An unmanned intelligent system virtual-real combined simulation training method and device

Abstract

This invention discloses a method and apparatus for joint virtual-real simulation training of an unmanned intelligent system. The method includes: firstly, acquiring a set of physical entity equipment configuration data, training task configuration data, and training parameter configuration data; then, performing modeling and scene generation processing on the set of physical entity equipment configuration data, the training task configuration data, and the training parameter configuration data to obtain a set of digital equipment model data and virtual training scene data; next, performing joint virtual-real training processing on the set of digital equipment model data and the virtual training scene data to obtain a set of training process data; and finally, evaluating and performing closed-loop optimization processing on the set of training process data to obtain a target training evaluation report and optimized training parameter configuration data.

Description

Technical Field

[0001] This invention relates to the field of simulation training technology for unmanned intelligent systems, and in particular to a method and apparatus for virtual-real joint simulation training of unmanned intelligent systems. Background Technology

[0002] With the widespread application of unmanned intelligent technology in military reconnaissance, logistics distribution, emergency rescue, and other fields, higher demands are placed on the operational proficiency of unmanned intelligent systems and the performance of multi-device collaborative operations. Training is a crucial link in ensuring the efficient operation of unmanned intelligent systems; however, traditional training methods have significant limitations. Pure virtual simulation training methods achieve operational training and algorithm verification by constructing virtual environments and virtual device models. Although low-cost and reusable scenarios, they lack real feedback from physical entities, such as aerodynamic drag, mechanical vibration, and sensor noise. The differences between virtual models and physical entities lead to a disconnect between training results and practical applications. Pure physical entity training directly uses real unmanned intelligent systems in actual sites. Although it can provide a realistic operational experience and physical feedback, it is limited by factors such as site, cost, and safety, and cannot simulate extreme environments and high-risk mission scenarios. The training scenarios are limited and have poor scalability.

[0003] Existing virtual-real hybrid simulation training methods mainly include semi-physical simulation methods and hardware-in-the-loop testing methods. Semi-physical simulation methods combine some physical entity modules with a virtual simulation environment, balancing physical realism and scene flexibility to a certain extent. However, they have the following shortcomings: First, the data interaction latency between the virtual environment and physical entities is high, and synchronization is poor. The lack of a quantitative synchronization error control mechanism easily leads to inconsistencies between the virtual and physical states, affecting the accuracy and reliability of simulation training. Second, scene generation relies on manual editing or preset templates, making it impossible to dynamically adapt training needs based on training objectives, operator skill levels, and historical training data. Third, the evaluation system is relatively simple, focusing only on task completion results and lacking multi-dimensional quantitative analysis of equipment operating status, operational behavior standardization, and collaborative cooperation, failing to provide accurate data support for subsequent training optimization. Fourth, the interfaces and communication protocols of physical equipment from different manufacturers are inconsistent, resulting in poor compatibility and making it difficult to achieve cross-platform, multi-model joint training. Summary of the Invention

[0004] The technical problem to be solved by this invention is to provide a virtual-real joint simulation training method and device for unmanned intelligent systems. By constructing a five-layer collaborative architecture consisting of a physical entity layer, a digital mapping layer, a virtual simulation layer, a data interaction layer, and a training control and evaluation layer, it achieves low-latency and high-precision synchronization between the virtual environment and physical entities, dynamic scene adaptive generation, multi-dimensional quantitative evaluation, and closed-loop parameter optimization, thereby improving the accuracy, personalization, and cross-platform compatibility of simulation training for unmanned intelligent systems.

[0005] To address the aforementioned technical problems, a first aspect of this invention discloses a method for joint virtual-real simulation training of an unmanned intelligent system, the method comprising: S1, obtain the physical entity device configuration data set, training task configuration data, and training parameter configuration data; S2, Modeling and scene generation processing are performed on the physical entity equipment configuration data information set, the training task configuration data information and the training parameter configuration data information to obtain the digital equipment model data information set and the virtual training scene data information; S3, perform virtual-real joint training processing on the digital equipment model data information set and the virtual training scene data information to obtain the training process data information set; S4, evaluate and perform closed-loop optimization on the training process data information set and the training parameter configuration data information to obtain target training evaluation report information and optimized training parameter configuration data information.

[0006] As an optional implementation, in the first aspect of the present invention, the step of modeling and scene generation processing of the physical entity equipment configuration data information set, the training task configuration data information, and the training parameter configuration data information to obtain a digital equipment model data information set and virtual training scene data information includes: S21, Perform digital equipment model construction processing on the physical entity equipment configuration data information set to obtain a digital equipment model data information set; the digital equipment model data information set includes several digital equipment model data information sets. S22, perform virtual training scene generation processing on the training task configuration data information and the training parameter configuration data information to obtain virtual training scene data information.

[0007] As an optional implementation, in the first aspect of the present invention, the step of performing digital equipment model construction processing on the physical entity equipment configuration data information set to obtain a digital equipment model data information set includes: S211, Perform geometric mapping processing on the device geometric parameter data information in any of the physical entity device configuration data information in the physical entity device configuration data information set to obtain the geometric mapping model data information of the physical entity device configuration data information; S212, Perform physical mapping processing on the sensor parameter data information and actuator parameter data information in the configuration data information of the physical entity device to obtain the physical mapping model data information of the configuration data information of the physical entity device. S213, Perform control mapping processing on the device type information in the configuration data information of the physical entity device to obtain the control mapping model data information of the configuration data information of the physical entity device. S214, integrate the geometric mapping model data information, the physical mapping model data information, and the control mapping model data information of the physical entity equipment configuration data information to obtain the digital equipment model data information corresponding to the physical entity equipment configuration data information.

[0008] As an optional implementation, in the first aspect of the present invention, the virtual training scene generation process of the training task configuration data information and the training parameter configuration data information to obtain virtual training scene data information includes: S221, Perform environmental parameter generation processing on the scene type information and initial difficulty level information in the training parameter configuration data information to obtain environmental parameter data information; S222, Perform task parameter generation processing on the operation level information in the training task configuration data information and the training parameter configuration data information to obtain task parameter data information; S223, perform interference parameter generation processing on the initial difficulty level information and historical training data information in the training parameter configuration data information to obtain interference parameter data information; S224, calculate and process the environmental parameter data, the task parameter data, and the interference parameter data to obtain the overall difficulty value of the scene; S225, Perform scene rationality verification processing on the comprehensive difficulty value of the scene to obtain virtual training scene data information.

[0009] As an optional implementation, in a first aspect of the present invention, the step of evaluating and performing closed-loop optimization on the training process data information set and the training parameter configuration data information to obtain target training evaluation report information and optimized training parameter configuration data information includes: S41, perform data preprocessing on the training process data information set to obtain a preprocessed training process data information set; S42, perform multi-dimensional index calculation processing on the preprocessed training process data information set to obtain a multi-dimensional evaluation index data information set; S43, perform comprehensive evaluation and calculation on the multi-dimensional evaluation index data information set and the training parameter configuration data information to obtain the target training evaluation report information; S44, perform closed-loop optimization processing on the target training evaluation report information and the training parameter configuration data information to obtain optimized training parameter configuration data information.

