, and finally the action node attacks the red party target node. Table 1 shows the OODA loop in the present invention.
[0074] Table 1 OODA loop and its meaning
[0075]
[0076] The rapid and effective extraction of OODA loops in the simulation system hypernetwork is the basis for evaluating the importance of nodes based on OODA loops. From the perspective of graph theory, the OODA loop search problem can be described as a subgraph isomorphism problem. In implementation, the VF3 algorithm can be used to search for OODA loops in the hypernetwork.
[0077] On the basis of successful extraction of OODA loop, a reasonable and effective evaluation of its comprehensive confrontation capability is the premise of evaluating the importance of hypernetwork nodes based on OODA loop. The comprehensive confrontation capability of OODA loop is mainly reflected in two aspects: 1) the damage capability of OODA loop to enemy targets; 2) the ability of OODA loop to maintain its confrontation capability, that is, the stability of OODA loop topology.
[0078] Under uncertainty, the damage capability of an OODA loop can be expressed as its probability of damage to the enemy target P op The comprehensive probability of damage to the target node by an action node in the OODA loop depends on four factors: 1) The quality of the target intelligence information received by the action node P S→D ; 2) The accuracy of the superior's accusation information D→A ; 3) Under the condition that the target information is completely determined and the superior accusation information is completely correct, the probability P of the action node hitting the target node h ; 4) Under the condition of hitting the target node, the probability P of the action node destroying the target node d .
[0079] Therefore, for a certain action node in a certain OODA loop, the comprehensive probability of damaging the target node in the loop is:
[0080] P A =P S→D ×P D→A ×P h ×P d (2)
[0081] Where k represents the number of action nodes in the OODA loop, P S→D Indicates the quality of target intelligence information received by the action node, P D→A Indicates the correctness of the accusation information received by the action node, P h represents the probability of the action node hitting the target node in the OODA loop, P d It represents the conditional damage probability of the action node to the target node, that is, the probability that the target node will be damaged when the action node hits the target node in the OODA loop.
[0082] Among them, the hit probability P of the action node to the target node h Including the probability P that the action node hits the target node by attacking h ′ and the probability P of the action node hitting the target node through electronic interference h ″. The conditional damage probability P of the action node to the target node h It consists of two parts: the conditional damage probability P of the action node destroying the target by attacking d′ and the conditional damage probability P of the action node destroying the target by electronic interference d "And there are
[0083] P h =1-(1-P h′ )(1-P h″ ) (3)
[0084] P d =1-(1-P d′ )(1-P d″ ) (4)
[0085] Among them, P h′ represents the probability of the action node hitting the target node by attacking, that is, under the condition that the target information is completely determined and the superior accusation information is completely correct, P h″ The probability of the action node hitting the target node through electronic interference, that is, under the condition that the target information is completely certain and the superior command information is completely correct. d′ P represents the conditional damage probability of the action node to the target node through attack, that is, the probability of the action node damaging the target under the condition that the action node hits the target based on the attack method, d″ The conditional probability of an action node damaging a target node through electronic interference is the probability of the action node damaging the target under the condition that the action node hits the target based on electronic interference.
[0086] The higher the strike accuracy of the action node, the smaller the circular probability deviation, and the smaller the area of the strike concentration area. If the strike area of the target node is larger, the hit probability is greater; the lower the strike accuracy of the action node, the larger the circular probability deviation, and the larger the area of the strike concentration area. If the strike area of the target node is smaller, it is more difficult for the action node to hit the target node, that is, the lower the hit probability. When the area of the strike concentration area is smaller than the target strike area, it indicates that the action node will definitely hit the target, that is, the hit probability is 1; when the area of the strike concentration area is larger than the target strike area, the probability of the action node hitting the target depends on the ratio of the target strike area to the area of the strike concentration area. In summary, the relationship between the hit probability of the action node on the target node, its strike accuracy, and the strike area of the target node can be described as follows in the mathematical form
[0087]
[0088] Among them, τ′ is the state quantity describing whether the target node is out of the attack range of the action node after the early warning maneuver,
[0089] Indicates the strike accuracy of the action node, namely the circular error probability (CEP); Indicates the area of the action node's strike concentration zone. The smaller the area, the higher the strike accuracy. Indicates the hit area of the target node. AT represents the initial distance between the action node A and the target node T, which can be calculated using the traditional distance calculation formula between two points; and respectively Indicates the maneuvering speed and warning time of the target node; Indicates the effective range of the action node.
