Simulation system intelligent decision-making method and system based on knowledge graph and federated learning

Abstract

The invention relates to the technical field of intelligent decision making, in particular to a simulation system intelligent decision making method and system based on a knowledge graph and federated learning, and the method comprises the steps that each simulation node constructs a knowledge graph sub-graph based on local dynamic data, and multi-modal semantic representation is generated through space-time modeling and event chain reasoning; the federal center initializes a simulation decision model architecture and issues the simulation decision model architecture to each node; each node uses a local knowledge graph to train a time sequence diagram network, extracts an equipment degradation path and abnormal propagation characteristics, and uploads gradient parameters in combination with homomorphic encryption; the federation center fuses the multi-node model through a dynamic weight aggregation algorithm to generate a global decision model; and driving knowledge graph evolution based on real-time data, and performing causal reasoning and decision optimization on an event chain through a federal model. According to the method, the problem of insufficient dynamic decision adaptability in a multi-source data island and privacy sensitive scene in a simulation system is solved.

Description

Technical Field

[0001] The present invention relates to the field of intelligent decision-making technology, and specifically to an intelligent decision-making method and system for a simulation system based on knowledge graph and federated learning. Background Art

[0002] In the prior art, the intelligent decision-making of simulation systems mostly relies on centralized knowledge graphs and independently trained decision models. Traditional federated learning frameworks have significant limitations when processing heterogeneous simulation data. Due to differences in spatiotemporal characteristics and protocol incompatibility, the constructed knowledge graph subgraphs have semantic gaps; data from different nodes are difficult to align in the spatiotemporal dimensions, and different protocols also make it difficult to unify semantics, making cross-node relational reasoning and knowledge fusion difficult. At the same time, existing privacy protection mechanisms will destroy the structural integrity of spatiotemporal correlation features while ensuring data security. Conventional encryption methods disrupt the original spatiotemporal correlation of data, resulting in the model being unable to accurately capture equipment degradation paths and abnormal propagation laws. In addition, static knowledge graphs cannot adapt to dynamic simulation environments, decision-making strategies lag behind real-time situation changes, and lack adaptability when responding to sudden failure chain reactions.

[0003] However, existing methods face technical contradictions in solving the above problems. Although the centralized knowledge graph construction method based on data sharing improves decision consistency through a unified data lake, it increases the risk of sensitive data leakage; and the federated learning framework that relies too much on local training protects privacy through parameter isolation, it cannot establish a causal network of cross-domain events due to the semantic fragmentation of the knowledge graph between nodes. In addition, when the traditional time series model runs in the encrypted domain, it is difficult to effectively extract the spatiotemporal dependencies of multimodal features because it cannot parse the spatiotemporal topological structure after the disturbance; dynamic graph evolution and federated model updates often adopt independent mechanisms and lack collaborative optimization design, causing decision optimization to lag behind real-time environmental changes. Existing technologies have not yet achieved an effective balance between the intensity of privacy protection, the efficiency of multi-source knowledge fusion, and the real-time nature of dynamic decision-making, which seriously restricts the intelligence level of complex simulation systems. Summary of the invention

[0004] In order to solve the problems of multi-source data islands in simulation systems and insufficient adaptability of dynamic decision-making in privacy-sensitive scenarios, the present invention provides an intelligent decision-making method and system for simulation systems based on knowledge graphs and federated learning.

[0005] In the first aspect, the present invention provides an intelligent decision-making method for a simulation system based on knowledge graph and federated learning, which adopts the following technical solution: An intelligent decision-making method for a simulation system based on knowledge graph and federated learning, comprising: Each simulation node builds a knowledge graph subgraph based on local dynamic data and generates multimodal semantic representation through spatiotemporal modeling and event chain reasoning; The federal center initializes the simulation decision model architecture and distributes it to each node; Each node uses the local knowledge graph to train the temporal graph network, extracts the device degradation path and abnormal propagation characteristics, and uploads the gradient parameters by combining homomorphic encryption; The federal center fuses the model parameters of multiple nodes through the dynamic weight aggregation algorithm to generate a global decision model; Based on real-time data-driven knowledge graph evolution, causal reasoning and decision optimization are performed on the event chain through the federal model.

