Distributed intelligent inference system and method for multi-source data fusion and dynamic scheduling

Abstract

The application discloses a multi-source data fusion and dynamic scheduling distributed intelligent deduction system and method, and relates to the technical field of deduction simulation. A deduction preprocessing module divides a military scenario into task stages and sets decision points, generates branch tasks according to tactical rules and classifies models; a deduction management module advances deduction from an initial state based on a simulation engine, generates branch tasks according to real-time states at the decision points, combines historical performance parameters and model types to distribute initial weight coefficients, realizes parallel deduction and dynamic model switching; a deduction optimization module continuously updates weight coefficients through branch performance parameters, filters optimal deduction paths and closes loop iteration state snapshots; a resource regulation module collects data in real time to support performance evaluation, dynamically optimizes resource allocation according to task load, significantly improves deduction efficiency and accuracy, and realizes intelligent resource scheduling and system adaptive optimization.

Description

Technical Field

[0001] This invention relates to the field of simulation technology, and in particular to a distributed intelligent simulation system for multi-source data fusion and dynamic scheduling. Background Technology

[0002] With the significant increase in the complexity of modern warfare, military simulation and exercise systems are gradually evolving from centralized, pre-scripted models to distributed, intelligent, and high-concurrency systems. This transformation mainly relies on three technological pillars: a multi-level joint simulation framework, a decision-making mechanism centered on intelligent agents and reinforcement learning, and a high-performance distributed parallel computing architecture. By establishing a cross-domain heterogeneous protocol interconnection environment and integrating methods such as digital twins and aggregation / deaggregation, the system achieves integrated modeling and dynamic integration of multi-domain combat elements, including land, sea, air, electronic, and cyber domains. This significantly enhances the system's expressive power and operational efficiency in simulating large-scale, multi-service joint operations, providing crucial support for military training, tactical verification, and equipment effectiveness analysis.

[0003] Current distributed intelligent simulation and inference systems are mainly based on multi-level simulation frameworks, distributed resource scheduling, and agent decision-making technologies. They coordinate models of different granularities through aggregation and deaggregation, and rely on heterogeneous protocols to achieve data interoperability across multiple platforms, constructing a unified inference environment. At the intelligent decision-making level, although hierarchical reinforcement learning and multi-model agent structures have been introduced to enhance the autonomy of entity behavior, their core learning architecture still largely adopts a centralized deep reinforcement learning model centered on a single server. This model requires simulation nodes to continuously interact with the central server at high frequency, generating significant communication overhead and making the central node a performance bottleneck, limiting system scalability and real-time decision-making capabilities. Although asynchronous communication, GPU acceleration, and database middleware have been combined to improve concurrent processing and simulation rendering performance, and network attack and defense simulation and post-mortem evaluation mechanisms have been integrated to enhance the realism of adversarial scenarios, the fundamental bottleneck of the centralized learning architecture still restricts system efficiency in ultra-large-scale inference scenarios. How to achieve an effective balance between intelligent decision-making capabilities and system scalability remains a key challenge for current technology.

[0004] For example, patent application CN113656963B discloses a distributed combat simulation system with real-time interactive control, comprising: a front-end graphical interface, various simulation logic units, and a simulation engine. The graphical front-end is responsible for real-time display and control of the simulation process; different simulation logic units in the simulation system are managed by different simulators as simulation components. Each simulator contains specific simulation units within the simulation scenario, and simulators are divided into physical simulators and functional simulators. These simulators can perform distributed collaborative simulations via a network; the simulation engine drives the execution of the entire simulation process and acts as the central hub for the front-end graphical interface to control the specific simulation units within the physical simulators.

[0005] For example, the simulation system, method, device, and storage medium for air-sea swarm confrontation disclosed in patent application CN113705102B include: deploying a deep reinforcement learning system on a server, deploying the simulation system on multiple computing nodes, connecting the multiple computing nodes to the server via a network, and running multiple simulation system instances in the simulation system of each computing node, thereby constructing a parallel distributed network architecture.

[0006] However, in the process of implementing the inventive technical solution in the embodiments of this application, it was found that the above-mentioned technology has at least the following technical problems:

[0007] In existing technologies, when conducting large-scale distributed combat simulations, the centralized deep reinforcement learning architecture with a single server as the central node is commonly used. This forces all simulation nodes to engage in continuous and high-frequency data interaction with the central server, resulting in enormous communication pressure. Consequently, the central node becomes a bottleneck in system performance, severely limiting the scalability of the simulation and the efficiency of real-time decision-making. Summary of the Invention

[0008] This application provides a distributed intelligent simulation system and method for multi-source data fusion and dynamic scheduling. It solves the problem that in large-scale distributed combat simulations, the centralized deep reinforcement learning architecture with a single server as the central node is commonly used. This results in all simulation nodes having to engage in continuous and high-frequency data interaction with the central server, causing huge communication pressure. The central node becomes a bottleneck in system performance, severely limiting the scalability of the simulation and the efficiency of real-time decision-making. The method significantly improves the efficiency and accuracy of the simulation and realizes intelligent resource scheduling and adaptive system optimization.

[0009] This application provides a distributed intelligent simulation system for multi-source data fusion and dynamic scheduling, including: a simulation preprocessing module, a simulation management module, a simulation optimization module, and a resource control module. The simulation preprocessing module is used to divide the military scenario of the distributed simulation into task phases, set decision points within each task phase based on tactical rules, generate branch tasks based on different decisions at each decision point, and perform initial model classification based on the quality attribute parameters of each simulation model. The task phases include an early warning phase, a mid-course interception phase, and a terminal defense phase. The simulation management module is used to advance the simulation from the initial state snapshot using the simulation engine of each simulation model. When a decision point is reached, multiple [items] are generated based on the current snapshot state. The system divides tasks into branches and assigns initial weight coefficients to each branch based on historical performance parameters and model type. Then, it distributes each branch task to computing nodes for parallel simulation. The system dynamically triggers model classification switching in real time based on the driving events of each simulation model. The simulation optimization module updates the weight coefficients of each branch task based on the branch performance parameters during the simulation process. Based on the subsequent weight coefficients of each branch task of each simulation model, it selects the final simulation simulation path of each simulation model and uses the final state of the path as the initial state snapshot of the next simulation cycle to close the loop iterative simulation process. The resource control module collects and stores simulation process data in real time to support performance evaluation and dynamically optimizes resource allocation strategies based on task load parameters.

[0010] This application also provides a distributed intelligent simulation method for multi-source data fusion and dynamic scheduling. The method divides the military scenario of the distributed simulation simulation into task phases, sets decision points within each task phase based on tactical rules, and generates branch tasks based on different decisions at each decision point. It performs initial model classification based on the quality attribute parameters of each simulation model. The simulation engine of each simulation model advances the simulation simulation from the initial state snapshot. When a decision point is reached, multiple branch tasks are generated based on the current snapshot state, and initial weight coefficients are assigned to each branch task based on historical performance parameters and model type. Subsequently, each branch task is distributed to computing nodes for parallel simulation. Model classification switching is dynamically triggered in real time based on the driving events of each simulation model. The weight coefficients of each branch task are updated based on the branch performance parameters during the simulation process. The final simulation path of each simulation model is selected based on the subsequent weight coefficients of each branch task, and the final state of this path is used as the initial state snapshot for the next simulation cycle in a closed-loop iterative simulation process. Simulation process data is collected and stored in real time to support performance evaluation, and resource allocation strategies are dynamically optimized based on task load parameters.