[0090] Under the condition of hitting the target, the probability of an action node damaging the target is the probability that its opponent falls into the vulnerable area of the target. Therefore, the conditional damage probability of an action node can be expressed as:
[0091]
[0092] in, Indicates the power of the action node against the unit, with a value range of [0,1]; It represents the sequence of vulnerable areas of target nodes when facing attacks of different power countermeasures; Indicates that the target node faces a force of The relative vulnerable area when confronting the part.
[0093] The probability P of an action node hitting the target under electronic jamming h ″ is expressed as
[0094]
[0095] Among them, τ″ is the state quantity that describes whether the target node is out of the electronic interference range of the action node after the early warning maneuver,
[0096] Indicates the interference frequency band range of the action node, Refers to the target's operating frequency band, Indicates the overlap between the target working frequency band and the action node interference frequency band; Indicates the interference radius of the action node.
[0097] Under the condition of hitting the target, since the present invention mainly considers active suppression interference, its essence is the confrontation between the power of the enemy and the friend. Therefore, the conditional damage probability of the action node to the target node is mainly related to the interference power of the action node and the transmission power of the target node. The ratio of the interference power of the action node to the maximum transmission power of the target node is called the suppression coefficient of the action node to the target node, denoted as ρ, then
[0098]
[0099] in, represents the interference power of the action node; Represents the maximum transmission power of the target node. The suppression coefficient determines the effect of suppressive electronic interference to a certain extent. Therefore, for the convenience of research, the conditional damage probability P″ of the action node to the target node under electronic interference can be d Description as follows
[0100]
[0101] Among them, ρ represents the suppression coefficient of the action node on the target node; ρ min Represents the minimum suppression coefficient that can interfere with the target node; ρ max It represents the minimum suppression coefficient that makes the target node lose its ability to resist 100%. min When ρ>ρ, the action node cannot effectively interfere with the target, and the conditional damage probability is 0; max When ρ , the action node completely suppresses the target node, causing the target node to lose its confrontation ability. At this time, the conditional damage probability is 1. min ≤ρ≤ρ max When the suppression coefficient increases, the action node can produce effective electronic interference to the target node, and the probability that the target node loses its confrontation capability increases with the increase of the suppression coefficient.
[0102] Therefore, it is necessary to reasonably evaluate the quality of the target intelligence information received by each action node in the OODA loop. S→D , and the correctness of the accusation information issued by each accusation node P D→A , is the key to describing the destructive capability of the OODA loop.
[0103] The target intelligence information received by the action node is actually the target intelligence information obtained by the sensor node that provides intelligence information support for the action node. The target intelligence information obtained by the sensor node comes from its reconnaissance and monitoring process of the target node on the one hand, and from the intelligence sharing process between sensor nodes on the other hand. Before the sensor nodes share the target intelligence information, the probability of each sensor node detecting and identifying the target node can be obtained. The calculation of the sensor node identification probability can be obtained by referring to the relevant industry standards of the corresponding equipment of each sensor node. This represents the quality of the target intelligence information obtained by each sensor node.
[0104] Through intelligence sharing, each sensor node can obtain the target intelligence information of other sensor nodes. By fusing the target intelligence information from multiple sensor nodes, more accurate, reliable and complete estimation and judgment of target attributes such as target location, state estimation and identity description can be obtained. However, on the other hand, as the number of intelligence sharing increases, the intelligence sharing time continues to grow, and the timeliness of information will gradually decrease, resulting in a decrease in the quality of the final target intelligence information. It can be seen that in the process of intelligence sharing, the information fusion capability of the sensor node and the total duration of the intelligence sharing process jointly determine the quality of the target intelligence information finally obtained by each sensor node. Among them, the information fusion capability mainly represents the ability of the sensor node to mine useful information from the target intelligence information of other nodes. The information fusion capability is only represented by the utilization rate of the sensor node to the intelligence information of other sensor nodes. The information utilization rate is usually obtained by the ratio of the utilized information to the total amount of information obtained, and the value range is 0-1; on the other hand, since the intelligence sharing link represents the intelligence sharing process between two sensor nodes, the number of intelligence sharing links in the OODA loop can represent the total duration of the intelligence sharing process. Based on the above analysis, the intelligence sharing process between two sensor nodes can be described as follows.