[0010] As an optional implementation, in the first aspect of the present invention, the step of performing multi-dimensional index calculation processing on the preprocessed training process data information set to obtain a multi-dimensional evaluation index data information set includes: S421, Process the device status time-series data information in the preprocessed training process data information set to obtain device status index data information; S422, Process the timing data of the operation instructions in the preprocessed training process data information set to obtain operation behavior index data information; S423, Process the task execution result data information in the preprocessed training process data information set to obtain task performance index data information; S424, Process the virtual-real synchronization deviation data information and the operation instruction timing data information in the preprocessed training process data information set to obtain the cooperative index data information; S425, integrate and process the equipment status index data information, the operation behavior index data information, the task effect index data information, and the cooperation index data information to obtain a multi-dimensional evaluation index data information set.

[0011] As an optional implementation, in the first aspect of the present invention, the step of comprehensively evaluating and calculating the multi-dimensional evaluation index data information set and the training parameter configuration data information to obtain target training evaluation report information includes: S431, The multi-dimensional evaluation index data information set is subjected to index normalization processing to obtain a normalized evaluation index data information set; S432, calculate and process the normalized evaluation index data information set and the evaluation index weight data information in the training parameter configuration data information to obtain the comprehensive training score value and the score data information set of each dimension. S433, integrate the comprehensive training score and the score data information set of each dimension to obtain the target training evaluation report information.

[0012] A second aspect of this invention discloses a virtual-real joint simulation training device for an unmanned intelligent system, the device comprising: The acquisition module is used to acquire physical entity device configuration data information set, training task configuration data information, and training parameter configuration data information; The modeling and scene generation module is used to model and generate scenes from the physical entity equipment configuration data information set, the training task configuration data information, and the training parameter configuration data information to obtain a digital equipment model data information set and a virtual training scene data information. The virtual-real joint training module is used to perform virtual-real joint training processing on the digital equipment model data information set and the virtual training scene data information to obtain the training process data information set. The evaluation and closed-loop optimization module is used to evaluate and perform closed-loop optimization on the training process data information set and the training parameter configuration data information to obtain target training evaluation report information and optimized training parameter configuration data information.

[0013] A third aspect of this invention discloses another virtual-real joint simulation training device for unmanned intelligent systems, the device comprising: processor; A memory containing executable program code is coupled to the processor; The processor calls the executable program code stored in the memory to execute some or all of the steps of the virtual-real joint simulation training method for unmanned intelligent systems disclosed in the first aspect of the present invention.

[0014] The fourth aspect of the present invention discloses a computer-readable storage medium storing computer instructions, which, when invoked, are used to execute some or all of the steps of the virtual-real joint simulation training method for unmanned intelligent systems disclosed in the first aspect of the present invention.

[0015] Compared with the prior art, the embodiments of the present invention have the following beneficial effects: In this embodiment of the invention, firstly, a set of physical entity equipment configuration data, training task configuration data, and training parameter configuration data are acquired; then, the set of physical entity equipment configuration data, the training task configuration data, and the training parameter configuration data are modeled and processed for scene generation to obtain a set of digital equipment model data and virtual training scene data; next, the set of digital equipment model data and the virtual training scene data are subjected to virtual-real joint training processing to obtain a set of training process data; finally, the set of training process data is evaluated and optimized in a closed loop to obtain a target training evaluation report and optimized training parameter configuration data. It can be seen that this application achieves precise three-dimensional mapping and real-time state synchronization of physical entities through digital equipment models, automatically generates virtual training scenes based on training objectives and operational levels using a dynamic scene generation engine, and dynamically adjusts training parameters using a multi-dimensional quantitative evaluation and closed-loop parameter optimization mechanism. This solves the technical problems of low virtual-real synchronization accuracy, poor scene adaptability, single evaluation dimensions, and insufficient cross-device compatibility in existing virtual-real combined simulation training, improving the accuracy, personalization, and cross-platform compatibility of unmanned intelligent system simulation training. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 This is a flowchart illustrating a virtual-real joint simulation training method for an unmanned intelligent system disclosed in an embodiment of the present invention; Figure 2 This is a schematic diagram of the five-layer collaborative technology architecture in a virtual-real joint simulation training method for an unmanned intelligent system disclosed in an embodiment of the present invention; Figure 3 This is a schematic diagram of the structure of a virtual-real joint simulation training device for an unmanned intelligent system disclosed in an embodiment of the present invention; Figure 4 This is a schematic diagram of another unmanned intelligent system virtual-real joint simulation training device disclosed in an embodiment of the present invention. Detailed Implementation

[0018] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] The terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish different objects, not to describe a specific order. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion.

[0020] Specific features, structures, or characteristics described in connection with the embodiments may be included in at least one embodiment of the invention. The appearance of this phrase in various places in the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment that is mutually exclusive with other embodiments.

[0021] Example 1 Please see Figure 1 , Figure 1 This is a flowchart illustrating a virtual-real joint simulation training method for an unmanned intelligent system disclosed in an embodiment of the present invention. Figure 1The described method is applied to unmanned intelligent system simulation training systems, such as local servers or cloud servers used for training unmanned intelligent systems, and the embodiments of the present invention are not limited thereto.

[0022] It should be noted that, for reference Figure 2 The virtual-real joint simulation training method of this invention is based on a five-layer collaborative architecture: a physical entity layer, a digital mapping layer, a virtual simulation layer, a data interaction layer, and a training control and evaluation layer. Each layer is connected through standardized interfaces. The physical entity layer, serving as the physical foundation for training, includes the physical entity of the unmanned intelligent system, sensor modules, actuator modules, and a ground control terminal. It executes training control commands, collects real-time physical state data and environmental data, uploads the collected data to the data interaction layer, and receives virtual environment data and control correction commands from the data interaction layer. The digital mapping layer constructs a digital equipment model that is precisely mapped to the physical entity layer, achieving real-time synchronization between the virtual and physical worlds. The digital equipment model acts as a bridge between the virtual simulation layer and the physical entity layer, enabling precise data mapping and interaction. In extreme scenarios or high-risk mission training, it can replace physical entities for virtual simulation. The virtual simulation layer consists of a dynamic scene generation engine, a virtual environment rendering module, and a task configuration module. It generates virtual training scenarios that flexibly adapt to training needs and operates collaboratively with the physical entity and digital equipment model. The data interaction layer, built on a real-time data interaction bus, is the core channel for realizing virtual-real integration. It is used for synchronous acquisition, parsing and conversion, real-time transmission, and synchronization of data from various layers, with an acquisition frequency of no less than 100Hz and an end-to-end latency of no more than 100ms. The training control and evaluation layer, as the system's control unit, is used to achieve closed-loop management and accurate evaluation of the training process, including training control, multi-dimensional evaluation, evaluation analysis, and collaborative training control.

[0023] like Figure 1 As shown, the method includes the following steps: S1, obtain the physical entity device configuration data set, training task configuration data, and training parameter configuration data.

[0024] It should be noted that the physical entity device configuration data information set includes several physical entity device configuration data information sets. Each physical entity device configuration data information set contains device type information, device communication protocol information, sensor parameter data information, actuator parameter data information, and device geometric parameter data information. The physical entity device configuration data information set refers to the set of configuration parameters of physical entities in an unmanned intelligent system that has been connected to the data interaction layer through a standardized interface. The physical entities in the unmanned intelligent system include, but are not limited to, drones, unmanned vehicles, and unmanned ships.

[0025] It should be noted that the training task configuration data includes task type information, task objective parameter data, task constraint parameter data, and task triggering condition data. The training task configuration data refers to the task configuration parameters preset by the training administrator according to training requirements.