[0011] One or more technical solutions provided in the embodiments of this application have at least the following technical effects or advantages:

[0012] 1. By generating branch tasks through structured task phases and tactical rule setting decision points, and performing initial classification based on model quality attributes, a hierarchical and rule-driven inference framework is constructed. Adaptive optimization and efficient resource utilization are achieved through dynamic weight allocation based on historical performance and real-time status, parallel inference, and event-driven model switching mechanisms. Autonomous optimization of the inference path and continuous improvement of result reliability are achieved through continuous weight updates of branch performance parameters, backtracking to select the optimal path, and closed-loop iteration of state snapshots. Finally, through full-cycle data acquisition and load-aware dynamic resource scheduling, the system's performance is assessable, resource allocation is elastically scalable, and the overall performance and decision support capabilities of large-scale distributed simulation are significantly enhanced.

[0013] 2. By partitioning simulation state data according to functional modules and allocating it to independent nodes, and combining it with a consistent hashing ring to achieve rapid data location and efficient storage, data migration and access latency are significantly reduced. By using an incremental snapshot mechanism to store and transmit only state difference data, network bandwidth and storage resource consumption are greatly reduced. By merging incremental data and baseline snapshots on independent nodes to reconstruct the complete state, and using shared memory segments to achieve zero-copy data sharing between multiple processes, millisecond-level fast cloning and efficient parallel simulation of branch tasks are achieved. Ultimately, this ensures the real-time performance, scalability, and resource utilization efficiency of the system in high-concurrency, multi-branch scenarios.

[0014] 3. By establishing hierarchical classification rules based on the average decision point change time and the number of decision points and branch tasks, the simulation model and computational granularity are accurately matched. By dynamically triggering classification switching through real-time monitoring of threat level, node load, and model error rate, and combining state preservation and Kalman filter compensation to achieve seamless state transition, the model granularity is ensured to adaptively match the battlefield situation and system load. Through unified coordination of the main thread, elastic allocation of the thread pool, and priority-based event scheduling, the efficient utilization of simulation resources, intelligent balancing of system load, and significant improvement in real-time performance and accuracy in complex military simulation environments are achieved.

[0015] 4. By combining model types with historical performance parameters, and utilizing multi-level predefined mapping tables to scientifically calculate the initial weight coefficients of each branch task, a reasonable initial resource allocation basis is established for parallel simulation. Then, during the simulation process, based on performance parameters such as branch probability values, probability variance, and resource consumption values, and combined with multiple sets of branch probability adjustment mapping tables, adjustment factors are dynamically generated, and weights are updated and normalized in a closed loop. This achieves adaptive optimization of simulation resources according to the battlefield situation and simulation performance. Furthermore, the system achieves precise focus on high-value, high-stability branches, significantly improving the efficiency, reliability, and economical use of resources in distributed simulation. Attached Figure Description

[0016] Figure 1 A schematic diagram of the structure of a distributed intelligent inference system for multi-source data fusion and dynamic scheduling provided in an embodiment of this application;

[0017] Figure 2 A schematic diagram of parallel and cloned multi-simulation processes in a distributed intelligent inference system for multi-source data fusion and dynamic scheduling provided in an embodiment of this application;

[0018] Figure 3 The directed acyclic graph of the deduction branches of the distributed intelligent deduction system for multi-source data fusion and dynamic scheduling provided in the embodiments of this application;

[0019] Figure 4 A thread model architecture diagram of a distributed intelligent inference system for multi-source data fusion and dynamic scheduling provided in an embodiment of this application;

[0020] Figure 5 An event processing flowchart of a distributed intelligent inference system for multi-source data fusion and dynamic scheduling provided in an embodiment of this application;

[0021] Figure 6 A flowchart of a distributed intelligent inference method for multi-source data fusion and dynamic scheduling provided in an embodiment of this application. Detailed Implementation

[0022] This application provides a distributed intelligent simulation system that integrates multi-source data fusion and dynamic scheduling. This solves the problem that in large-scale distributed combat simulations, the commonly used centralized deep reinforcement learning architecture with a single server as the central node forces all simulation nodes to engage in continuous, high-frequency data interaction with the central server, resulting in significant communication pressure. This leads to the central node becoming a system performance bottleneck, severely limiting the scalability of the simulation and the efficiency of real-time decision-making. The overall approach is as follows:

[0023] In distributed simulation and war game scenarios, the mission is first divided into phases based on the operational process, and key decision points are set within each phase based on tactical rules. For each decision point, multiple branch tasks are generated based on the current tactical situation, and the model is initially classified based on the quality attribute parameters of the simulation model. During the simulation, the simulation engine starts running from the initial state snapshot. When it advances to a decision point, it dynamically generates multiple branch tasks based on the current state snapshot, and assigns initial weight coefficients to each branch based on historical performance parameters and model type. Subsequently, the tasks are distributed to different computing nodes for parallel execution. During the process, the model type is switched in real time based on the driving events generated by the simulation model. The system dynamically updates the weight coefficients based on the performance parameters of each branch collected during the simulation, and selects the final simulation path of each model based on the weights, using its final state as the initial state snapshot for the next simulation cycle, thus forming a closed-loop iteration. At the same time, the system collects and stores data from the entire simulation process in real time to support performance evaluation, and dynamically optimizes resource allocation strategies based on task load parameters to ensure efficient system operation.

[0024] To better understand the above technical solutions, the following will provide a detailed explanation of the technical solutions in conjunction with the accompanying drawings and specific implementation methods.

[0025] like Figure 1The diagram shows the structure of a distributed intelligent simulation system for multi-source data fusion and dynamic scheduling provided in this application embodiment. This system includes: a simulation preprocessing module, a simulation management module, a simulation optimization module, and a resource control module. The simulation preprocessing module divides the military scenario of the distributed simulation simulation into task phases, sets decision points within each task phase based on tactical rules, generates branch tasks based on different decisions at each decision point, and performs initial model classification based on the quality attribute parameters of each simulation model. The task phases include an early warning phase, a mid-course interception phase, and a terminal defense phase. The simulation management module uses the simulation engine of each simulation model to advance the simulation simulation from the initial state snapshot. When a decision point is reached, multiple branch tasks are generated based on the current snapshot state, and each branch task is assigned a specific configuration based on historical performance parameters and model type. The system assigns initial weight coefficients to tasks, then distributes each branch task to computing nodes for parallel simulation. Model classification switching is dynamically triggered in real-time based on driving events of each simulation model. The simulation optimization module updates the weight coefficients of each branch task based on its branch performance parameters during the simulation process. It then selects the final simulation path for each simulation model based on the subsequent weight coefficients of each branch task, using the final state of that path as the initial state snapshot for the next simulation cycle, thus creating a closed-loop iterative simulation process. The simulation engine is the core computing component that loads the current simulation cycle's state snapshot into the simulation model and drives the interaction of simulation simulation events. The resource control module, as a core component of the distributed intelligent simulation system, is responsible for real-time collection and storage of simulation process data to support system performance evaluation. It also dynamically optimizes resource allocation strategies based on task load parameters, ultimately achieving seamless system expansion from a single node to tens of thousands of nodes.