[0105] Assume that the sensor node S i and S j There is a monitoring relationship between the two sensor nodes and the target node T, and the two sensor nodes share the target intelligence information. Assume that before sharing, the probability of each sensor node detecting and identifying the target node is and After sharing the target intelligence information, the probability of each sensor node detecting and identifying the target node is as follows:
[0106]
[0107]
[0108] in, and After sharing intelligence, the sensor nodes S i and S j The probability of detecting and identifying the target node T; and They represent the sensor nodes S i and S j information fusion capability, that is, information utilization rate.
[0109] On this basis, the impact of intelligence sharing duration on the quality of intelligence information must also be considered.
[0110] It should be noted that the quality of the target intelligence information received by the action node is P S→DIt is the quality of the target intelligence information obtained by the last sensor node in the current OODA loop.
[0111] Based on this, the quality P of the target intelligence information obtained by the sensor nodes in the OODA loop after intelligence sharing can be calculated. S→D :
[0112]
[0113]
[0114] Among them, the i-th sensor node is the last sensor node in the OODA loop that sends target intelligence information to the accusation node. represents the information utilization rate of the i-th sensor node, R(S i ) represents the serial number set of sensor nodes that share intelligence with the i-th sensor node, represents the probability of the i-th sensor node detecting and identifying the target node, It represents the probability that the i-th sensor node detects and identifies the target node after the intelligence sharing process; λ1 represents the attenuation coefficient of the timeliness of intelligence information. The smaller λ1 is, the faster the attenuation speed of the timeliness of information is; It represents the number of intelligence sharing links in the OODA loop, which to a certain extent represents the duration of the intelligence sharing process.
[0115] For example, if the OODA loop is T→S→S→D→S→D→A→T, then the third sensor node is the last sensor node in the OODA loop to send target intelligence information to the command node.
[0116] In the individual decision-making process, the correctness of the accusation information output by the accusation node is P D It can be expressed as:
[0117]
[0118] in, Indicates the superior accusation node D of the control node D up The correctness of the output accusation information is that when the accusation node is the highest accusation node in its OODA loop, That is, when l D =1, P D =P S→D ×η D ;P S→D Indicates the quality of target intelligence information received by the action node; η DIt represents the information utilization capability of the command node D in the current state. The value range is 0-1. The information utilization capability can be approximately calculated according to the task difficulty coefficient. The collaborative decision-making process between the command nodes at the same level is similar to the intelligence sharing process between the sensor nodes. The only difference is that what is shared between the command nodes is the decision opinions of each command node on the same confrontation task, which is essentially knowledge sharing. Therefore, the collaborative decision-making process between the command nodes at the same level can be described by analogy with the modeling idea of the intelligence sharing process. Similar to the intelligence sharing process between sensor nodes, the information utilization capability in the collaborative decision-making process is an important factor in determining the correctness of the command information output by each command node. In addition, since the confrontation situation changes rapidly, if the decision-making time is too long, the decision made at the previous moment may not be applicable to the confrontation situation at the current moment, thus missing the opportunity. Therefore, the impact of the duration of the collaborative decision-making process on the correctness of the command information must be considered. Since the collaborative decision-making link represents the collaborative decision-making relationship between the command nodes, the number of collaborative decision-making links between the command nodes at the same command level can be used to represent the duration of the collaborative decision-making process between the command nodes at this level.
[0119] Assume that the two command nodes D1 and D2 in the OODA loop can make collaborative decisions, and the information utilization capabilities of the two command nodes in the current state are and After individual decision-making, the correctness of the results output by each command node is and Then after collaborative decision-making, the correctness of the accusation information output by each accusation node is:
[0120]
[0121]
[0122] in, and They respectively represent the correctness of the accusation information output by the accusation nodes D1 and D2 after collaborative decision-making.