[0006] Furthermore, the construction of the knowledge graph subgraph by each simulation node based on local dynamic data includes: Collect the dynamic data of the device operation status, environmental parameters and interaction events in the simulation system; Perform spatio-temporal alignment processing on the dynamic data to construct an event sequence in the spatio-temporal dimension; Establish cross-modal association relationships of device-event-environment through event chain reasoning, and use the preset event chain reasoning rules to analyze the logical connections between device state changes, event occurrences and environmental factors.

[0007] Furthermore, the spatio-temporal alignment processing includes: Establish a unified spatio-temporal coordinate system; Divide the space of the simulation system into several grids, assign a unique code to each grid, map the spatial coordinates of the device and the event into these grids, and form a spatial coding vector; Generate a spatio-temporal fusion feature tensor. Divide the decision-making period at a certain time interval, fuse the time and space information, and generate a tensor containing spatio-temporal features for subsequent analysis and processing.

[0008] Furthermore, the federal center initializes the simulation decision model architecture and distributes it to each node, including: Design a federated learning model architecture based on the temporal graph network; Send the initial model parameters and training specifications to each edge node through a secure communication protocol.

[0009] Furthermore, each node uses the local knowledge graph to train the temporal graph network, and the extraction of the device degradation path and abnormal propagation includes: Use the degradation state of the device as a node, the situation of abnormal propagation as an edge, and assign corresponding weights to the edge to construct a temporal graph containing the node of the device degradation state and the weight of the abnormal propagation edge; Design a bi-directional graph attention network to capture the characteristics of the abnormal propagation path; Combine the time convolutional layer to extract the device degradation trend characteristics.

[0010] Furthermore, the uploading of the gradient parameters by combining homomorphic encryption includes: After the node completes the timing diagram network training, the gradient parameters obtained through training are calculated by the backpropagation algorithm; The Paillier homomorphic encryption algorithm is used for parameter encryption; The encrypted gradient parameters are uploaded to the federal center.

[0011] Furthermore, the dynamic weight aggregation algorithm includes: Calculate the confidence weights of the model parameters of each node, where the confidence weights are based on the KL divergence between the local data distribution and the global distribution; Set the time decay coefficient according to the freshness of the node data; Use differential privacy technology for secure parameter aggregation; Use an adaptive momentum optimizer to update the parameters of the global model according to the aggregated gradient parameters to improve the performance of the model.

[0012] Furthermore, the knowledge graph evolution includes: Real-time monitor device status jump events and generate a graph update trigger signal; Process the relationship reasoning of new nodes and edges through an incremental graph neural network; Use a lightweight graph pruning algorithm to maintain the spatio-temporal consistency of the knowledge graph.

[0013] Furthermore, the causal reasoning and decision optimization include: Construct a causal graph model including equipment degradation, anomaly propagation, and environmental impact; Use the federal decision-making model to generate multiple possible decision-making schemes and evaluate them; Select the set of Pareto-optimal decision-making strategies, that is, the set of strategies that cannot further improve a certain goal without damaging other goals.

[0014] In a second aspect, an intelligent decision-making system of a simulation system based on a knowledge graph and federated learning includes: A data acquisition module, deployed on each simulation node, for obtaining device operation status, environmental parameters, and dynamic data of interaction events; A knowledge graph construction module, connected to the data acquisition module, for generating spatio-temporally aligned multi-modal semantic representations and constructing a local knowledge graph subgraph; A federated model training module, connected to the knowledge graph construction module, for training the local model and performing regular evaluations in the global model; A privacy computing module, connected to the model training and evaluation module, integrating homomorphic encryption and differential privacy components, for secure parameter transmission and aggregation to ensure data privacy; The Federal Coordination Module, connected to the Privacy Computing Module, includes a central server and node agents, and is responsible for model architecture initialization, parameter distribution, and multi-node training synchronization; The Atlas Evolution Module, connected to the Data Acquisition Module, integrates an incremental graph neural network and a version control unit, and is used to maintain the spatio-temporal topological consistency of the dynamic knowledge atlas; The Dynamic Decision Engine Module, connected to the Federal Coordination Module and the Atlas Evolution Module, includes a causal reasoning unit and a reinforcement learning optimizer, and is used to generate Pareto optimal combat strategies; The Decision Output Module, connected to the Dynamic Decision Engine Module, outputs various decision-making information and optimal decisions predicted by the federal decision-making model.