[0026] In terms of performance evaluation, the system has constructed a complete closed-loop evaluation process. This process first establishes data acquisition standards, clarifying the types, frequencies, and accuracy of various data collection methods based on the specific objectives and application scenarios of military simulation. This covers operational unit status, location, movement trajectories, and environmental factors such as terrain and weather. Simultaneously, the system establishes a unified standard output file format to ensure that heterogeneous data from different sources can be consistently parsed and processed. The next stage is algorithm mapping, where appropriate calculation algorithms are selected based on evaluation indicators and data characteristics. The mapping relationship between indicators and data is configured visually using Luna scripts, and algorithm performance is continuously optimized based on feedback from the simulation engine. The dynamic data recording unit is responsible for capturing key data during the simulation process in real time, including operational unit status and interaction information. It has single-sample data processing capabilities, enabling preliminary processing, analysis, and storage of raw process data, and supports cross-batch aggregation and statistics of multiple sample data. Finally, through efficient file storage and management technologies, persistent data storage and rapid retrieval are achieved, providing a solid data foundation for evaluation and analysis.

[0027] In terms of dynamic resource allocation, the system uses task load parameters as the core to achieve elastic resource orchestration. Task load parameters are a set of multi-dimensional indicators, including task computational load (the demand for computing resources such as CPUs or GPUs), data throughput (the total amount and rate of input and output data during task execution, which directly affects the allocation of network bandwidth and storage input / output resources), task priority and its dependencies (used to determine scheduling order and resource preemption strategies), and task deadlines (hard or soft time constraints, related to the timeliness and real-time guarantee of resource allocation). Based on these parameters, the system models the resource allocation problem as a multi-objective constrained optimization model. The objective function includes resource costs (usually quantified based on cloud service billing models), task end-to-end latency (the total time from task submission to result return), and system energy consumption (such as server power consumption). Each objective is flexibly balanced through dynamic weighting coefficients. Constraints cover the latest task completion time, upper limits for various resource capacities such as memory, video memory, and network bandwidth, and service quality requirements such as response time thresholds and availability levels. The optimization process employs an improved third-generation non-dominated sorting genetic algorithm to solve for the Pareto optimal solution set, characterizing the trade-off between resource efficiency, timeliness, and cost. Then, a sorting method approximating the ideal solution is used to select the comprehensively optimal scheduling scheme from the solution set. This strategy can significantly improve resource utilization, reduce overall operating costs, and support large-scale system expansion while meeting the real-time and reliability requirements of the simulation task.

[0028] In this embodiment, the present invention constructs a distributed simulation and deduction mechanism, employing a parallel processing framework and a high-speed simulation engine to decouple the functions of multi-task parallel execution and task distribution from the simulation and deduction process. This decoupling design separates task scheduling from simulation execution, significantly improving the system's modularity and functional independence. It not only enhances the system's reusability and scalability under different scenarios and scales but also effectively ensures the high-speed operation and real-time response capability of the simulation, thus supporting the high-frequency, low-latency simulation requirements for various combat schemes and plans. Regarding the construction of the evaluation system, the system introduces a dynamic indicator configuration mechanism based on Luna scripts. Users can flexibly define indicator calculation logic and aggregation rules by writing scripts, effectively adapting to evaluation needs under different tactical backgrounds. The system relies on a parallel data acquisition framework to acquire real-time status, event, and interaction data generated by the simulation engine and automatically completes indicator calculation and real-time generation of a multi-dimensional evaluation system according to pre-defined script rules. This mechanism not only enables comprehensive, efficient, and reliable quantitative evaluation of operational plans, tactics, and intelligent decision-making models, but also supports dynamic adjustment of evaluation priorities during the simulation process, thus providing a complete and reliable technical foundation for in-depth analysis of simulation results, strategy iteration, and decision optimization.

[0029] Furthermore, the quality attribute parameters include the average decision point change time, the number of decision points, and the number of branch tasks. The steps for initial model classification based on the quality attribute parameters of each simulation model include: comparing the average decision point change time of each simulation model with the first decision point change time and the second decision point change time, respectively. If the average decision point change time of a simulation model is lower than the first decision point change time, then the model is classified in an auxiliary manner based on the number of decision points and the number of branch tasks of the simulation model: if the number of decision points of the simulation model is lower than the decision point number threshold and the number of branch tasks of the simulation model is lower than the task number threshold, then the simulation model is marked as a unit-level model; otherwise, the simulation model is marked as a component-level model. If the average decision point change time of a simulation model is not lower than the first decision point change time but lower than the second decision point change time, then the simulation model is marked as a component-level model. If the average decision point change time of a simulation model is not lower than the second decision point change time, then the simulation model is marked as a group-level model.

[0030] In this embodiment, scientifically classifying simulation models based on quality attribute parameters can significantly improve the operational efficiency and resource adaptation accuracy of distributed simulation inference systems. This classification method relies on three core parameters: average decision point change time (the average time difference between adjacent decision points in the historical simulation cycle); the number of decision points (the total number of decision points set in the inference preprocessing module); and the number of branch tasks (the total number of branch tasks set in the inference preprocessing module). The specific classification process is as follows: First, the average decision point change time of each model is compared with preset first and second decision point change time thresholds. If the average decision point change time of a model is lower than the first threshold, further auxiliary judgment is made based on its number of decision points and the number of branch tasks. When both of these quantities are lower than the corresponding thresholds, the model is marked as a unit-level model, suitable for high-precision, small-time-step micro-scenarios such as radar signal processing units or ballistic fine-tuning algorithms. If either quantity exceeds the threshold, it is marked as a component-level model. If the average decision point change time falls between the first and second thresholds, the model is directly classified as component-level. These models typically describe independent functional modules such as radar, missile seekers, or communication components, with moderate time steps and appropriate computational granularity. If the average decision point change time is not lower than the second threshold, the model is labeled as group-level, representing a complete combat unit such as an anti-missile system, responsible for global state management and macro-level resource scheduling, and with a larger time step. The advantage of this hierarchical mechanism lies in automatically identifying the simulation granularity and computational characteristics of the model through quantitative indicators, providing a crucial basis for system resource allocation and scheduling strategies. By dividing the model into three levels—unit-level, component-level, and group-level—the system can achieve precise matching and dynamic optimization of computational resources, ensuring the real-time requirements of high-precision models while avoiding excessive fine-grained computational power consumption by macro-level models, thereby comprehensively improving the efficiency, scalability, and overall simulation confidence of large-scale distributed simulations.