[0123] On this basis, considering the impact of the duration of the collaborative decision-making process on the correctness of the accusation information, the final correctness of the accusation information output by each accusation node can be obtained.
[0124] It should be noted that the correctness of the accusation information received by the action node is the correctness of the accusation information output by the last accusation node to send the accusation information, which is specifically expressed as follows:
[0125]
[0126]
[0127] in, Indicates the correctness of the initial output charge information of the i-th charge node, Indicates the charge node D i (D i ∈D op ) After the collaborative decision-making process, the final correctness of the output accusation information, represents the correctness of the initial output accusation information of the jth accusation node, λ2 represents the attenuation coefficient of the timeliness of the accusation information, the smaller λ2 is, the faster the timeliness of the information decays; represents the number of collaborative decision-making links in the corresponding command level in the OODA loop, represents the information utilization capability of the ith accusation node, R(D i ) represents the serial number set of the accusation nodes that make collaborative decisions with the ith accusation node. i It is the last charging node to send charging information to the action node.
[0128] For example, if the OODA loop is T→S→S→D→S→D→A→T, then the second sensor node is the last sensor node in the loop to send target intelligence information to the command node.
[0129] The target damage probability P of any OODA loop op It can be calculated using the following formula:
[0130]
[0131] Among them, P Ai represents the comprehensive damage probability of the i-th action node in the OODA loop to the target node in the loop, and k is the number of action nodes in the OODA loop.
[0132] The destructive capability of the OODA loop is closely related to the properties of all nodes and edges in the loop. Therefore, maintaining the stability of the OODA loop topology is the basic condition for maintaining its confrontation capability. The OODA loop consists of four types of nodes, namely target nodes, sensor nodes, command nodes and action nodes, as well as many edges between nodes. The stability of its topology is inseparable from the stability of every node and every edge in the loop.
[0133] Since the target node is the object of the OODA loop, its stability is regarded as 1. Therefore, the stability of the sensor node, the command node and the action node in the OODA loop is mainly analyzed. There are two main factors affecting node stability: internal factors and external factors. Among them, internal factors mainly refer to the impact of internal fault factors of the node on the stability of the node; while external factors refer to the impact of the correlation between nodes on the stability of the node, and the latter is mainly considered here. When the node remains stable internally, whether the material resource demand of the node is adequately guaranteed is the main factor affecting its stability. Therefore, the present invention measures the stability of the node by the material guarantee strength of the simulation system for each node. According to the material guarantee needs of the node and the guarantee capabilities of the material guarantee department, the stability of each node can be described as follows:
[0134]
[0135] Among them, Q i represents the stability of the i-th node in the OODA loop, Indicates the strength of the jth material of the i-th node in the OODA loop. In a special ideal situation, the material guarantee of the i-th node is sufficient, indicating that the i-th node does not need any material resource guarantee, that is, its stability Q i =1.
[0136] The communication link is used to characterize the static connection relationship between nodes. Therefore, when the nodes at both ends are not failed, the stability of the communication link is considered to be 1. The OODA loop mainly involves 7 types of links, namely, reconnaissance and surveillance links, intelligence sharing links, intelligence support links, command and control links, collaborative decision-making links, status reporting links, and damage links. Once the OODA loop is formed, the reconnaissance and surveillance relationship and damage relationship of other nodes in the loop to the target node are established and unchanged. Therefore, the stability of the reconnaissance and surveillance link and the damage link is considered to be 1. In addition, the intelligence sharing link, intelligence support link, command and control link, collaborative decision-making link, and status reporting link all characterize the information interaction relationship between nodes, which is essentially the information transmission process between nodes. Therefore, the stability of the above 5 types of links mainly depends on the information transmission capacity between the nodes at both ends. The number of effective communication paths between nodes is an effective indicator to characterize the stability of information interaction between nodes. Therefore, the stability calculation method of the information interaction link is as follows.
[0137]
[0138] Among them, Q ij represents the stability of the link between the i-th node and the j-th node, α represents the coefficient of the influence of the number of effective communication paths between nodes in the OODA loop on the stability of the link, and CL ijRepresents the number of valid communication paths between the i-th node and the j-th node of the OODA loop.