[0026] It should be noted that the training parameter configuration data includes scenario type information, initial difficulty level information, operator skill level information, evaluation indicator weight data, and historical training data. The training parameter configuration data refers to the training control parameters generated by the training system based on the training objectives and operator skill level, or configured by the training administrator. The historical training data refers to the collection of evaluation results and process data generated by the operator or training subject in previous training sessions, which may include historical comprehensive training scores, scores for each dimension, weakness identifiers, and training time, used to calculate historical training difficulty deviation values ​​and generate interference parameters.

[0027] S2, Modeling and scene generation processing are performed on the physical entity equipment configuration data information set, the training task configuration data information, and the training parameter configuration data information to obtain the digital equipment model data information set and the virtual training scene data information.

[0028] It should be noted that the modeling and scene generation process includes digital equipment model construction and virtual training scene generation. The digital equipment model data information set includes several digital equipment model data information sets, each of which contains geometric mapping model data information, physical mapping model data information, control mapping model data information, and state synchronization configuration data information. The digital equipment model data information corresponds one-to-one with the physical entity equipment configuration data information.

[0029] It should be noted that the virtual training scenario data includes environmental parameter data, task parameter data, and interference parameter data. The environmental parameter data includes terrain data, meteorological condition data, light intensity values, and electromagnetic interference intensity values. The interference parameter data includes sensor noise intensity values, simulated communication delay values, and simulated equipment fault configuration data.

[0030] S3, perform virtual-real joint training processing on the digital equipment model data information set and the virtual training scene data information to obtain the training process data information set.

[0031] It should be noted that the virtual-real joint training process includes training initiation, real-time data interaction and synchronization, and collaborative control processing. The training process data information set includes several training process data information sets, each of which contains device status timing data information, operation command timing data information, task execution result data information, and virtual-real synchronization deviation data information.

[0032] S4, evaluate and perform closed-loop optimization on the training process data information set and the training parameter configuration data information to obtain target training evaluation report information and optimized training parameter configuration data information.

[0033] It should be noted that the evaluation and closed-loop optimization process includes multi-dimensional evaluation calculation and training parameter closed-loop optimization. The target training evaluation report information includes a comprehensive training score, data sets of scores for each dimension, strengths analysis data, and suggestions for improving weaknesses data. The optimized training parameter configuration data is used as the training parameter configuration data for the next round of training, realizing a closed-loop iteration of training-evaluation-optimization.

[0034] As can be seen, the virtual-real joint simulation training method for unmanned intelligent systems described in the embodiments of the present invention solves the technical problems of low virtual-real synchronization accuracy, poor scene adaptability, single evaluation dimension and insufficient cross-device compatibility in existing virtual-real joint simulation training by constructing a precise mapping between physical entities and digital equipment models, and combining dynamic scene generation and multi-dimensional closed-loop evaluation optimization. This improves the accuracy, personalization and cross-platform compatibility of unmanned intelligent system simulation training.

[0035] In an optional embodiment, the step of modeling and generating scenarios from the physical entity equipment configuration data set, the training task configuration data information, and the training parameter configuration data information to obtain a digital equipment model data set and virtual training scenario data information includes: S21, Perform digital equipment model construction processing on the physical entity equipment configuration data information set to obtain a digital equipment model data information set; the digital equipment model data information set includes several digital equipment model data information sets. It should be noted that the configuration data information of the physical entity equipment corresponds one-to-one with the data information of the digital equipment model.

[0036] It should be noted that the digital equipment model construction process refers to performing geometric mapping, physical mapping, and control mapping on each physical entity device to construct a digital model that accurately corresponds to the physical entity.

[0037] S22, perform virtual training scene generation processing on the training task configuration data information and the training parameter configuration data information to obtain virtual training scene data information.

[0038] It should be noted that the virtual training scene generation process refers to the automatic generation of virtual training scenes by the dynamic scene generation engine based on the training objectives, operational level, and historical training data.

[0039] As can be seen, the virtual-real joint simulation training method for unmanned intelligent systems described in the embodiments of the present invention combines the precise digitization of physical entities with the scenario-based training requirements by constructing digital equipment models and generating virtual training scenarios respectively, providing a high-fidelity digital foundation and a flexible and adaptable training environment for subsequent virtual-real joint training.

[0040] In another optional embodiment, the step of performing digital equipment model construction processing on the physical entity equipment configuration data information set to obtain a digital equipment model data information set includes: S211, Perform geometric mapping processing on the device geometric parameter data information in any of the physical entity device configuration data information in the physical entity device configuration data information set to obtain the geometric mapping model data information of the physical entity device configuration data information; It should be noted that the geometric mapping process refers to constructing a 1:1 scale 3D geometric model of the physical entity based on its 3D design drawings and laser scanning data, thus reproducing the device's appearance, structural dimensions, and motion joints. This geometric mapping process can be performed using 3ds Max, Blender, or SolidWorks; however, this embodiment of the invention does not impose any specific limitations.

[0041] It should be noted that S211-S214 processes any one of the physical entity equipment configuration data information in the physical entity equipment configuration data information set, ultimately obtaining the digital equipment model data information of that physical entity equipment configuration data information. In S21, the obtained digital equipment model data information set is obtained by performing S211-S214 on all physical entity equipment configuration data information in the physical entity equipment configuration data information set, integrating all the obtained geometric mapping model data information together, which is the geometric mapping model data information set.

[0042] S212, Perform physical mapping processing on the sensor parameter data information and actuator parameter data information in the configuration data information of the physical entity device to obtain the physical mapping model data information of the configuration data information of the physical entity device. It should be noted that the physical mapping process refers to reproducing the motion laws, mechanical responses, and sensor output characteristics of a device in a virtual environment by acquiring the physical characteristic parameters of the physical entity. These physical characteristic parameters include aerodynamic coefficients, inertial parameters, motor output characteristics, and sensor noise model parameters. The physical mapping process can be performed using Gazebo, Unity, or MATLAB / Simulink, etc., but the specific implementation in this invention is not limited to any particular method.

[0043] S213, Perform control mapping processing on the device type information in the configuration data information of the physical entity device to obtain the control mapping model data information of the configuration data information of the physical entity device. It should be noted that the control mapping process refers to replicating the control logic of the physical entity, including PID control algorithm parameters, path planning algorithm parameters, and obstacle avoidance algorithm parameters, to ensure that the control behavior of the digital equipment model is consistent with that of the physical entity. The control mapping process can be performed using ROS2, ArduPilot, or MATLAB / Simulink, etc. Specifically, this embodiment of the invention does not limit the specific implementation.

[0044] S214, integrate the geometric mapping model data information, the physical mapping model data information and the control mapping model data information of the physical entity equipment configuration data information to obtain the digital equipment model data information corresponding to the physical entity equipment configuration data information; It should be noted that the integration process refers to fusing and encapsulating the geometric mapping model, physical mapping model, and control mapping model, and configuring state synchronization configuration data to generate a complete digital equipment model. The state synchronization configuration data includes a synchronization sampling frequency value and a maximum permissible synchronization error threshold. The synchronization sampling frequency value is not less than 100Hz, and the maximum permissible synchronization error threshold is not greater than 100ms. The state synchronization configuration data can be user-defined or obtained from historical data; specifically, this embodiment of the invention does not impose limitations.

[0045] As can be seen, the virtual-real joint simulation training method for unmanned intelligent systems described in the embodiments of the present invention achieves accurate three-dimensional digitization of physical entities by performing geometric mapping, physical mapping and control mapping on each physical entity and integrating them into a complete digital equipment model, thus providing a high-fidelity digital foundation for real-time state synchronization and virtual simulation in virtual-real joint training.