[0015] Thirdly, the present invention provides a computer-readable storage medium, in which multiple instructions are stored, and the instructions are suitable for being loaded and executed by a processor of a terminal device for the method for constructing a knowledge atlas related to the situation and action of maritime equipment.

[0016] Fourthly, the present invention provides a terminal device, including a processor and a computer-readable storage medium. The processor is used to implement each instruction; the computer-readable storage medium is used to store multiple instructions, and the instructions are suitable for being loaded and executed by the processor for the method for constructing a knowledge atlas related to the situation and action of maritime equipment.

[0017] In summary, the present invention has the following beneficial technical effects: 1. The intelligent decision-making method and system for a simulation system based on a knowledge atlas and federated learning proposed by the present invention can effectively integrate and analyze multi-source data of the simulation system. By constructing knowledge atlas subgraphs through each simulation node, fusing multi-node models by the federation center, etc., a decision-making model that can intelligently reflect the dynamic situation of the simulation system is generated. Compared with the traditional method that relies on a centralized knowledge atlas and an independently trained decision-making model, the present invention breaks the multi-source data silos and greatly improves the data utilization efficiency; compared with the method that simply relies on local training, the constructed decision-making model has more comprehensive information and higher accuracy, and can better adapt to the complex and changeable simulation environment.

[0018] 2. The spatio-temporal modeling, event chain reasoning, dynamic weight aggregation algorithm, etc. adopted by the present invention provide a unified and intelligent processing method for complex and changeable simulation data. Spatio-temporal modeling and event chain reasoning endow the data with multi-modal semantic representations, which is conducive to mining the logical relationships behind the data; the dynamic weight aggregation algorithm can effectively fuse multi-node models and improve the performance of the global decision-making model. This not only solves the problems such as semantic gap and inability to accurately capture the equipment degradation path existing in the traditional method when dealing with heterogeneous simulation data, but also deeply mines and effectively integrates the data, improving the scientificity and accuracy of subsequent decision-making.

[0019] 3. The present invention is based on real-time data-driven knowledge graph evolution, and performs causal reasoning and decision optimization on the event chain through a federated model. This process completely covers the entire process from data collection, knowledge graph construction, model training to decision optimization, enabling the constructed decision-making system to comprehensively and accurately reflect various situations of the simulation system and the causal relationships between events. It can provide rich and accurate knowledge support for the decision-making of the simulation system, helping decision-makers make more reasonable and effective decisions.

[0020] 4. The present invention combines privacy protection mechanisms such as uploading gradient parameters with homomorphic encryption and using differential privacy technology to achieve efficient knowledge fusion and dynamic decision-making while ensuring data privacy. In privacy-sensitive scenarios, it can not only effectively protect the sensitive data of each node from being leaked, but also achieve multi-node collaboration and knowledge sharing through federated learning, enabling the decision-making system to adapt to the dynamic changes of the simulation system in real time, greatly improving the timeliness and adaptability of decision-making, and helping to enhance the overall intelligent level and application value of the simulation system. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 is the flowchart of the method of Embodiment 1 of the present invention; Figure 2 is the schematic structural diagram of the system of Embodiment 2 of the present invention; DETAILED DESCRIPTION OF THE EMBODIMENTS

[0022] The present invention will be further described in detail below with reference to the accompanying drawings.

[0023] Embodiment 1 Referring to Figure 1 , a method for intelligent decision-making of a simulation system based on a knowledge graph and federated learning in this embodiment includes: S1. Each simulation node constructs a knowledge graph subgraph based on local dynamic data, and generates multi-modal semantic representations through spatio-temporal modeling and event chain reasoning, including: S11. Collect dynamic data of device operation status, environmental parameters and interaction events in the simulation system; S12. Perform spatio-temporal alignment processing on the dynamic data to construct an event sequence in the spatio-temporal dimension, including: S121. Establish a unified spatio-temporal coordinate system; Establish a spatio-temporal reference to ensure the consistency of data of each combat unit in terms of time and space. In terms of time synchronization, the timestamps of heterogeneous data sources are calibrated through a sliding window to ensure that the clock deviation of each combat unit is controlled within a very small range. In terms of the selection of the spatial coordinate system, taking the geometric center position of the simulation environment as the origin, a plane rectangular coordinate system is established, so that the spatial positions of each of our units and targets can be accurately represented.