[0139] For an OODA loop, its topological stability Q op It is expressed as:
[0140]
[0141] Among them, Q i Indicates the stability of the i-th node in the OODA loop, with a value range of [0,1]; N op Indicates the number of nodes in the OODA loop; Q ij represents the link l between the i-th node and the j-th node ij The stability of the value range is [0,1], and when l ij = 0, Q i =1.
[0142] Then, the initial comprehensive confrontation capability of each OODA loop is calculated in the following way:
[0143]
[0144] Among them, P op represents the target damage probability of the OODA loop, Q op represents the stability of the topological structure of the OODA loop, w1 and w2 represent the weights of the target damage probability and topological structure stability of the OODA loop respectively.
[0145] After obtaining the comprehensive importance of the OODA loop, to evaluate the importance of a certain node in it, we can refer to the idea of node deletion and describe it with the change in the OODA comprehensive importance of the entire network before and after deleting the node.
[0146] Considering the characteristics of the simulation system, we cannot only delete the nodes to be evaluated and the edges directly connected to the nodes to be evaluated, which is likely to lead to a low final evaluation result, because the failure of some nodes may cause cascading failures of other nodes and edges. Therefore, in order to achieve a comprehensive and effective evaluation of node value, it is necessary to perform cascading failure analysis on the hypernetwork of the simulation system.
[0147] After deleting the nodes to be evaluated, the relevant failed nodes are analyzed through cascading failure analysis, and then these nodes and directly connected edges are deleted from the super network to update the super network. The updated comprehensive confrontation capability of each OODA loop is calculated based on the updated super network. It should be noted that if an OODA loop has a failed node, the comprehensive confrontation capability of the OODA loop is considered to be 0. Otherwise, its comprehensive confrontation capability is re-evaluated according to the above process to obtain the updated comprehensive confrontation capability.
[0148] According to the comprehensive confrontation capabilities of each OODA loop before and after the node v to be evaluated is deleted, the support degree of the node to be evaluated for each OODA loop can be calculated, which is expressed as follows:
[0149]
[0150] in, Indicates the node v to be evaluated with the i-th target node T i is the support of the jth OODA loop of the target, Indicates that the i-th target node T i The initial comprehensive confrontation capability of the target j-th OODA loop, Indicates that the i-th target node T i The updated comprehensive confrontation capability of the target j-th OODA loop.
[0151] Comprehensively consider the support of the node v to be evaluated for each OODA loop to obtain the importance evaluation result of the node to be evaluated
[0152]
[0153] in, Indicates the node v to be evaluated with the i-th target node T i is the support of the jth OODA loop of the target, Indicates that the i-th target node T i is the weight of the jth OODA loop of the target, Value(v) represents the importance evaluation result of node v, Indicates that the i-th target node T i is the number of the jth OODA loop of the target, N T Indicates the number of target nodes.
[0154] Take the i-th target node T i The weight of the j-th OODA loop for the target is:
[0155]
[0156] in, Represents the target node T i The weight of Indicates that the i-th target node T i The jth OODA loop for the target Initial comprehensive confrontation capability. It can be evaluated by existing methods such as expert scoring.
[0157] During implementation, each node can be deleted in turn, and the importance of the node can be evaluated based on the changes in the comprehensive confrontation capability of the OODA loop in the network before and after deleting the node.
[0158] As the simulation system continues to change, the importance of each node will change. The credibility of the system at each historical simulation moment can be calculated by formula (1).
[0159] After calculating the credibility of the system at each historical simulation moment, the historical simulation time series data of the simulation system and the corresponding credibility are extracted to construct a training sample set. The historical simulation time series data of the simulation system includes the number of each type of simulation entity at each simulation moment and the attribute information of each simulation node. For example, the average probability of sensor node recognition of the target, the resolution of sensor node recognition of the target, the accuracy of sensor node recognition of the target, the scale of information sent by sensor node, the delay of sensor node sending information, the average information utilization ability of the command node, the information response ability of the command node (the time from receiving sensor information to sending instructions), the strike capability of the action node, the anti-interference ability of the action node, the average spatial concealment probability of the target node, the average false survival probability of the target node, the average warning time of the target node (s), the average maneuvering speed of the target node (km / h), the average vulnerable area sequence of the target node, the average working frequency band of the target node (MHz), etc.