[0046] In another optional embodiment, the virtual training scenario generation process performed on the training task configuration data information and the training parameter configuration data information to obtain virtual training scenario data information includes: S221, Perform environmental parameter generation processing on the scene type information and initial difficulty level information in the training parameter configuration data information to obtain environmental parameter data information; It should be noted that the environmental parameter generation process refers to matching and generating corresponding terrain data, meteorological condition data, light intensity, and electromagnetic interference intensity from a preset scene parameter library based on the scene type and initial difficulty level. The preset scene parameter library is a collection of several pre-stored sets of environmental parameter configuration data. Each set of environmental parameter configuration data corresponds to a combination of scene type and difficulty level, including terrain data, meteorological condition data, light intensity values, and electromagnetic interference intensity values ​​under that combination. The environmental parameter generation process matches the scene type information and the initial difficulty level information with the indexes in the scene parameter library. The scene type information includes urban terrain, mountainous terrain, marine terrain, and desert terrain, etc.

[0047] S222, Perform task parameter generation processing on the operation level information in the training task configuration data information and the training parameter configuration data information to obtain task parameter data information; It should be noted that the task parameter data includes the task start coordinates, task end coordinates, target point distribution, and time limit. The task parameter generation process can utilize the RRT algorithm, genetic algorithm, or a pre-defined task template library, etc. Specifically, this embodiment of the invention does not impose any limitations.

[0048] S223, perform interference parameter generation processing on the initial difficulty level information and historical training data information in the training parameter configuration data information to obtain interference parameter data information; It should be noted that the interference parameter generation process refers to automatically configuring sensor noise intensity, communication delay simulation, and equipment failure simulation parameters based on the initial difficulty level and weak points data in historical training data. Low interference parameters are generated for operators with lower skill levels; high interference parameters, including strong wind interference, sudden equipment failure, and complex electromagnetic interference, are generated for operators with higher skill levels. The above interference parameter generation process can use Gaussian noise models, Poisson processes, or Monte Carlo simulations, etc., and the specific implementation in this invention is not limited thereto.

[0049] It should be noted that the aforementioned historical training data information can be user-defined or obtained from a historical database through an API interface. Specifically, this embodiment of the invention does not impose any limitations.

[0050] S224, using a dynamic scene difficulty adaptive calculation model, calculates and processes the environmental parameter data, the task parameter data, and the interference parameter data to obtain the comprehensive scene difficulty value; The dynamic scene difficulty adaptive calculation model is as follows: ; ; In the formula, The overall difficulty value of the scenario. This represents the normalized difficulty value of the p-th environmental sub-parameter in the environmental parameter data, where P is the number of environmental sub-parameters. This represents the normalized difficulty value of the q-th task sub-parameter in the task parameter data, where Q is the number of task sub-parameters. R is the normalized difficulty value of the r-th interference sub-parameter in the interference parameter data information, where R is the number of interference sub-parameters. , and These are the weighting coefficients for environmental parameters, task parameters, and disturbance parameters, respectively. To adapt and adjust parameters based on historical training, The historical training difficulty deviation value in the historical training data information is used to adjust the scene difficulty based on historical training performance.

[0051] It should be noted that the dynamic scene difficulty adaptive calculation model is used to uniformly quantify environmental parameter data, task parameter data, interference parameter data, and historical training difficulty deviation values ​​obtained based on the historical training data into the comprehensive scene difficulty value. This provides a basis for scene rationality verification, thereby achieving controllable difficulty and personalized adaptation during the virtual training scene generation stage. The environmental parameter weight coefficient ranges from [0.2, 0.5], with a preferred value of 0.3. The task parameter weight coefficient ranges from [0.2, 0.5], with a preferred value of 0.4. The interference parameter weight coefficient ranges from [0.1, 0.4], with a preferred value of 0.3. The historical training adaptation adjustment parameter ranges from [0, 0.3], with a preferred value of 0.1. The normalized difficulty value ranges from [0, 1], where 0 represents the lowest difficulty and 1 represents the highest difficulty. The historical training difficulty deviation value is calculated as follows: the average value of the historical comprehensive training score is extracted from the historical training data information, and the difference between the average value and the preset target training score is normalized and mapped to the preset deviation range to obtain the historical training difficulty deviation value; when the average value is lower than the target training score, the historical training difficulty deviation value is positive and the scene difficulty is adjusted upward; when the average value is higher than the target training score, the historical training difficulty deviation value is negative and the scene difficulty is adjusted downward.

[0052] S225, Perform scene rationality verification on the comprehensive difficulty value of the scene to obtain virtual training scene data information; It should be noted that the scenario rationality verification process refers to comparing whether the overall difficulty value of the scenario falls within the preset difficulty range corresponding to the current initial difficulty level information. If it falls within the range, the environmental parameter data information, the task parameter data information, and the interference parameter data information constitute the virtual training scenario data information. If it does not fall within the range, the overall difficulty value of the scenario is recalculated after scaling each interference sub-parameter in the interference parameter data information, and this process is repeated until the overall difficulty value of the scenario falls within the preset difficulty range. Each interference sub-parameter in the interference parameter data information includes sensor noise intensity values, communication delay simulation values, and scalable items (such as fault trigger probability, fault duration, etc.) in the equipment fault simulation configuration data information. The scaling adjustment refers to multiplying each of the above interference sub-parameters by the same scaling factor or by their respective scaling factors and then resubmitting them into the dynamic scenario difficulty adaptive calculation model to calculate the overall difficulty value of the scenario. The preset difficulty range corresponds to the difficulty level. For example, the difficulty range corresponding to the beginner level is [0.2, 0.4], the difficulty range corresponding to the intermediate level is [0.4, 0.7], and the difficulty range corresponding to the advanced level is [0.6, 1.0]. The specific range values ​​are not limited in this embodiment of the invention.

[0053] As can be seen, the virtual-real joint simulation training method for unmanned intelligent systems described in the embodiments of the present invention automatically generates virtual training scenarios by combining a dynamic scene difficulty adaptive calculation model with historical training data, and performs scene rationality verification, thereby achieving personalized adaptation of training scenarios and precise control of difficulty, avoiding the problem of traditional scene generation relying on manual editing.

[0054] In an optional embodiment, the step of performing virtual-real joint training processing on the digital equipment model data information set and the virtual training scene data information to obtain the training process data information set includes: S31, Perform training initialization processing on the digital equipment model data information set and the virtual training scene data information to obtain the initial training state data information; It should be noted that the training initialization process involves sending start commands to the physical entity layer, digital mapping layer, and virtual simulation layer, causing the physical entities to enter a ready-to-operate state, synchronizing the digital equipment model with the initial state of the physical entities, and loading the virtual training scene data information into the virtual simulation layer. The initial training state data information includes the initial position data, initial attitude data, and initial equipment state data of each physical entity.

[0055] S32, perform real-time data interaction and synchronization processing on the initial training state data information and the digital equipment model data information set to obtain a virtual-real synchronization state data information set; the virtual-real synchronization state data information set includes virtual-real synchronization state data information at several moments. S33, perform coordinated control and instruction execution processing on the virtual-real synchronization state data information set, the digital equipment model data information set, and the virtual training scene data information to obtain the training process data information set.

[0056] It should be noted that the collaborative control and instruction execution processing includes sub-processes such as collaborative task allocation, collaborative control and instruction execution, anomaly detection and handling, and collaborative execution monitoring. The virtual training scenario data information includes task parameter data information, used for task decomposition and sub-task complexity calculation during collaborative task allocation. During training, device status timing data, operation instruction timing data, task execution result data, and virtual-real synchronization deviation data are continuously collected to form the training process data information set.