[0024] S122. Divide the space of the simulation system into several grids, assign a unique code to each grid, and map the spatial coordinates of devices and events into these grids to form a spatial coding vector; According to the spatial range and accuracy requirements of the simulation system, divide the entire space into grids with uniform size. For a simulation system on a two-dimensional plane, divide it into 100×100 grids. Use the row-column numbering method to assign a unique code to each grid. For the spatial coordinates of devices and events, map them to the corresponding grids according to their positions. If the coordinates of a device fall within the grid in the 3rd row and 5th column, the spatial coding vector of this device is 0305.

[0025] S123. Generate a spatio-temporal fusion feature tensor. Divide the decision-making period at a certain time interval, fuse the time and space information to generate a tensor containing spatio-temporal features for subsequent analysis and processing; According to the actual requirements of the simulation, determine an appropriate decision-making period, which can be 1 minute as a decision-making period. Combine the time information within each decision-making period with the spatial coding vector. For the information such as device state changes and events occurring within each decision-making period, integrate their corresponding time and space information to form a multi-dimensional spatio-temporal fusion feature tensor.

[0026] S13. Establish a cross-modal association relationship of device - event - environment through event chain reasoning, and use the preset event chain reasoning rules to analyze the logical connection between device state changes, event occurrences, and environmental factors; According to the domain knowledge and historical data of the simulation system, preset a series of event chain reasoning rules. These rules can be expressed in the form of condition - conclusion. Input the collected device operation status, environmental parameters, and interaction event data into the event chain reasoning engine, and perform reasoning according to the preset rules. Through reasoning and analysis, find out the causal relationship and sequence between device state changes, event occurrences, and environmental factors, and establish a cross-modal association relationship of device - event - environment.

[0027] S2. The federal center initializes the simulation decision model architecture and distributes it to each node, including: S21. Design a federated learning model architecture based on a sequence diagram network; Design a model architecture suitable for battlefield decision-making. The input layer is responsible for processing the spatio-temporal feature vectors collected from various combat units and converting them into a format that the model can handle. The core module consists of multiple parts. The bidirectional graph convolutional network is used to analyze the cooperation relationships between various combat units, find out their mutual influences and dependencies; the spatio-temporal attention mechanism can capture the situation changes in key areas of the battlefield and focus on important information; the LSTM prediction layer can deduce the action trajectories of the enemy based on historical data and provide forward-looking information for our decision-making. The output layer generates a tactical advice vector according to the analysis results of the core module, and this vector contains various combat instructions.

[0028] S22. Distribute the initial model parameters and training specifications to each edge node through a secure communication protocol; To ensure the secure transmission of model parameters, advanced encryption technologies are adopted. Quantum key distribution technology is selected to transmit the initial model. This technology has extremely high security and can effectively prevent information from being stolen. Before transmission, the national cryptographic algorithm is used to encrypt the model parameters, and at the same time, the hash algorithm is used to generate a check code for verifying the integrity of the data. After each simulation unit receives the encrypted model, it decrypts it with the corresponding key and verifies whether the data has been tampered with during transmission through the check code.

[0029] S3. Each node uses the local knowledge graph to train the temporal graph network, extracts the device degradation path and abnormal propagation characteristics, and uploads the gradient parameters in combination with homomorphic encryption, including: S31. Use the degradation state of the device as nodes and the situation of abnormal propagation as edges, and assign corresponding weights to the edges to construct a temporal graph containing device degradation state nodes and abnormal propagation edge weights; At the edge node, construct a combat temporal graph. The nodes in the graph represent various states and parameters related to combat, and the edges represent the relationships between these nodes. These relationships can be represented by numerical values between 0 and 1, and the larger the value, the closer the relationship. By constructing a combat temporal graph, the changes and mutual relationships of various factors during the combat process can be intuitively displayed.