[0160] Specifically, a sliding window with a first length is used to slide on the historical simulation time series data, and a training sample is obtained in each sliding step to construct a training sample set, wherein the input data of the sample is the simulation time series data and the corresponding credibility time series data in the sliding window, and the label of the sample is the credibility of multiple moments after the sliding window;
[0161] Considering that the credibility of the past time will affect the credibility of the future moment, the credibility of the past time has a certain reference significance for predicting the credibility of the future. Therefore, the input data of the sample includes not only the simulation time series data in the sliding window, but also the credibility of the corresponding moment, which constitutes the credibility time series data. The label of the sample includes the credibility of multiple time steps after the sliding window, which facilitates multi-step prediction.
[0162] Specifically, the constructed multi-step time series prediction neural network model includes:
[0163] A first encoding module, used for extracting features from input simulation time series data to obtain a first feature;
[0164] A second encoding module is used to extract features from the input credibility time series data to obtain a second feature;
[0165] A feature fusion module, used for fusing the first feature and the second feature to obtain a fusion feature;
[0166] The prediction unit is used to predict the multi-step credibility values of the complex simulation system based on the fusion features.
[0167] During implementation, the first encoding module is used to encode the simulation time series data in the input data of the sample, and the second encoding module is used to encode the credibility time series in the input data of the sample, capturing the long-term trend characteristics of the credibility, and obtaining the first feature and the second feature respectively. Feature fusion fuses the first feature and the second feature, and the prediction unit performs credibility prediction based on the fused feature to more accurately predict the credibility of the system at multiple moments in the future.
[0168] During implementation, the feature fusion module may concatenate the first feature and the second feature to obtain a fused feature.
[0169] Specifically, in order to effectively extract features from the simulation time series data, the first encoding module includes a multi-layer gated recurrent unit and a weighting module;
[0170] Each layer of gated recurrent units is used to extract hidden features based on input data; the input of the first layer of gated recurrent units is the simulation time series data; the hidden features extracted by each layer of gated recurrent units except the last layer are the input of the next layer of gated recurrent units; the hidden features output by the last layer of gated recurrent units are transmitted to the weighted module.
[0171] The weighted module performs weighted fusion on the hidden features output by the last layer of gated recurrent units at each time step to obtain the first feature.
[0172] Since different time steps have different influence weights on future moments, the hidden features of each time step are extracted through a multi-layer gated recurrent unit, and the hidden features of each time step are weighted and fused through a weighted module to obtain the first feature.
[0173] During implementation, the weighted module calculates the inner product of the second feature and the hidden feature of each time step, and calculates the correlation between the hidden feature of each time step and the second feature. The greater the correlation, the greater the impact on the credibility and the greater the weight. The inner product result is normalized as the weight, and the hidden feature of each time step is weighted and fused to obtain the first feature.
[0174] Specifically, the second encoding module includes a single-layer gated recurrent unit, and the following formula is used to calculate the hidden features extracted by the single-layer gated recurrent unit:
[0175]
[0176]
[0177] Z t =σ(X t W xz +H t-1 W hz +b z ) (30)
[0178] R t =σ(X t W xr +H t-1 W hr +b r ) (31)
[0179] Among them, H t represents the hidden layer features output at the tth time step, Z t represents the update gate of the tth time step, H t-1 represents the hidden layer features output at the t-1th time step, represents the candidate hidden state at the tth time step, X t represents the input data at the tth time step, R t represents the reset gate at the tth time step, W xh , W xz and W xr Both represent weight parameters, b h 、b z and b r All represent bias parameters, σ(·) represents the sigmoid function, and tanh(·) represents the tanh activation function.
[0180] By resetting the gate to capture the short-term dependencies in the time series data, and by updating the gate to capture the long-term dependencies in the time series data, the characteristics of the time series data can be better captured, thereby extracting more effective features.
[0181] The prediction unit may use an existing regression network to regress multiple credibility values. That is, the prediction unit outputs credibility values of multiple time steps to be predicted.