[0057] As can be seen, the virtual-real joint simulation training method for unmanned intelligent systems described in the embodiments of the present invention achieves high-fidelity joint training of virtual environment and physical entity through training initialization, real-time data interaction synchronization and collaborative control, ensuring the complete acquisition of training process data.

[0058] In another optional embodiment, the real-time data interaction and synchronization processing of the initial training state data information and the digital equipment model data information set to obtain a virtual-real synchronization state data information set includes: S321, Perform multi-source data acquisition and processing on the initial training state data information to obtain the original multi-source data information set; It should be noted that the multi-source data acquisition and processing refers to the synchronous acquisition of sensor and actuator data from the physical entity layer, model state data from the digital mapping layer, and scene data from the virtual simulation layer through the data interaction layer, with an acquisition frequency of no less than 100Hz. This multi-source data acquisition and processing supports mainstream communication protocols such as ROS2, MQTT, and TCP / IP, and can be integrated with unmanned intelligent systems of different types and manufacturers without requiring large-scale modifications to the physical entities.

[0059] S322, perform data parsing and format conversion on the original multi-source data information set to obtain a standardized data information set; It should be noted that the data parsing and format conversion processing refers to converting heterogeneous data from multiple sources, such as analog signals from sensors, digital signals from control terminals, and image data from virtual scenes, into a standardized data format through device adaptation middleware. The device adaptation middleware is a software module deployed at the data interaction layer to connect unmanned intelligent systems from different manufacturers and models with the data interaction layer. It can automatically identify device types and communication protocols, complete data format conversion and protocol adaptation, and achieve cross-platform, multi-device data unification without requiring large-scale modifications to physical entities.

[0060] S323, perform timestamp alignment processing on the standardized data information set to obtain a time-aligned data information set; It should be noted that the timestamp alignment process refers to attaching a uniform timestamp to the data packets of each data source and aligning the data from different data sources according to the timestamp, thereby eliminating time offsets caused by differences in transmission paths.

[0061] S324, The time-aligned data information set is processed to obtain the virtual-real synchronization deviation value; It should be noted that the above calculation and processing can be carried out using a virtual-real state deviation measurement calculation model, or using methods such as weighted Euclidean distance, Mahalanobis distance, or root mean square error combination. Specifically, the embodiments of the present invention do not limit the specifics.

[0062] The calculation model for the virtual-real state deviation measurement is as follows: ; ; In the formula, This is the virtual-real synchronization deviation value. This refers to the position vector corresponding to the physical entity in the time-aligned data set. The position vector corresponding to the digital equipment model in the time-aligned data information set. The velocity vector corresponding to the physical entity in the time-aligned data set. The velocity vector is the corresponding velocity vector of the digital equipment model in the time-aligned data information set. This refers to the attitude angle vector corresponding to the physical entity in the time-aligned data information set. The attitude angle vector is the digital equipment model corresponding to the time-aligned data information set. , and These are the position normalization factor, velocity normalization factor, and attitude angle normalization factor, respectively, used to unify deviations of different dimensions into a dimensionless space. , and These are the position deviation weighting coefficient, velocity deviation weighting coefficient, and attitude angle deviation weighting coefficient, respectively. The Euclidean norm of a vector is denoted by .

[0063] It should be noted that the position deviation weighting coefficient ranges from [0.3, 0.5], with a preferred value of 0.4. The velocity deviation weighting coefficient ranges from [0.2, 0.4], with a preferred value of 0.3. The attitude angle deviation weighting coefficient ranges from [0.2, 0.4], with a preferred value of 0.3. The position normalization factor, velocity normalization factor, and attitude angle normalization factor are determined based on the maximum motion range, maximum velocity, and maximum attitude angle variation range of the physical entity, respectively.

[0064] It should be noted that the virtual-real state deviation measurement calculation model is used to normalize and weightedly fuse the deviations in position, velocity, and attitude angle between the physical entity and the digital equipment model in the time-aligned data information set to obtain a single virtual-real synchronization deviation value. This provides a basis for determining whether motion prediction compensation should be triggered subsequently, thereby controlling the virtual-real synchronization error within a preset threshold range.

[0065] S325, determine whether the virtual-real synchronization deviation value is greater than the preset maximum allowable synchronization error threshold, and obtain the first judgment result; When the first judgment result is yes, motion prediction compensation processing is performed on the digital equipment model corresponding to the digital equipment model data information set to obtain the compensated virtual-real synchronization state data information, and the compensated virtual-real synchronization state data information is determined to be the virtual-real synchronization state data information at the current moment. When the first judgment result is negative, the time-aligned data information set is determined to be the virtual-real synchronization state data information at the current moment.

[0066] It should be noted that the preset maximum allowable synchronization error threshold is a threshold used to determine whether the virtual-real synchronization deviation exceeds an acceptable range. When the virtual-real synchronization deviation exceeds this threshold, motion prediction compensation needs to be triggered. The maximum allowable synchronization error threshold corresponds to the state synchronization configuration data information in the digital equipment model data information, and is used to control the state synchronization error between the virtual environment and the physical entity within an acceptable range. Typically, the corresponding time synchronization error is no greater than 100ms. The acquisition, parsing, alignment, deviation measurement, and compensation processes from S321 to S325 are continuously executed to obtain the virtual-real synchronization state data information set. The motion prediction compensation process uses linear extrapolation or Kalman filtering algorithms to perform short-term prediction of the state of the physical entity, with the prediction time window not exceeding the maximum allowable synchronization error threshold. The above motion prediction compensation process can use Kalman filtering, extended Kalman filtering, or unscented Kalman filtering; specifically, this embodiment of the invention does not limit the specific methods used.

[0067] As can be seen, the virtual-real joint simulation training method for unmanned intelligent systems described in the embodiments of the present invention achieves low-latency and high-precision synchronization between the virtual environment and the physical entity through a complete process of multi-source data acquisition, standardization conversion, timestamp alignment, virtual-real deviation measurement and state compensation, and controls the virtual-real synchronization deviation within a preset threshold range, thus solving the problem of virtual-real asynchrony in the prior art.

[0068] In another optional embodiment, the step of coordinating and executing the virtual-real synchronization state data information set, the digital equipment model data information set, and the virtual training scenario data information to obtain the training process data information set includes: S331, Process the digital equipment model data information set and the virtual training scene data information to obtain a collaborative task allocation data information set; It should be noted that the above processing can be obtained using a collaborative task balancing calculation model, or it can be calculated using a greedy algorithm, a genetic algorithm, or a particle swarm optimization algorithm. Specifically, the embodiments of the present invention do not limit the specifics.

[0069] It should be noted that the above processing refers to decomposing the task parameter data information in the virtual training scenario data information into several sub-tasks to form a complete sub-task set, and solving the sub-task allocation scheme that balances the task load of each device based on the collaborative task balancing allocation calculation model. The optimized sub-task set of each device is then encoded as the collaborative task allocation data information set. The collaborative task allocation data information set includes several collaborative task allocation data information, which correspond one-to-one with each device participating in collaborative training. The device identifier information in the m-th collaborative task allocation data information corresponds to the m-th device, and the sub-task content data information is the set of sub-tasks allocated to the m-th device in the optimization result of the calculation model. The included subtasks and collaborative constraint data information are determined by the constraints of the computational model. When training on a single device, the collaborative task allocation data set may contain only the subtask allocation information corresponding to one device.