[0030] S32. Design a bidirectional graph attention network to capture the characteristics of abnormal propagation paths; Construct a bidirectional graph attention network, which includes two propagation directions: forward and backward. In each propagation direction, the graph attention mechanism is used to calculate the attention weights between nodes to capture the important relationships between nodes. Through bidirectional propagation, the characteristics of abnormal propagation paths can be captured more comprehensively. In the forward propagation, calculate the attention weights of each node to its neighbor nodes; in the backward propagation, calculate the attention weights of neighbor nodes to the current node. The attention weights in the two directions are fused to obtain the final node representation.

[0031] S33. Combine the temporal convolutional layer to extract the device degradation trend features; Based on the bidirectional graph attention network, add a temporal convolutional layer. The temporal convolutional layer can perform convolutional operations on the time series data of the device to extract the long-term dependence features of the device degradation trend. Select appropriate convolutional kernel sizes and strides to perform convolutional processing on the device status data. Integrate the features extracted by the temporal convolutional layer with the features extracted by the bidirectional graph attention network to obtain a more comprehensive device feature representation.

[0032] S34. After the node completes the training of the temporal graph network, use the backpropagation algorithm to calculate the gradient parameters obtained from the training; Use the local knowledge graph data on the node to train the temporal graph network. During the training process, use the backpropagation algorithm to calculate the gradient of the loss function with respect to the model parameters. Calculate the value of the loss function based on the training data and the output of the model. Then, starting from the loss function, use the chain rule to calculate the gradient of each model parameter in reverse. Store the calculated gradient parameters locally for subsequent encryption and uploading.

[0033] S35. Adopt the Paillier homomorphic encryption algorithm for parameter encryption; Select the Paillier homomorphic encryption algorithm to encrypt the calculated gradient parameters. First, generate the public key and private key of the Paillier encryption algorithm. The public key is used to encrypt the gradient parameters, and the private key is securely saved by the node. Convert the gradient parameters into a data format suitable for encryption, and use the public key to perform the encryption operation on the gradient parameters. The encrypted gradient parameters can still perform certain mathematical operations while maintaining data privacy, meeting the requirements of federated learning.

[0034] S36. Upload the encrypted gradient parameters to the federated center; Through a secure network channel, upload the encrypted gradient parameters to the federated center. During the upload process, ensure the integrity and reliability of the data. A data verification mechanism can be used to verify the uploaded data to prevent the data from being tampered with during transmission. After receiving the encrypted gradient parameters, the federated center performs corresponding processing and verification.

[0035] S4. The federated center fuses the model parameters of multiple nodes through the dynamic weight aggregation algorithm to generate a global decision-making model, including: S41. Calculate the confidence weights of the model parameters of each node, where the confidence weights are based on the KL divergence between the local data distribution and the global distribution; Calculate the KL divergence between the local data distribution and the global data distribution for each node. The KL divergence can measure the degree of difference between two probability distributions. The smaller the KL divergence, the closer the local data distribution is to the global data distribution, and the higher the confidence weight of the node.

[0036] S42. Set the time decay coefficient according to the freshness of node data; Set a time tag for the data of each node to record the data collection time. Calculate the freshness of the data based on the difference between the data collection time and the current time. Set a time decay function to calculate the time decay coefficient according to the freshness of the data. Multiply the time decay coefficient by the confidence weight of the node to obtain the final node weight.

[0037] S43. Use differential privacy technology for secure parameter aggregation; When performing parameter aggregation, use differential privacy technology to protect the privacy of node data. During the aggregation process, add Gaussian noise to the aggregation result so that even if an attacker obtains the aggregated parameters, they cannot infer the data information of a single node. Determine the variance of the Gaussian noise according to the parameters and sensitivity of differential privacy. Multiply the gradient parameters of each node by its final weight and then sum them, and then add Gaussian noise to obtain the aggregated gradient parameters.

[0038] S44. Use an adaptive momentum optimizer to update the parameters of the global model according to the aggregated gradient parameters to improve the performance of the model; Select an adaptive momentum optimizer to update the parameters of the global model. Update the parameters of the global model according to the aggregated gradient parameters and the parameters of the optimizer. During the update process, the adaptive momentum optimizer can automatically adjust the learning rate to accelerate the convergence speed of the model. Through multiple iterative updates, continuously optimize the parameters of the global model to improve the performance of the model.