[0031] Furthermore, the steps for advancing the simulation deduction from the initial state snapshot using the simulation engines of each simulation model include: partitioning the simulation state data contained in the snapshot point according to functional modules such as satellite data, radar status, and resource inventory; and using a task scheduler such as K8s (Kubernetes, a container orchestration system) to allocate each partition to the corresponding independent nodes in the computing resource cluster, reducing the amount of data migration during cloning. The computing resource cluster is a collection of resources composed of virtual machines that provide distributed parallel computing capabilities; and constructing a consistent hash ring for the computing resource cluster: calculating the hash value of each independent node identifier and mapping it to the hash ring to ensure that cloning requests can quickly locate the data source, reducing the number of network hops. For any state partition data, its snapshot point information hash value is calculated and mapped to the same consistent hash ring. The system then locates the independent node corresponding to the first node hash value along the ring clockwise, designating this node as the storage location for the partition data to achieve fast data addressing and access. When the simulation engine reaches a decision point, branch cloning is triggered, initiating an incremental snapshot mechanism. The system calculates the differences between the current simulation state and the previous baseline snapshot, compresses the data using efficient serialization protocols such as Protobuf (Protocol Buffers), reduces network transmission latency during cloning, and adds the incremental snapshot and corresponding branch task metadata to the distributed task queue, thereby reducing storage pressure. For example, snapshot 05 only stores data that differs from snapshot 04, such as newly added target trajectories or resource consumption. At snapshot 03, the system clones the current snapshot state to generate branches 4 and 5. Through pre-allocated memory pools and shared storage such as Redis clusters, snapshot cloning time can be reduced to milliseconds. Furthermore, metadata such as decision point identifiers and branch weights from snapshots can be stored independently in a relational database, facilitating rapid association of paths and decision logic during subsequent result analysis. Independent nodes retrieve branch tasks and their associated incremental data from the task queue, merge the incremental data with the locally stored baseline snapshot, and reconstruct the complete state of the branch starting point. Distributed locks or version control mechanisms ensure the consistency of key states such as target location and remaining interceptor missile counts during branch simulation. Memory mapping technology configures the reconstructed complete state data into a shared memory segment accessible to multiple simulation processes within the same node, achieving zero-copy data sharing across processes and avoiding the physical copying overhead of large-scale state data. Independent simulation process instances are created based on the state stored in the shared memory segment, and each simulation process instance starts from the branch starting point and executes subsequent path simulation calculations in parallel. Memory mapping technology allows multiple processes to share the same physical memory region, avoiding data copying overhead. For example, radar data from snapshot 05 can be directly mapped to the address space of a new process, achieving second-level cloning.

[0032] like Figure 2The diagram illustrates the parallel and cloning of multiple simulation processes in a distributed intelligent simulation system with multi-source data fusion and dynamic scheduling provided in this embodiment. During the distributed simulation, the system utilizes each simulation engine to initiate the simulation from an initial state snapshot. By partitioning the simulation state data contained in the snapshot according to functional modules such as satellite data, radar status, and resource inventory, and using a task scheduler to allocate each partition to independent nodes in a computing resource cluster composed of virtual machines, the migration overhead during data cloning is effectively reduced. By constructing a consistent hash ring and mapping the hash values ​​of nodes and state partitions onto the ring, rapid data addressing and location are achieved, significantly reducing the number of network jumps. When the simulation reaches a decision point, branch cloning is triggered. The system initiates an incremental snapshot mechanism, calculates the difference data between the current state and the previous baseline snapshot, and uses an efficient serialization protocol for compression processing, storing only the changed parts, such as newly added target trajectories or resource consumption, significantly reducing network transmission latency and storage pressure. The difference data and branch task metadata are then added to the distributed task queue.

[0033] The core of branch parallelism lies in decoupling the multi-path simulation tasks at decision points into independent processes. For example, the maneuver response decision point in the mid-course interception phase may generate three branches: maintain the current intercept trajectory, switch to the backup seeker, or release decoy flares. Each branch requires independent computing resources. The task distribution layer dynamically allocates branch tasks to available nodes through message queues and can prioritize scheduling high-priority branches based on indicators such as threat level. After independent nodes obtain tasks and incremental data from the queue, they merge them with the local baseline snapshot to reconstruct the complete state and use distributed locks or version control mechanisms to ensure the consistency of critical states such as target location and the number of remaining interceptor missiles. Through memory mapping technology, the reconstructed state is configured as a shared memory segment accessible to multiple simulation processes, achieving zero-copy data sharing across processes, avoiding large-scale physical copying of data, and then creating independent simulation process instances based on the shared state to execute subsequent simulations in parallel. This invention effectively supports millisecond-level snapshot cloning and second-level process creation through fine-grained state management, efficient data compression and addressing, dynamic resource allocation, and consistency control, significantly improving system parallel efficiency and resource utilization, and providing high real-time and high reliability support for large-scale complex military simulations.

[0034] The core of distributed simulation lies in transforming business logic into a process model that operates collaboratively across multiple nodes, thereby achieving efficient and scalable simulation. Taking the mid-course interception phase as an example, the system decomposes the simulation task into multiple specialized processes: the perception layer process is responsible for radar signal processing and target trajectory prediction based on GPU (Graphics Processing Unit) acceleration; the decision layer process analyzes the battlefield situation in real time, triggers key decision points, and generates multi-branch task queues; the execution layer process is responsible for simulating physical actions such as controlling the launch of interceptor missiles and adjusting seeker parameters. Efficient data interaction between processes is achieved through lightweight communication protocols such as gRPC (gRPC Remote Procedure Calls). For instance, the perception layer process pushes target coordinates to the decision layer in real time, the decision layer generates branch tasks based on the rule engine, and the execution layer process consumes the queues and feeds back the simulation results.

[0035] like Figure 3 The diagram shows a directed acyclic graph (DAG) of the deduction branches in the distributed intelligent deduction system for multi-source data fusion and dynamic scheduling provided in this application embodiment. The deduction process can be structurally abstracted as a DAG, where nodes represent system snapshot states and edges represent decision branches. For example, reaching snapshot 02 from snapshot 01 via branch 2, or reaching snapshot 03 via branch 3, constitutes different deduction paths. When a branch deduction fails, the system can backtrack along the DAG to the nearest decision point to regenerate the branch, ensuring the fault tolerance and continuity of the deduction. The branch cloning process relies on several key data logics: causal consistency ensures that the cloned branch fully inherits all causal dependencies of the parent snapshot. For example, if the radar lock-on state of snapshot 05 depends on the target guidance result of snapshot 04, this dependency must be maintained during branch cloning; the version tracing mechanism requires each branch to record a complete version chain, such as snapshot 01 extending through branch 2 to snapshot 02 and then through branch 4, so that the entire decision path can be clearly traced during result analysis; the conflict resolution mechanism handles the competition problem caused by multiple branches modifying the same state through optimistic locking or transaction mechanisms. For example, when multiple branches attempt to occupy the same radar resource simultaneously, the system can coordinate resource allocation and maintain state consistency. This architecture, through multi-process collaboration, a directed acyclic graph organization, and strict data logic management, significantly improves the reliability, backtrackability, and parallel efficiency of distributed simulation and inference, providing a solid technical foundation for intelligent inference in complex military scenarios.