[0070] The collaborative task allocation data information set is the set of subtasks for each device obtained by solving the computational model. Obtained through encoding.

[0071] The collaborative task balanced allocation calculation model is as follows: ; ; ; In the formula, Let m be the task load value of the m-th device. The total number of devices participating in collaborative training is consistent with the number of digital equipment model data information in the digital equipment model data information set. The computational model solves for the set of subtasks assigned to the m-th device. smallest The data is encoded as the subtask content data of each collaborative task allocation data in the collaborative task allocation data information set. The task complexity value of the j-th subtask is determined by the task parameter data information in the virtual training scenario data information based on the task path length, the number of targets, and the number of constraints. The set of all subtasks is obtained by decomposing the task parameter data information in the virtual training scenario data. The optimization objective of the collaborative task balancing allocation calculation model is to minimize the maximum task load value among all devices, so as to balance the task load of each device as much as possible.

[0072] It should be noted that the aforementioned set of all subtasks The acquisition method is as follows: Task parameter data information in the virtual training scene data information is processed by task decomposition. The task parameter data information includes task start coordinate data, task end coordinate data, target point distribution data, and time limit value. The overall training task is decomposed into several sub-tasks according to a preset sub-task partitioning strategy, which includes partitioning by target point, partitioning by path segment, or partitioning by time window. The set of indices of each sub-task is determined as the set of all sub-tasks. The task complexity value The method for obtaining the task complexity value is as follows: For the j-th subtask in the set of all subtasks, the task complexity value of the j-th subtask is calculated by weighting or weighted summing based on the path length, the number of target points involved, and the number of constraints. The path length is determined by the geometric distance or planned path length between the task start point and the task end point or target point corresponding to the subtask. The number of constraints refers to the number of time limits, no-fly zones, or route constraints associated with the subtask. The above-mentioned collaborative task allocation processing can use a greedy algorithm, genetic algorithm, or particle swarm optimization algorithm to solve the computational model to obtain the collaborative task allocation data information set. Specifically, this embodiment of the invention does not limit the specific implementation.

[0073] It should be noted that the collaborative task balanced allocation calculation model is used to solve the subtask allocation scheme that makes the task load of each device as balanced as possible, given the complete set of subtasks and the complexity of each subtask. The optimized set of subtasks of each device is encoded into the collaborative task allocation data information set to support the task decomposition and allocation of multi-device collaborative training.

[0074] S332, Perform collaborative control and instruction execution processing on the virtual-real synchronization state data information set and the collaborative task allocation data information set to obtain the training process data information set; It should be noted that the aforementioned collaborative control and command execution processing refers to the operator sending operation commands to the physical entity through a control terminal. These commands are synchronized to the digital equipment model and virtual simulation layer via a data interaction layer. The physical entity executes the operation commands, the digital equipment model synchronously simulates the operation effect, and the virtual simulation layer presents the movement and operational status of the physical entity in a virtual scene, achieving collaborative operation of physical operation, virtual presentation, and real feedback. During training, equipment status timing data, operation command timing data, task execution result data, and virtual-real synchronization deviation data are continuously collected to form the training process data information set.

[0075] In an optional embodiment, after S332, the following may also be included: S333, calculate and process the device status time-series data information in the training process data information set to obtain an abnormal deviation detection value, and compare the abnormal deviation detection value with the preset abnormal detection threshold to obtain abnormal detection result information; It should be noted that the above calculation and processing can be performed using an anomaly detection calculation model, or using statistical outlier detection methods, isolated forest algorithms, or local outlier factor algorithms. Specifically, the embodiments of the present invention do not limit the specific methods.

[0076] It should be noted that the anomaly detection and processing refers to real-time monitoring of equipment status data and data transmission status during training and operation to detect whether there are equipment malfunctions or data transmission anomalies.

[0077] It should be noted that the comparison process refers to comparing the abnormal deviation detection value with a preset abnormal detection threshold. When the abnormal deviation detection value is greater than the abnormal detection threshold, an anomaly is determined to exist and an anomaly type identifier is recorded. When the abnormal deviation detection value is not greater than the abnormal detection threshold, the current training is determined to be normal. The above comparison conclusions and relevant detection data are integrated into the anomaly detection result information. The anomaly detection result information includes an anomaly identifier, an anomaly type identifier, an abnormal deviation detection value, a comparison time stamp, and an anomaly triggering device identifier. The anomaly identifier indicates whether the training process within the current sliding window is normal or abnormal. The anomaly type identifier indicates the anomaly category (such as device hardware failure or data transmission anomaly). The comparison time stamp is the time identifier when the comparison process is performed. When an anomaly exists, the anomaly triggering device identifier indicates the device or data source that caused the abnormal deviation.

[0078] The abnormal deviation detection calculation model is as follows: ; In the formula, This refers to the abnormal deviation detection value. This represents the number of data points within the sliding window. The device status value is the nth data point within the sliding window of the device status time sequence data information in the training process data information set. The average value of the device status time sequence data over the reference period. The standard deviation of the device status timing data information within the reference time period is given. To prevent division by a constant.

[0079] It should be noted that the number of data points within the sliding window ranges from [10, 100], with a preferred value of 50. When the abnormal deviation detection value exceeds the preset abnormal detection threshold, an anomaly is identified, triggering an alarm and adjusting the training state according to the anomaly type; when the abnormal deviation detection value is not greater than the preset abnormal detection threshold, training is considered normal. The abnormal detection threshold ranges from [2.0, 5.0], with a preferred value of 3.0. When an anomaly is detected, the training control module performs the following processing based on the anomaly type: if it is a hardware failure, the training of the corresponding physical entity is paused, and the digital equipment model performs virtual simulation instead; if it is a data transmission anomaly, the system switches to a backup communication channel and resynchronizes the data.

[0080] It should be noted that the abnormal deviation detection calculation model is used to measure the deviation between the device status time-series data information in the training process data information set and the statistical quantity of the reference time period within the sliding window, and obtain the abnormal deviation detection value, which provides input for comparison processing, thereby supporting real-time abnormal detection and security protection in the training process.

[0081] S334, Perform collaborative execution monitoring processing on the collaborative task allocation data information set and the virtual-real synchronization status data information set to obtain collaborative training process data information; It should be noted that the aforementioned collaborative execution monitoring and processing refers to real-time monitoring of the task execution progress and collaborative cooperation status of each device during multi-device joint training. When the task execution of a device deviates from the expected trajectory, a collaborative adjustment instruction is sent to other relevant devices to ensure the overall consistency of multi-device collaborative operation. The specific execution method of the collaborative execution monitoring and processing is as follows: Based on the sub-task content data information and collaborative constraint condition data information of each collaborative task allocation data information set, the expected execution trajectory or expected progress time node corresponding to each device is determined; the position, attitude, and task execution status of each device at each moment are extracted from the virtual-real synchronization status data information set and compared with the expected execution trajectory or expected progress time node; when the deviation between the current state of a device and the expected execution trajectory is greater than a preset collaborative deviation threshold, it is determined that the task execution of the device has deviated from the expectation, a collaborative adjustment instruction is generated and sent to the device and related devices with collaborative constraints, the collaborative adjustment instruction includes the adjustment target, priority, and effective time; the task execution progress, collaborative cooperation status, and issued collaborative adjustment instruction records of each device are continuously collected and integrated into the collaborative training process data information. The collaborative training process data information includes task execution progress data information of each device, collaborative cooperation status data information of each device, and collaborative adjustment instruction record data information, which are used as part of the training process data information set for subsequent evaluation.