[0039] S5. Based on real-time data-driven knowledge graph evolution, perform causal reasoning and decision optimization on the event chain through a federated model, including: S51. Real-time monitor device status jump events and generate a graph update trigger signal; Set up a real-time monitoring module in the simulation system to monitor the status of the device in real time. When it detects a jump in the status of the device, generate a graph update trigger signal. The signal contains the identification of the device, status change information, occurrence time, etc. Send the trigger signal to the knowledge graph evolution module to start the update process of the knowledge graph.

[0040] S52. Process the relationship reasoning of new nodes and edges through an incremental graph neural network; After receiving the graph update trigger signal, use the incremental graph neural network to process the relationship reasoning of newly added nodes and edges. The incremental graph neural network can quickly learn the relationships of newly added nodes and edges without retraining the entire model. Add the newly added device status information, event information, etc. as new nodes and edges to the knowledge graph. Use the incremental graph neural network to reason about the relationships of the newly added nodes and edges, and update the structure and attributes of the knowledge graph.

[0041] S53. Adopt a lightweight graph pruning algorithm to maintain the spatio-temporal consistency of the knowledge graph; As the knowledge graph is continuously updated, some redundant or inconsistent information may appear. Use a lightweight graph pruning algorithm to optimize the knowledge graph. This algorithm can identify and delete nodes and edges that have little impact on the spatio-temporal consistency of the knowledge graph. Determine the nodes and edges to be pruned according to the importance indicators of the nodes and edges. Through the pruning operation, maintain the spatio-temporal consistency of the knowledge graph and improve the quality and query efficiency of the knowledge graph.

[0042] S54. Construct a causal graph model including equipment degradation, anomaly propagation, and environmental impact; Based on the device status information, event information, and environmental parameters in the knowledge graph, construct a causal graph model including equipment degradation, anomaly propagation, and environmental impact. Take the degradation state of the device, anomaly propagation events, and environmental factors as nodes, and the causal relationships between them as edges. Determine the direction and weight of the edges according to event chain reasoning and historical data.

[0043] S55. Use the federated decision-making model to generate multiple possible decision-making schemes and evaluate them; Input the causal graph model and real-time data into the federated decision-making model. The federated decision-making model can generate multiple possible decision-making schemes according to the causal relationships and real-time data. Then evaluate each decision-making scheme, considering factors such as the cost, effect, and risk of the decision. Use a multi-objective optimization algorithm to evaluate the decision-making schemes and obtain the evaluation scores of each scheme.

[0044] S56. Select the set of Pareto-optimal decision-making strategies, that is, the set of strategies that cannot further improve a certain objective without harming other objectives; According to the evaluation scores of the decision-making schemes, select the set of Pareto-optimal decision-making strategies. The set of Pareto-optimal decision-making strategies refers to the set of strategies that cannot further improve a certain objective without harming other objectives. By comparing the evaluation scores of different decision-making schemes, find the schemes that perform well in multiple objectives and form the set of Pareto-optimal decision-making strategies. Use this set as the final decision result and apply it to the simulation system.

[0045] This embodiment provides an intelligent decision-making method for a simulation system based on a knowledge graph and federated learning. The use of this method is in the context that existing simulation decision-making technologies are difficult to cope with the difficulties of multi-source heterogeneous data fusion, insufficient data privacy protection, and inability to adapt to dynamic changes in real time. In order to make accurate intelligent decisions quickly in real-time situation data, each simulation node first collects local dynamic data such as device operating status, environmental parameters, and interaction events, and performs spatio-temporal alignment processing on the data through operations such as establishing a unified spatio-temporal coordinate system, dividing spatial grids, and generating spatio-temporal fusion feature tensors, constructing an event sequence in the spatio-temporal dimension, and then using event chain reasoning to establish cross-modal association relationships between devices, events, and environments to generate a sub-knowledge graph. The federated center designs a federated learning model architecture based on a temporal graph network and distributes the initial model parameters and training specifications to each edge node through a secure communication protocol. Each node uses the local knowledge graph to train the temporal graph network, extracts device degradation paths and abnormal propagation features, calculates gradient parameters through the backpropagation algorithm, and uploads them to the federated center after encryption using the Paillier homomorphic encryption algorithm. The federated center calculates the confidence weights of the model parameters of each node, sets a time decay coefficient, performs secure parameter aggregation using differential privacy technology, and then updates the global model parameters using an adaptive momentum optimizer to generate a global decision-making model. At the same time, based on real-time data-driven knowledge graph evolution, real-time monitoring of device state jump events triggers graph updates, uses incremental graph neural networks to process the relationship reasoning of new nodes and edges, and uses a lightweight graph pruning algorithm to maintain the spatio-temporal consistency of the knowledge graph. A causal graph model including device degradation, abnormal propagation, and environmental impact is constructed, and multiple possible decision-making schemes are generated and evaluated using the federated decision-making model, and finally a set of Pareto-optimal decision-making strategies is selected, greatly improving the real-time performance and accuracy of decision-making. And the use of this method and system is also to make accurate and real-time intelligent decisions on the massive real-time situation data of a large number of combat units and a large number of combat rules in actual scenarios, and then can provide accurate, real-time, and effective decision-making support for application systems such as battlefield command and control systems based on knowledge graphs, improving the effectiveness and real-time performance of battlefield decision-making application systems.