[0036] Furthermore, historical performance parameters include simulation success rate, target achievement time, and resource consumption rate. The steps for assigning initial weight coefficients to each branch task based on historical performance parameters and model type include: inputting the model type into a predefined model base weight mapping table to obtain the corresponding model base weights; querying the corresponding predefined mapping table according to the historical performance parameters of each branch task to obtain weight adjustment values: inputting the simulation success rate into the success rate adjustment coefficient mapping table to obtain the corresponding weight adjustment coefficients; inputting the average target achievement time into a predefined time efficiency adjustment value mapping table to obtain the corresponding time efficiency adjustment value; inputting the average resource consumption rate into a predefined resource efficiency adjustment value mapping table to obtain the corresponding resource efficiency adjustment value; coupling the time efficiency adjustment value and resource efficiency adjustment value corresponding to each branch task, and then multiplying the coupling result with the weight adjustment coefficients to obtain the weight adjustment value corresponding to each branch task; processing the model base weights based on the weight adjustment values ​​corresponding to each branch task to obtain the initial weight coefficients for the weight adjustment values ​​corresponding to each branch task.

[0037] In this embodiment, the mapping table establishes a deterministic correspondence between one input value and another output value. It transforms patterns summarized from historical data into a series of precise, executable mathematical relationships, thereby driving the entire simulation system to perform automated and intelligent resource allocation and path selection. The pre-defined mapping tables in this invention include success rate adjustment coefficient mapping tables, time efficiency adjustment value mapping tables, and resource efficiency adjustment value mapping tables. Assigning initial weight coefficients to each branch task based on historical performance parameters and model type significantly improves the scientific rigor and decision-making efficiency of the simulation system. By quantitatively evaluating the historical performance and inherent characteristics of the model, differentiated initial weights are assigned to multiple branch tasks in parallel simulations, thereby achieving a more accurate and reasonable initial configuration of simulation resources. Mapping model type to basic weights reflects its inherent design capabilities. Introducing multiple historical performance parameters, including simulation success rate, target achievement time, and resource consumption rate, as dynamic adjustment criteria effectively integrates the model's static attributes and dynamic performance. By querying a predefined mapping table and coupling the processing time efficiency and resource efficiency adjustment values, a comprehensive weight adjustment value is finally generated. This system not only takes into account the efficiency and resource economy of the deduction path, but also assigns higher priority to high-value and high-success-probability branches at the initial stage of deduction, guiding the system to optimize resource allocation, avoid waste of computing power, and thus comprehensively improve the convergence speed and reliability of the deduction process.

[0038] like Figure 4The diagram shows the thread model architecture of the distributed intelligent inference system for multi-source data fusion and dynamic scheduling provided in this application embodiment. The steps for dynamically triggering model classification switching based on driving events of each simulation model in real time include: real-time monitoring of driving events of each simulation model; triggering model classification switching when the threat level of a driving event exceeds a predefined level; triggering model classification switching when the node load of a driving event exceeds a predefined load threshold; and triggering model classification switching when the error rate of the simulation model obtained based on Monte Carlo inference exceeds a preset model error rate. The steps for model classification switching include: saving the current simulation model type. The system outputs the state and uses it as the initial input value for the switched model type. A Kalman filter is used to compensate for state transitions during model classification switching, ensuring seamless transfer of state data. The main thread manages the global state updates and event distribution of the entity-level model, and uses synchronization mechanisms such as spinlocks to control the execution of each sub-thread, ensuring state consistency. A thread pool dynamically allocates computational tasks corresponding to component-level and sub-component-level models, enabling elastic management of computational resources. The event queue is dynamically scheduled based on the priority parameters of each driving event, including probe events, engagement events, and interference events.

[0039] Specifically, priority parameters include threat level, event completion time, and allowable delay time. The steps for dynamically scheduling the event queue based on the priority parameters and event type of each driving event include: retrieving events from the event queue; calculating the priority index of each driving event based on the priority parameters; if the priority index of a driving event exceeds a predefined priority index threshold, the corresponding event is assigned to an exclusive thread for processing; if the priority index of a driving event does not exceed the predefined priority index threshold, dynamic allocation processing is performed based on the driving event type. Figure 5The diagram shows the event processing flowchart of the distributed intelligent simulation system for multi-source data fusion and dynamic scheduling provided in this application embodiment. If the driving event type is a detection event, it is assigned to the detection thread, i.e., the target tracking thread; if the driving event type is an engagement event, an interception command is initiated and assigned to the engagement thread, i.e., the ballistic calculation and damage assessment thread; if the driving event type is a jamming event, it is assigned to the jamming thread, i.e., the thread that performs electronic warfare simulation and communication parameter adjustment; the remaining tasks are assigned to sub-threads that handle the corresponding event types, and corresponding instructions are issued to the thread pool through the processor; to cope with high... In concurrent scenarios, when the number of pending events within a computing node exceeds a preset threshold, the system automatically initiates performance optimization operations. These operations primarily involve two key technologies: Within the computing node, lock-free data structures based on CAS (Compare-And-Swap) instructions, such as circular queues and hash tables, are employed. Hardware-level atomic operations replace traditional lock mechanisms, ensuring thread safety and low-latency response during multi-threaded concurrent access, completely avoiding blocking and performance degradation caused by lock contention. Between distributed computing nodes, RDMA (Remote Direct Memory Access) network technology is enabled, bypassing the operating system kernel and protocol stack to achieve remote direct memory access between nodes. This reduces network communication latency to the microsecond level and significantly reduces CPU (Central Processing Unit) overhead, thereby ensuring the ultimate efficiency of cross-node data exchange. These two technologies work synergistically to improve the system's throughput, real-time performance, and overall stability under high load pressure.

[0040] The target tracking thread is dedicated to handling detection-driven events. Its core function is to continuously receive, fuse, and process raw observation data from various sensors. Through algorithms such as Kalman filtering and multi-source information fusion, it achieves accurate estimation and prediction of the enemy target's trajectory, speed, and azimuth, and continuously updates target trajectory information to provide real-time and accurate intelligence support for subsequent firepower decisions. The ballistic calculation and damage assessment thread is responsible for handling engagement-related events. Upon receiving an interception command, this thread calculates the most accurate trajectory in real time based on the current target's predicted trajectory, the interceptor missile's dynamics model, and environmental parameters. The system optimizes ballistic interception and generates guidance commands. After simulated engagement, this thread calculates and evaluates the damage level and effect on the target using a physical effect model based on parameters such as warhead power, fuse detonation conditions, and miss distance. The thread responsible for electronic warfare simulation and communication parameter adjustment focuses on handling interference events. It calculates the interference effect of friendly electronic warfare equipment on specific sensors or communication links by simulating electromagnetic wave propagation, signal modulation, and interference patterns. It also dynamically adjusts parameters such as frequency hopping mode and transmission power of friendly communication networks to counter enemy electromagnetic interference and ensure uninterrupted communication. These threads are typically implemented based on a multi-threaded programming model. After the main control scheduling logic determines the event type, it allocates specific event objects to pre-created, continuously running dedicated threads via processor instructions. These threads reside in memory, receive tasks through thread-safe message queues, continuously retrieve tasks from the queue during the event loop, execute their core algorithms, and return the results to shared memory or message queues for consumption by other threads or nodes. This achieves efficient and professional processing of high-concurrency, high-real-time simulation events.