[0182] Specifically, the training loss of the multi-step time series prediction neural network model is calculated using the following formula:
[0183] Loss = γ1L mse +γ2L d +γ3L t (32)
[0184] Among them, L mse represents the mean square error loss, L d represents the distance loss between the prediction credibility sequence and the label credibility sequence, L t represents the time difference loss between the prediction credibility sequence and the label credibility sequence, and γ1, γ2 and γ3 represent weights.
[0185] During implementation, the training loss not only considers the difference between the label and the predicted value, but also the difference in data distance and time and space. Therefore, the training loss is calculated comprehensively by the mean square error loss, distance loss and time difference loss.
[0186] Specifically, the following formula is used to calculate the distance loss between the prediction credibility sequence and the label credibility sequence:
[0187]
[0188] Among them, Y i represents the label credibility sequence of the i-th sample, represents the prediction credibility sequence of the i-th sample, represents the distance matrix between the label credibility sequence and the prediction credibility sequence, <·,·> represents the DTW distance, γ represents the parameter, represents a set of k×k matrices, where k is the length of the label credibility sequence.
[0189] The distance loss mainly considers the similarity between two sequences, where The existing distance calculation formula can be used to calculate the distance matrix of two sequences. The value of the hth row and jth column in the matrix is the label credibility sequence Y i The hth element of The distance to the j-th element of , for example, can be the Euclidean distance.
[0190] It should be noted that the element value of matrix A is 0 or 1. If the value of the hth row and jth column in the matrix is 1, it means that the label credibility sequence Y i The hth element of The jth element of is the similarity point. If it is 0, it means the label credibility sequence Y i The hth element of The j-th element of is not a similarity point.
[0191] Since the predicted value may be different from the label value in the time dimension, the loss in the time dimension needs to be considered.
[0192] Specifically, the time difference loss is calculated using the following formula:
[0193]
[0194]
[0195] Among them, Y i represents the label credibility sequence of the i-th sample, represents the prediction credibility sequence of the i-th sample, represents the distance matrix between the label credibility sequence and the prediction credibility sequence, <·,·> represents the DTW distance, γ represents the parameter, represents a set of k×k matrices, k is the length of the label credibility sequence, and Ω is a k×k matrix.
[0196] Among them, the value of the hth row and jth column of the matrix Ω is:
[0197]
[0198] The training loss is calculated comprehensively by using mean square error loss, distance loss, and time difference loss, so that the credibility of the trained model can be evaluated more accurately, thus improving the efficiency of training.
[0199] When the training reaches a preset accuracy or a preset number of times, a trained credibility prediction model can be obtained.
[0200] For the simulation system to be predicted, the simulation time series data and credibility of the first length before the current time moment can be obtained, and the future time series data to be predicted can be input into the credibility prediction model, so that the credibility of the simulation system at multiple future moments can be quickly and accurately predicted, thereby improving the efficiency and accuracy of credibility evaluation.
[0201] Those skilled in the art will appreciate that all or part of the processes of the above-mentioned embodiments can be implemented by instructing related hardware through a computer program, and the program can be stored in a computer-readable storage medium, wherein the computer-readable storage medium is a disk, an optical disk, a read-only storage memory, or a random access memory, etc.
[0202] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by any technician familiar with the technical field within the technical scope disclosed by the present invention should be covered within the protection scope of the present invention.
Claims
1. A training method for a multi-step credibility assessment model of a simulation system, characterized in that: The following steps are involved: Construct a hypernetwork corresponding to each historical simulation moment of the simulation system; Calculate the credibility of the simulation system at each historical simulation moment based on the hypernetwork; Extracting historical simulation time series data of the simulation system, sliding a sliding window with a first length on the historical simulation time series data, obtaining a training sample in each sliding step to construct a training sample set, wherein the input data of the sample is the simulation time series data and the corresponding credibility time series data in the sliding window, and the label of the sample is the credibility of multiple moments after the sliding window; A multi-step time series prediction neural network model is constructed, and the time series prediction neural network model is trained based on the training sample set to obtain a credibility prediction model.