[0046] Embodiment 2 Referring to Figure 2 , this embodiment provides a knowledge graph construction system for the association between the situation and actions of maritime equipment, including: A data collection module, deployed on each simulation node, for obtaining dynamic data of device operating status, environmental parameters, and interaction events; A knowledge graph construction module, connected to the data collection module, for generating spatio-temporally aligned multi-modal semantic representations and constructing a local sub-knowledge graph; A federated model training module, connected to the knowledge graph construction module, for training the local model and performing regular evaluations in the global model; The privacy computing module is connected to the federated model training and evaluation module, integrating homomorphic encryption and differential privacy components, for secure parameter transmission and aggregation to ensure data privacy; The federated coordination module is connected to the privacy computing module, including a central server and node agents, responsible for model architecture initialization, parameter distribution, and multi-node training synchronization; The graph evolution module is connected to the data acquisition module, integrating incremental graph neural networks and version control units, for maintaining the spatio-temporal topological consistency of the dynamic knowledge graph; The dynamic decision-making engine module is connected to the federated coordination module and the graph evolution module, including a causal reasoning unit and a reinforcement learning optimizer, for generating Pareto optimal combat strategies; The decision output module is connected to the dynamic decision-making engine module, outputting various decision-making information and the optimal decision predicted by the federated decision model.

[0047] A computer-readable storage medium stores multiple instructions, which are adapted to be loaded and executed by a processor of a terminal device for the method for constructing a knowledge graph related to the situation and action of maritime equipment as described above.

[0048] A terminal device includes a processor and a computer-readable storage medium. The processor is used to implement each instruction; the computer-readable storage medium is used to store multiple instructions, which are adapted to be loaded and executed by the processor for the method for constructing a knowledge graph related to the situation and action of maritime equipment as described above.

[0049] The above are all preferred embodiments of the present invention. Without restricting the protection scope of the present invention accordingly, therefore: All equivalent changes made according to the structure, shape, and principle of the present invention shall be covered within the protection scope of the present invention.

Claims

1. An intelligent decision-making method for a simulation system based on knowledge graph and federated learning, characterized in that: include: Each simulation node builds a knowledge graph subgraph based on local dynamic data and generates multimodal semantic representation through spatiotemporal modeling and event chain reasoning; The federation center initializes the simulation decision model architecture and sends it to each node; Each node uses the local knowledge graph to train the time-series graph network, extracts the device degradation path and abnormal propagation characteristics, and uploads the gradient parameters in combination with homomorphic encryption; The federation center fuses multi-node model parameters through a dynamic weight aggregation algorithm to generate a global decision model; Based on the evolution of knowledge graph driven by real-time data, causal reasoning and decision optimization are performed on the event chain through the federated model.

2. According to claim 1, a simulation system intelligent decision-making method based on knowledge graph and federated learning is characterized in that: Each simulation node constructs a knowledge graph subgraph based on local dynamic data, including: Collect dynamic data of equipment operation status, environmental parameters and interactive events in the simulation system; Perform time-space alignment processing on dynamic data to construct event sequences in time and space dimensions; Establish cross-modal association relationships through event chain reasoning and construct dynamic knowledge graph subgraphs.