[0041] The priority index of each driving event is obtained as follows: the threat level, event completion time and preset critical allowable delay time are respectively compared with the preset critical threat level, critical event completion time and allowable delay time to calculate the proportion convergence; the proportion convergence calculation results are weighted by the preset threat level weight ratio, event completion time weight ratio and allowable delay time weight ratio respectively; the weighted results are coupled to obtain the priority index of each driving event.

[0042] In this embodiment, the threat level is typically obtained through real-time analysis of battlefield situational data, such as calculations based on factors like target type, speed, heading, and the value of friendly assets using a rule engine or predictive model. The event completion time is estimated based on historical simulation data or the standard operating time for this type of task under a specific scenario. The allowable delay time is a hard constraint pre-set by operational doctrine, equipment performance limits, and specific tactical rules, representing the maximum delay tolerance that this type of event can handle. This invention employs a scientific multi-parameter weighted algorithm to quantify the urgency of events: by calculating the approximate convergence of the threat level, event completion time, and allowable delay time with their corresponding critical values, and applying preset weight ratios for weighted coupling, a comprehensive priority index is generated, thereby achieving an accurate assessment of the importance of driving events. Based on the dual criteria of index and event type, the system can perform intelligent scheduling: allocating extremely high-priority events to dedicated threads to ensure immediate response to critical tasks; and precisely distributing routine events to dedicated thread pools according to their type. For example, detection events are handled by the target tracking thread, engagement events trigger interception commands and are assigned to the ballistic calculation and damage assessment thread, and interference events are handled by the electronic warfare simulation thread, ensuring that various events are processed professionally and in parallel. Furthermore, the system automatically activates high-performance optimization strategies under high load pressure, including using lock-free data structures within nodes to reduce synchronization overhead and enabling remote direct data access between nodes to reduce communication latency, thereby effectively maintaining the system's stability and low-latency characteristics under extreme loads. Overall, this mechanism, through quantitative evaluation, classified scheduling, and elastic optimization, achieves efficient, reliable, and real-time processing of complex simulation events, comprehensively enhancing the system's overall performance and mission assurance capabilities.

[0043] Furthermore, the branch performance parameters include branch probability value, resource consumption value, and probability variance. The steps for updating the weight coefficients of each branch task based on the branch performance parameters of each branch task during the simulation include: recording the average probability value obtained by the Monte Carlo method through random sampling to simulate battlefield uncertainty as the branch probability value; recording the quantified value obtained by the independent node executing the branch task during the simulation based on the comprehensive evaluation of hardware performance parameters as the resource consumption value; recording the variance of the branch probability values ​​obtained by the same branch during the simulation as the probability variance; obtaining the initial weight coefficients of each branch task based on the model type of each simulation model; if the branch probability value of a certain branch task in a certain simulation model does not exceed the preset branch probability threshold, the difference between the branch probability threshold and the branch probability value is recorded as the deviation probability value, and the deviation probability value is input into the preset branch probability adjustment first mapping table to obtain the corresponding branch probability adjustment factor to penalize low probability branches; if the branch probability value of a certain branch task in a certain simulation model exceeds the preset branch probability threshold, and the probability variance exceeds the preset probability variance threshold, it indicates that the result is not stable enough. If the difference between the probability variance and the probability variance threshold is recorded as the deviation probability variance, and the deviation probability variance is input into the preset second mapping table of branch probability adjustment to obtain the corresponding branch probability adjustment factor for inhibitory adjustment; if the branch probability value of a certain branch task in a simulation model exceeds the preset branch probability threshold, but the probability variance does not exceed the preset probability variance threshold, then it is determined whether the resource consumption value exceeds the preset resource consumption threshold. If so, the difference between the resource consumption value and the resource consumption threshold is recorded as the deviation resource consumption value, and the deviation resource consumption value is input into the preset third mapping table of branch probability adjustment to obtain the corresponding branch probability adjustment factor for economic adjustment; otherwise, a weighted coupling process is performed based on the branch probability value, resource consumption value, probability variance and its corresponding preset weight factor to generate the branch probability adjustment factor; the initial weight coefficient of each branch task is multiplied by its corresponding branch probability adjustment factor to obtain the adjusted weight coefficient of each branch task, and then normalization is performed based on the adjusted weight coefficient of each branch task to obtain the subsequent weight coefficient of each branch task in each simulation model.

[0044] The hardware performance parameters include CPU usage time, peak memory usage, and total network I / O (total network input / output). CPU usage time can be obtained by querying the process-level performance counters provided by the operating system kernel, which records the actual processor time consumed by a specific simulation task on the computing node. Peak memory usage is obtained by monitoring the maximum physical memory usage of the process throughout the simulation cycle, usually collected and reported in real time by the runtime environment's memory management subsystem or an external monitoring agent. Total network I / O is obtained by reading the statistics of the operating system's network stack, which cumulatively counts the number of bytes of network packets sent and received during the execution of the simulation task. The resource consumption values ​​are obtained as follows: CPU usage time, peak memory usage, and total network I / O are compared with preset critical CPU usage time, critical peak memory usage, and critical total network I / O, respectively, to calculate their approximation. The approximation results are then weighted using preset weight ratios for CPU usage time, peak memory usage, and total network I / O. Finally, the weighted results are coupled, i.e., summed, to obtain the resource consumption. This invention comprehensively characterizes the predictability, economy, and stability of branching tasks by introducing three key performance parameters: branch probability value, resource consumption value, and probability variance.

[0045] In this embodiment, the invention first determines the branch probability value based on the average probability value obtained from Monte Carlo random sampling, obtains the resource consumption value based on a comprehensive evaluation of hardware performance indicators, and calculates the probability variance through historical inference data to measure the volatility of the results. Through multi-level condition judgment and mapping table lookup, accurate and automated evaluation and weight allocation of branch tasks are achieved, ensuring that system resources are concentrated on high-value, highly stable, and economically feasible inference paths, thereby comprehensively improving inference efficiency and the reliability of results.