2. The training method of the multi-step credibility assessment model of the simulation system according to claim 1, characterized in that: The training loss of the multi-step time series prediction neural network model is calculated using the following formula: Loss=γ1L mse +γ2L d +γ3L t Among them, L mse represents the mean square error loss, L d represents the distance loss between the prediction credibility sequence and the label credibility sequence, L t represents the time difference loss between the prediction credibility sequence and the label credibility sequence, and γ1, γ2 and γ3 represent weights.
3. The training method of the multi-step credibility assessment model of the simulation system according to claim 2, characterized in that: The distance loss between the prediction credibility sequence and the label credibility sequence is calculated using the following formula: Among them, Y i represents the label credibility sequence of the i-th sample, represents the prediction credibility sequence of the i-th sample, represents the distance matrix between the label credibility sequence and the prediction credibility sequence, <·,·> represents the DTW distance, γ represents the parameter, represents a set of k×k matrices, where k is the length of the label credibility sequence.
4. The training method of the multi-step credibility assessment model of the simulation system according to claim 2, characterized in that: The time difference loss is calculated using the following formula: Among them, Y i represents the label credibility sequence of the i-th sample, represents the prediction credibility sequence of the i-th sample, represents the distance matrix between the label credibility sequence and the prediction credibility sequence, <·,·> represents the DTW distance, γ represents the parameter, represents a set of k×k matrices, k is the length of the label credibility sequence, and Ω is a k×k matrix.
5. The training method of the multi-step credibility evaluation model of the simulation system according to claim 1, characterized in that: The multi-step time series prediction neural network model includes: A first encoding module, used for extracting features from input simulation time series data to obtain a first feature; A second encoding module, used for extracting features from the input credibility time series data to obtain a second feature; A feature fusion module, used for fusing the first feature and the second feature to obtain a fusion feature; The prediction unit is used to predict the multi-step credibility values of complex simulation systems based on fusion features.
6. The training method of the multi-step credibility evaluation model of the simulation system according to claim 5, characterized in that: The first encoding module includes a multi-layer gated recurrent unit and a weighting module; Each layer of gated recurrent units is used to extract hidden features based on input data; The input of the first layer of gated recurrent units is the simulation time series data; the hidden features extracted by each layer of gated recurrent units except the last layer are the input of the next layer of gated recurrent units; the hidden features output by the last layer of gated recurrent units are transmitted to the weighting module; The weighted module performs weighted fusion on the hidden features output by the last layer of gated recurrent units at each time step to obtain the first feature.
7. The training method of the multi-step credibility evaluation model of the simulation system according to claim 1, characterized in that: Calculating the credibility of the simulation system at each historical simulation moment based on the hypernetwork includes: Calculate the importance of each node in the simulation system at each historical simulation moment; The credibility of the simulation system at each historical simulation moment is calculated based on the importance of each node.
8. The training method of the multi-step credibility evaluation model of the simulation system according to claim 7, characterized in that: The importance of each node in the simulation system at each historical simulation moment is calculated in the following way: Based on the system's hyper-network search for OODA loops at each historical simulation moment, the initial comprehensive confrontation capability of each OODA loop is calculated; For each node, treat it as a failed node and update the super network based on the network cascade failure analysis; calculate the updated comprehensive confrontation capability of each OODA loop based on the updated super network; Based on the initial comprehensive confrontation capability and the updated comprehensive confrontation capability of each OODA loop, the importance of the failed node is obtained.
9. The training method of the multi-step credibility evaluation model of the simulation system according to claim 8, characterized in that: The importance of the failed node is calculated using the following formula: in, represents the initial comprehensive confrontation capability of the jth OODA loop, The updated comprehensive confrontation capability of the jth OODA loop, represents the weight of the jth OODA loop, Value(v) represents the importance evaluation result of node v, and N T represents the number of target nodes in the simulation system, Indicates the number of OODA loops where the i-th target node is located.
10. The training method of the multi-step credibility evaluation model of the simulation system according to claim 8, characterized in that: The initial comprehensive confrontation capability of each OODA loop is calculated in the following way: Among them, P op represents the target damage probability of the OODA loop, Q op represents the stability of the topological structure of the OODA loop, w1 and w2 represent the weights of the target damage probability and topological structure stability of the OODA loop respectively.
Citation Information
Cited By
Method, device and equipment for evaluating credibility of equipment digital model
CN119475824A