[0082] S335, the collaborative training process data information is added to the training process data information set to obtain an updated training process data information set, and the updated training process data information set is determined as the training process data information set.

[0083] As can be seen, the virtual-real joint simulation training method for unmanned intelligent systems described in the embodiments of the present invention realizes the joint training of multiple unmanned intelligent systems, supports cross-platform and multi-model collaborative operation training, and ensures the safety and continuity of the training process through real-time abnormal deviation detection and automated abnormal handling mechanisms, thus solving the problems of insufficient support for multi-device collaborative training and training interruption in the prior art.

[0084] In another optional embodiment, the step of evaluating and performing closed-loop optimization on the training process data information set and the training parameter configuration data information to obtain target training evaluation report information and optimized training parameter configuration data information includes: S41, perform data preprocessing on the training process data information set to obtain a preprocessed training process data information set; It should be noted that the data preprocessing includes data cleaning and standardization. Data cleaning removes outliers and missing values ​​from the training data set, while standardization unifies data from different dimensions to the same unit of measurement. The data preprocessing can utilize Z-score standardization, Min-Max normalization, or median absolute deviation; however, this embodiment of the invention does not impose specific limitations.

[0085] S42, perform multi-dimensional index calculation processing on the preprocessed training process data information set to obtain a multi-dimensional evaluation index data information set; S43, perform comprehensive evaluation and calculation on the multi-dimensional evaluation index data information set and the training parameter configuration data information to obtain the target training evaluation report information; S44, perform closed-loop optimization processing on the target training evaluation report information and the training parameter configuration data information to obtain optimized training parameter configuration data information.

[0086] As can be seen, by implementing the virtual-real joint simulation training method for unmanned intelligent systems described in the embodiments of the present invention, the training process data is preprocessed, multi-dimensional indexes are calculated, comprehensive evaluation is performed, and closed-loop optimization is performed. This achieves accurate quantitative evaluation of training effects and automatic optimization of training parameters, providing a data-driven optimization scheme for subsequent training.

[0087] In an optional embodiment, the step of performing multi-dimensional index calculation processing on the preprocessed training process data information set to obtain a multi-dimensional evaluation index data information set includes: S421, Process the device status time-series data information in the preprocessed training process data information set to obtain device status index data information; It should be noted that the device status index data includes device failure rate, battery life utilization rate, sensor data accuracy rate, and actuator response speed. The device failure rate is the ratio of the number of device failures during training to the total operating time. The battery life utilization rate is the ratio of actual power consumption to total available power. The sensor data accuracy rate is the ratio of the effective data volume of the sensors to the total data volume collected by the sensors. The actuator response speed is the average time from receiving a command to completing a response. The above processing can be performed using statistical counting methods, sliding window aggregation, or temporal feature extraction algorithms; specifically, this embodiment of the invention does not limit the specific processing.

[0088] S422, Process the timing data of the operation instructions in the preprocessed training process data information set to obtain operation behavior index data information; It should be noted that the operational behavior indicator data includes operational command response time, operational error, and fault handling time. The operational command response time is the average time from when the operator perceives the event to when they issue the operational command. The operational error is the deviation between the target state and the actual execution state of the operational command. The fault handling time is the average time from when the equipment malfunction occurs to when the operator completes the fault handling. The above processing can be performed using timestamp difference statistics, state comparison algorithms, or event log analysis methods; specifically, this embodiment of the invention does not limit the specific methods used.

[0089] S423, Process the task execution result data information in the preprocessed training process data information set to obtain task performance index data information; It should be noted that the task performance metrics data include task completion rate, task completion time, and target achievement accuracy. The task completion rate is the ratio of the number of completed tasks to the total number of tasks. The task completion time is the actual time taken to complete each task. The target achievement accuracy is the deviation between the actual target location and the planned target location. The above processing can be performed using completion count statistics, time accumulation algorithms, or Euclidean distance and geometric deviation calculation methods; specifically, this embodiment of the invention does not limit the specific methods used.

[0090] S424, Process the virtual-real synchronization deviation data information and the operation instruction timing data information in the preprocessed training process data information set to obtain the cooperative index data information; It should be noted that the cooperative coordination index data includes path planning rationality value, obstacle avoidance response speed value, and multi-device cooperative coordination degree value. The path planning rationality value is the ratio of the actual driving path length to the theoretical shortest path length. The obstacle avoidance response speed value is the average time from detecting an obstacle to completing the obstacle avoidance action. The multi-device cooperative coordination degree value is the normalized value of the synchronization deviation of the actions of multiple devices in the cooperative task. The above processing can be performed using path length accumulation and shortest path comparison, time difference statistics, or multi-source timing synchronization deviation measurement methods. Specifically, this embodiment of the invention does not limit the specific methods.

[0091] S425, integrate and process the equipment status index data information, the operation behavior index data information, the task effect index data information, and the cooperation index data information to obtain a multi-dimensional evaluation index data information set.

[0092] It should be noted that the above integration process involves merging the equipment status indicator data, the operational behavior indicator data, the task effect indicator data, and the collaborative cooperation indicator data to obtain a multi-dimensional evaluation indicator data set.

[0093] As can be seen, the virtual-real joint simulation training method for unmanned intelligent systems described in the embodiments of the present invention achieves a comprehensive quantitative analysis of the training process by calculating evaluation indicators from four dimensions: equipment status, operational behavior, task effect, and collaborative cooperation. This avoids the one-sidedness of traditional evaluation systems that only focus on task completion results.

[0094] In another optional embodiment, the step of comprehensively evaluating and calculating the multi-dimensional evaluation index data set and the training parameter configuration data to obtain the target training evaluation report information includes: S431, The multi-dimensional evaluation index data information set is subjected to index normalization processing to obtain a normalized evaluation index data information set; It should be noted that the index normalization process refers to using the Min-Max normalization method to map the values ​​of each evaluation index to the [0, 1] interval, thereby eliminating the dimensional differences between different indices. For positive indices, the closer the normalized value is to 1, the better the performance; for negative indices, their inverse normalized value is used.

[0095] S432, using the hierarchical analysis calculation model, calculate and process the normalized evaluation index data information set and the evaluation index weight data information in the training parameter configuration data information to obtain the comprehensive training score value and the score data information set of each dimension. It should be noted that the set of score data for each dimension includes several dimension score data, each corresponding one-to-one with the evaluation dimension. When the value of i ranges from 1 to 4, the set of score data for each dimension includes four dimension score data, corresponding to the device status dimension, operation behavior dimension, task effect dimension, and collaborative cooperation dimension, respectively. Each dimension score data includes dimension identifier information and dimension score value. The dimension identifier information is determined by the dimension division contained in the evaluation index weight data information in the training parameter configuration data information. The evaluation index weight data information has already divided each dimension and assigned dimension names or numbers in the analytic hierarchy process, and the dimension score value of the i-th dimension is calculated. When the dimension is specified, the dimension name or number corresponding to that dimension is written into the dimension identification information; or, the dimension identification information is determined according to the dimension index i and a preset dimension mapping relationship, where i=1, 2, 3, 4 correspond to the identifiers of the device status dimension, operation behavior dimension, task effect dimension, and collaborative cooperation dimension, respectively. The dimension score is the dimension score value of the i-th dimension in the calculation model. It is obtained by calculating the normalized values ​​of each evaluation indicator under this dimension and then summing them according to the indicator weights.