[0046] Furthermore, the steps for selecting the final simulation inference path of each simulation model based on the subsequent weight coefficients of each branch task of each simulation model include: starting from the initial state of the simulation model, traversing each decision point sequentially along the inference branch architecture, and performing an adaptive backtracking operation for the current decision point: Step 1: Obtain the subsequent weight coefficients of each branch task under this decision point; Step 2: If the subsequent weight coefficient of any branch task exceeds the preset branch weight coefficient threshold, then sort each branch task in descending order according to the corresponding subsequent weight coefficient, select the branch task at the top of the sorted order to be included in the final inference path, and advance to the next branch at the end of the branch. Step 3: If the subsequent weight coefficients of all branch tasks do not exceed the branch weight coefficient threshold, the current path deduction is determined to have failed, and a backtracking mechanism is executed: backtracking backward along the current deduction path to the previous decision point, and marking the selected branches in that decision point as invalid, re-executing Step 1 and Step 2, and re-selecting a path from the remaining valid branches; repeating the adaptive backtracking operation, if the simulation model successfully advances to the deduction termination state, the recorded path is taken as the final simulation deduction path of the simulation model; if backtracking to the starting state and all branches are marked as invalid, the simulation model deduction is determined to have failed.

[0047] In this embodiment, the present invention significantly improves the reliability and intelligent decision-making level of simulation path selection by introducing an adaptive backtracking strategy based on weight coefficients. Dynamic path selection is performed based on quantified weights rather than preset rules. Starting from the initial state, the system prioritizes the branch with the highest subsequent weight coefficient at each decision point, ensuring that the simulation always extends along the current optimal path. When the weights of all branches have not reached the threshold, the system automatically triggers a backtracking mechanism: returning to the previous decision point, disabling failed branches, and re-evaluating the remaining options, thereby effectively avoiding local failures and exploring potential feasible paths. This process is repeated until the termination state is successfully reached or all possibilities are exhausted, ultimately outputting a complete simulation path with high weight and high credibility, or explicitly declaring the simulation a failure. This not only improves the success rate and result quality of simulations under complex scenarios but also enhances the system's adaptability to uncertainties and abnormal situations, providing more scientific and robust decision support for military simulations.

[0048] like Figure 6The diagram shows a flowchart of a distributed intelligent simulation method for multi-source data fusion and dynamic scheduling provided in this application embodiment. The method involves dividing the military scenario of the distributed simulation into task phases, setting decision points within each task phase based on tactical rules, generating branch tasks based on different decisions at each decision point, and initially classifying models based on their quality attribute parameters. The simulation engine of each model is used to advance the simulation from the initial state snapshot. When a decision point is reached, multiple branch tasks are generated based on the current snapshot state, and initial weight coefficients are assigned to each branch task based on historical performance parameters and model type. Subsequently, each branch task is distributed to computing nodes for parallel simulation, and model classification switching is dynamically triggered in real time based on the driving events of each simulation model. The weight coefficients of each branch task are updated based on the branch performance parameters during the simulation process. The final simulation path of each simulation model is selected based on the subsequent weight coefficients of each branch task, and the final state of this path is used as the initial state snapshot for the next simulation cycle in a closed-loop iterative simulation process. Simulation process data is collected and stored in real time to support performance evaluation, and resource allocation strategies are dynamically optimized based on task load parameters.

[0049] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0050] This invention is described with reference to flowchart illustrations and / or block diagrams of systems, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0051] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0052] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0053] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0054] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A distributed intelligent inference system for multi-source data fusion and dynamic scheduling, characterized in that, It includes a simulation preprocessing module, a simulation management module, a simulation optimization module, and a resource control module; The simulation preprocessing module is used to divide the military scenario of the distributed simulation simulation into task phases, set decision points in each task phase based on tactical rules, generate branch tasks according to different decisions at each decision point, and perform initial model classification based on the quality attribute parameters of each simulation model. The task phases include early warning phase, mid-course interception phase and terminal defense phase. The simulation management module is used to advance the simulation simulation from the initial state snapshot using the simulation engine of each simulation model. When the decision point is reached, multiple branch tasks are generated according to the current snapshot state, and initial weight coefficients are assigned to each branch task based on historical performance parameters and model type. Then, each branch task is distributed to computing nodes for parallel simulation and the model classification switching is dynamically triggered in real time based on the driving events of each simulation model. The simulation optimization module is used to update the weight coefficients of each branch task based on the branch efficiency parameters of each branch task during the simulation process, and to select the final simulation simulation path of each simulation model based on the subsequent weight coefficients of each branch task of each simulation model. The final state of the path is used as the initial state snapshot of the next simulation cycle for the closed-loop iterative simulation process. The branch efficiency parameters include branch probability value, resource consumption value and probability variance. The step of updating the weight coefficients of each branch task based on the branch performance parameters of each branch task during the simulation includes: The initial weight coefficients for each branch task are obtained based on the model type of each simulation model. If the branch probability value of a certain branch task in a certain simulation model does not exceed the preset branch probability threshold, the difference between the branch probability threshold and the branch probability value is recorded as the deviation probability value. The deviation probability value is input into the preset branch probability adjustment first mapping table to obtain the corresponding branch probability adjustment factor. If the branch probability value of a certain branch task in a certain simulation model exceeds the preset branch probability threshold, and the probability variance exceeds the preset probability variance threshold, then the difference between the probability variance and the probability variance threshold is recorded as the deviation probability variance. The deviation probability variance is input into the preset branch probability adjustment second mapping table to obtain the corresponding branch probability adjustment factor. If the branch probability value of a certain branch task in a certain simulation model exceeds the preset branch probability threshold, and the probability variance does not exceed the preset probability variance threshold, then it is determined whether the resource consumption value exceeds the preset resource consumption threshold. If so, the difference between the resource consumption value and the resource consumption threshold is recorded as the deviation resource consumption value. The deviation resource consumption value is input into the preset third mapping table for branch probability adjustment to obtain the corresponding branch probability adjustment factor. Otherwise, the branch probability adjustment factor is obtained based on the branch probability value, resource consumption value, and probability variance. The adjusted weight coefficients of each branch task are obtained by multiplying the initial weight coefficients of each branch task with their corresponding branch probability adjustment factors. Then, the adjusted weight coefficients of each branch task are normalized to obtain the subsequent weight coefficients of each branch task in each simulation model. The resource control module is used to collect and store simulation process data in real time to support performance evaluation and dynamically optimize resource allocation strategies based on task load parameters.

2. The distributed intelligent inference system for multi-source data fusion and dynamic scheduling as described in claim 1, characterized in that, The quality attribute parameters include average decision point change time, number of decision points, and number of branch tasks; The steps for initial model classification based on the quality attribute parameters of each simulation model include: The average decision point change time of each simulation model is compared with the change time of the first and second decision points, respectively. If the average decision point change time of a simulation model is lower than the change time of the first decision point, then model-assisted classification is performed based on the number of decision points and the number of branch tasks of that simulation model. If the number of decision points in the simulation model is lower than the decision point number threshold and the number of branch tasks in the simulation model is lower than the task number threshold, then the simulation model is marked as a unit-level model. Otherwise, mark the simulation model as a component-level model; If the average decision point change time of a simulation model is not less than the first decision point change time, but less than the second decision point change time, then the simulation model is marked as a component-level model. If the average decision point change time of a simulation model is not less than the second decision point change time, then the simulation model is marked as a group-level model.