3. According to claim 2, a simulation system intelligent decision-making method based on knowledge graph and federated learning is characterized in that: The spatiotemporal alignment process includes: Establish a unified space-time coordinate system; The space of the simulation system is divided into several grids, each grid is assigned a unique code, and the spatial coordinates of the equipment and events are mapped to these grids to form a spatial coding vector; Generate spatiotemporal fusion feature tensor; The decision cycle is divided into certain time intervals, and the time and space information are integrated to generate a tensor containing time and space features, which is convenient for subsequent analysis and processing.

4. According to claim 1, a simulation system intelligent decision-making method based on knowledge graph and federated learning is characterized in that: The federation center initializes the simulation decision model architecture and sends it to each node, including: Design a federated learning model architecture based on a temporal graph network; The initial model parameters and training specifications are sent to each edge node through a secure communication protocol.

5. According to claim 1, a simulation system intelligent decision-making method based on knowledge graph and federated learning is characterized in that: Each node uses the local knowledge graph to train the time series graph network to extract the equipment degradation path and abnormal propagation features, including: The degradation state of the equipment is taken as a node, the abnormal propagation situation is taken as an edge, and the edges are assigned corresponding weights to construct a time series graph containing equipment degradation state nodes and abnormal propagation edge weights; Design a bidirectional graph attention network to capture the characteristics of abnormal propagation paths; Combined with the temporal convolution layer to extract the device degradation trend characteristics.

6. According to claim 1, a simulation system intelligent decision-making method based on knowledge graph and federated learning is characterized in that: The uploading of gradient parameters in combination with homomorphic encryption includes: After the node completes the timing graph network training, the gradient parameters obtained through the training are calculated through the back propagation algorithm; The Paillier homomorphic encryption algorithm is used for parameter encryption; Upload the encrypted gradient parameters to the federation center.

7. The intelligent decision-making method for a simulation system based on knowledge graph and federated learning according to claim 1, characterized in that: The dynamic weight aggregation algorithm includes: Calculate the confidence weight of each node model parameter, where the confidence weight is based on the KL divergence of the local data distribution and the global distribution; Set the time decay coefficient according to the freshness of node data; Use differential privacy technology to aggregate security parameters; Using the adaptive momentum optimizer, the parameters of the global model are updated according to the aggregated gradient parameters to improve the performance of the model.

8. The intelligent decision-making method for a simulation system based on knowledge graph and federated learning according to claim 1, characterized in that: The knowledge graph evolution includes: Monitor device status transition events in real time and generate graph update trigger signals; Process the relationship reasoning of newly added nodes and edges through incremental graph neural network; A lightweight graph pruning algorithm is used to maintain the spatiotemporal consistency of the knowledge graph.

9. The intelligent decision-making method for a simulation system based on knowledge graph and federated learning according to claim 1, characterized in that: The causal reasoning and decision optimization include: Construct a causal graph model that includes equipment degradation, anomaly propagation, and environmental impact; Using the federated decision-making model, multiple possible decision-making scenarios are generated and evaluated; Choose a Pareto-optimal set of decision strategies, that is, a set of strategies that cannot further improve one goal without compromising other goals.

10. An intelligent decision-making system for a simulation system based on knowledge graph and federated learning, comprising: The data acquisition module is deployed in each simulation node to obtain the dynamic data of equipment operation status, environmental parameters and interactive events; The knowledge graph construction module is connected to the data acquisition module to generate spatiotemporally aligned multimodal semantic representations and construct local knowledge graph subgraphs; The federated model training module is connected to the knowledge graph building module to train the local model and perform regular evaluation on the global model; The privacy computing module is connected to the model training and evaluation module, integrating homomorphic encryption and differential privacy components for secure parameter transmission and aggregation to ensure data privacy; The federated coordination module is connected to the privacy computing module and includes a central server and node agents. It is responsible for model architecture initialization, parameter distribution, and multi-node training synchronization. The graph evolution module is connected to the data acquisition module and integrates the incremental graph neural network and version control unit to maintain the spatiotemporal topological consistency of the dynamic knowledge graph; The dynamic decision engine module is connected to the federation coordination module and the graph evolution module, and contains a causal reasoning unit and a reinforcement learning optimizer to generate Pareto optimal combat strategies; The decision output module is connected to the dynamic decision engine module to output various decision information and optimal decisions predicted by the federal decision model.

Images