3. The distributed intelligent inference system for multi-source data fusion and dynamic scheduling as described in claim 1, characterized in that, The steps of using the simulation engines of each simulation model to advance the simulation deduction from the initial state snapshot include: The simulation state data contained in the snapshot point is partitioned according to the functional module, and each partition is assigned to the corresponding independent node in the computing resource cluster. The snapshot point is the time node used to record the state snapshot during the distributed simulation and simulation process. The state snapshot includes the initial state snapshot and the incremental snapshot generated at each decision point. Construct a consistent hash ring for the computing resource cluster: calculate the hash value of each independent node identifier and map it onto the hash ring. For any state partition data, calculate the hash value of its snapshot point information and map it onto the same consistent hash ring. Locate the independent node corresponding to the first node hash value clockwise along the ring and designate that node as the storage location of the partition data. When the simulation engine reaches a decision point, it initiates an incremental snapshot mechanism to calculate the difference between the current simulation state and the previous baseline snapshot. After compressing the data, it adds the corresponding branch task metadata to the distributed task queue. Independent nodes retrieve branch tasks and their associated incremental data from the task queue, merge the incremental data with the local storage baseline snapshot, and reconstruct the complete state of the branch starting point. The reconstructed complete state data is configured as a shared memory segment accessible to multiple simulation processes within the same node. Independent simulation process instances are created based on the state stored in the shared memory segment. Each simulation process instance starts from the branch origin and executes the deduction calculation of the subsequent path in parallel.

4. The distributed intelligent inference system for multi-source data fusion and dynamic scheduling as described in claim 1, characterized in that, The historical performance parameters include simulation success rate, target achievement time, and resource consumption rate. The steps for assigning initial weight coefficients to each branch task based on historical performance parameters and model type include: The basic weights of the model are obtained based on the model type; Each branch task's historical performance parameters are used to query the corresponding predefined mapping table to obtain the corresponding weight adjustment value. The weight adjustment values ​​corresponding to each branch task are coupled with the basic weights of the model to obtain the initial weight coefficients of the weight adjustment values ​​corresponding to each branch task.

5. The distributed intelligent inference system for multi-source data fusion and dynamic scheduling as described in claim 1, characterized in that, The steps for dynamically triggering model classification switching based on driving events of each simulation model in real time include: Real-time monitoring of driving events for each simulation model; When the threat level of a driving event exceeds the predefined level, the node load of the driving event exceeds the predefined load threshold, or the simulation model error rate exceeds the preset model error rate, the model classification switch is triggered. The steps for switching model classifications are as follows: Save the output state of the current simulation model type and use this output state as the initial input value of the model type after switching; The main thread manages the global state update and event dispatch functions of the entity-level model and synchronously controls the execution process of each sub-thread. Dynamically allocate computational tasks corresponding to component-level and sub-component-level models using thread pools; The event queue is dynamically scheduled based on the priority parameters and types of each driving event. The driving event types include detection events, engagement events, and interference events.

6. The distributed intelligent inference system for multi-source data fusion and dynamic scheduling as described in claim 5, characterized in that, The priority parameters include threat level, event completion time, and allowed delay time; The step of dynamically scheduling the event queue according to the priority parameters and types of each driving event includes: Retrieve events from the event queue and calculate the priority index of each driving event based on the priority parameter; If the priority index of a certain driving event exceeds the predefined priority index threshold, the corresponding event will be assigned to an exclusive thread for processing. If the priority index of a certain driving event does not exceed the predefined priority index threshold, then dynamic allocation processing is performed according to the driving event type: If the driving event type is a probe event, it is assigned to the probe thread; If the driving event type is an engagement event, then an interception instruction is initiated and assigned to the engagement thread; If the driving event type is a disturbance event, it is assigned to the disturbance thread; The remaining tasks are assigned to sub-threads that handle the corresponding event types, and the corresponding processing tasks are submitted to the thread pool through the processor. When the number of pending events within a compute node exceeds a preset threshold, performance optimization operations are initiated, including: Lock-free data structures are used within the compute nodes; Enable remote direct data access between distributed computing nodes.

7. The distributed intelligent inference system for multi-source data fusion and dynamic scheduling as described in claim 6, characterized in that, The priority index of each driving event is obtained as follows: The threat level, event completion time, and preset critical allowable delay time are respectively compared with the preset critical threat level, critical event completion time, and allowable delay time to calculate the approximate degree. The results of the proportion convergence calculation are weighted using preset threat level weight ratio, event completion time weight ratio, and allowable delay time weight ratio, respectively. The weighting results are coupled to obtain the priority index of each driving event.

8. The distributed intelligent inference system for multi-source data fusion and dynamic scheduling as described in claim 1, characterized in that, The steps for selecting the final simulation derivation path of each simulation model based on the subsequent weight coefficients of each branch task of each simulation model include: Starting from the initial state of the simulation model, the system sequentially traverses each decision point along the deduction branch architecture, performing an adaptive backtracking operation for the current decision point: Step 1: Obtain the subsequent weight coefficients of each branch task at this decision point; Step 2: If the subsequent weight coefficient of a branch task exceeds the preset branch weight coefficient threshold, sort each branch task in descending order according to the corresponding subsequent weight coefficient, select the branch task at the top of the sort order and include it in the final deduction path, and advance to the next state point at the end of the branch. Step 3: If the subsequent weight coefficients of all branch tasks do not exceed the branch weight coefficient threshold, then execute the backtracking mechanism: backtrack along the current deduction path to the previous decision point, mark the selected branches in that decision point as invalid, re-execute Step 1 and Step 2, and re-select a path from the remaining valid branches; Repeat the adaptive backtracking operation. If the simulation model successfully advances to the simulation termination state, the recorded path is taken as the final simulation path of the simulation model. If the simulation model fails to run, it is determined that the simulation model has failed if it is backtracked to the initial state and all branches are marked as invalid.

9. A distributed intelligent inference method for multi-source data fusion and dynamic scheduling, applied to the distributed intelligent inference system for multi-source data fusion and dynamic scheduling as described in any one of claims 1-8, characterized in that: The military scenario in the distributed simulation is divided into task phases. Based on tactical rules, decision points are set in each task phase, and branch tasks are generated according to the different decisions at each decision point. The models are initially classified based on the quality attribute parameters of each simulation model. The simulation engine of each simulation model is used to advance the simulation from the initial state snapshot. When the decision point is reached, multiple branch tasks are generated according to the current snapshot state. Initial weight coefficients are assigned to each branch task based on historical performance parameters and model type. Then, each branch task is distributed to computing nodes for parallel simulation. The model classification switching is dynamically triggered in real time based on the driving events of each simulation model. The weight coefficients of each branch task are updated based on the branch efficiency parameters of each branch task during the simulation process. The final simulation path of each simulation model is selected based on the subsequent weight coefficients of each branch task of each simulation model. The final state of the path is used as the initial state snapshot of the next simulation cycle for the closed-loop iterative simulation process. Real-time acquisition and storage of simulation process data to support performance evaluation, and dynamic optimization of resource allocation strategies based on task load parameters.

Images