[0096] The hierarchical analysis calculation model is as follows: ; ; ; ; In the formula, The comprehensive training score is... Let be the dimension score of the i-th dimension, where i ranges from 1 to 4, corresponding to the device status dimension, operation behavior dimension, task effect dimension, and collaboration dimension, respectively. is the dimension weight coefficient of the i-th dimension. Let i be the number of evaluation metrics in the i-th dimension. is the index weight coefficient of the j-th evaluation index in the i-th dimension. It is the normalized value of the j-th evaluation indicator in the i-th dimension of the normalized evaluation indicator data set. Let be the fuzzy membership value of the j-th evaluation indicator in the i-th dimension, representing the degree of membership of the evaluation level corresponding to the indicator value. is the kurtosis parameter of the fuzzy membership function. Let be the fuzzy membership center parameter for the i-th dimension and j-th evaluation index.

[0097] It should be noted that the dimensional weight coefficients are determined using the analytic hierarchy process (AHP) based on the evaluation index weight data. Specifically, the process involves constructing pairwise comparison matrices between each dimension, calculating the eigenvectors as weights, and performing a consistency check. Weights are considered valid when the consistency ratio is less than 0.1. The kurtosis parameter ranges from [5, 20], with a preferred value of 10. The fuzzy membership center parameter ranges from [0.3, 0.7], with a preferred value of 0.5.

[0098] It should be noted that the hierarchical analysis calculation model is used to weight and fuse the normalized evaluation index data information set according to the dimension weights and index weights determined by the evaluation index weight data information, and after fuzzy membership degree calculation, to obtain the comprehensive training score value and the score data information set of each dimension. This provides data support for subsequent report generation and closed-loop optimization, and realizes multi-dimensional quantitative evaluation of training effect.

[0099] S433, integrate the comprehensive training score and the score data information set of each dimension to obtain the target training evaluation report information; It should be noted that the integration process refers to sorting and analyzing the scores of each dimension indicator based on the comprehensive training score and the score data information set of each dimension, marking indicators with scores higher than a preset advantage threshold as advantages, marking indicators with scores lower than a preset weakness threshold as weaknesses, generating corresponding weakness improvement suggestion data information, and integrating them into the target training evaluation report information. The specific method for integrating the information into the target training evaluation report is as follows: The comprehensive training score is written as the comprehensive score portion of the target training evaluation report; the score data information of each dimension (including dimension identifiers and dimension score values) in the score data information set of each dimension is sorted by dimension order or by score from high to low, and written as the score data set of each dimension indicator in the target training evaluation report; the score data set of each dimension is traversed, and dimensions with score values ​​higher than a preset advantage threshold and their scores are included in the advantage analysis data information and written as the advantage analysis data information in the target training evaluation report; dimensions with score values ​​lower than a preset weakness threshold are marked as weaknesses, and corresponding weakness improvement suggestion data information is generated for each weakness, and summarized as the weakness improvement suggestion data information in the target training evaluation report; the comprehensive score portion, the score data set of each dimension indicator, the advantage analysis data information, and the weakness improvement suggestion data information are combined into a complete target training evaluation report according to a preset report structure or report template. The generation method for the weakness improvement suggestion data information is as follows: For each dimension marked as a weakness, the difference between the dimension score and the preset weakness threshold is calculated based on the dimension identifier information and the dimension score value in the dimension score data information; based on the difference and the dimension identifier information, improvement suggestion text corresponding to the dimension and score range is matched from a preset improvement suggestion template library, where the improvement suggestion template library pre-stores the correspondence between each dimension, each score range and the improvement suggestion content; or, the improvement intensity level is determined based on the range to which the difference belongs, and then targeted weakness improvement suggestion data information is generated according to the dimension type represented by the dimension identifier information. For example, when the equipment status dimension is weak, suggestions to strengthen equipment inspection and maintenance are generated; when the operation behavior dimension is weak, suggestions to increase operation standard training are generated; when the task effect dimension is weak, suggestions to strengthen task goal achievement training are generated; when the collaboration dimension is weak, suggestions to strengthen multi-device collaborative exercises are generated. The advantage threshold ranges from [0.7, 0.9], with a preferred value of 0.8. The weakness threshold ranges from [0.3, 0.5], with a preferred value of 0.4.

[0100] As can be seen, the virtual-real joint simulation training method for unmanned intelligent systems described in the embodiments of the present invention determines the weights of each dimension through the analytic hierarchy process and calculates the comprehensive training score by combining the fuzzy comprehensive evaluation model. This achieves an objective and accurate quantitative evaluation of the training effect, avoids the subjectivity of manual evaluation, and provides data support for subsequent training optimization.

[0101] In another optional embodiment, the closed-loop optimization process of the target training evaluation report information and the training parameter configuration data information to obtain optimized training parameter configuration data information includes: S441, perform optimization direction parsing processing on the weakness improvement suggestion data information in the target training evaluation report information to obtain the optimization direction data information set; It should be noted that the optimization direction data information set includes several optimization direction data information sets, each of which contains the identifier information of the dimension to be optimized, the identifier information of the indicator to be optimized, and the optimization direction value. The above-mentioned optimization direction parsing process can be performed using keyword matching, a preset improvement suggestion and optimization direction mapping table, or natural language parsing methods. Specifically, this embodiment of the invention does not limit the specific methods used.

[0102] S442, calculate and process the optimization direction data information set, the training parameter configuration data information and the target training evaluation report information to obtain the parameter adjustment amount data information set; It should be noted that the above calculations can be performed using a closed-loop optimization calculation model with training parameters, or using gradient descent, Adam optimization algorithm, or proportional-integral-derivative adjustment method. Specifically, the embodiments of the present invention are not limited.

[0103] The closed-loop optimization calculation model for the training parameters is as follows: ; ; In the formula, This is the adjustment amount for the k-th training parameter. To optimize the learning rate parameter, The optimization direction value is the value corresponding to the kth training parameter in the optimization direction data set. The preset target training score value, This is the comprehensive training score. The dimension score value of the dimension to which the k-th training parameter belongs in the target training evaluation report information. The average score of all dimensions in the training evaluation report information for the target. This is the dimensional deviation amplification factor, used to make more significant adjustments to the parameters of weak dimensions. To prevent division by a constant. Configure the current value of the k-th training parameter in the training parameter configuration data information. The training parameter values ​​are adjusted from the data information set of the parameter adjustment amount. and Let be the minimum and maximum allowed values ​​for the k-th training parameter, respectively. The function is used to limit the adjusted parameter value to a allowed range.

[0104] It should be noted that the optimized learning rate parameter ranges from [0.01, 0.2], with a preferred value of 0.05. When the optimized learning rate parameter is large, the parameter adjustment range is large, and the training difficulty changes significantly; when the optimized learning rate parameter is small, the parameter adjustment range is small, and the training difficulty changes gradually. The dimensionality bias amplification coefficient ranges from [0.5, 2.0], with a preferred value of 1.0. The division-by-zero constant ranges from [1e-6, 1e-4], with a preferred value of 1e-5. The target training score ranges from [0.6, 0.95], with a preferred value of 0.8.

[0105] It should be noted that the training parameter closed-loop optimization calculation model is used to automatically calculate the adjustment amount of each training parameter based on the gap between the current comprehensive training score and the target training score, the deviation of the scores of each dimension from the average score, and the optimization direction data information, so as to realize the closed-loop adjustment of training parameter configuration such as training difficulty and scene parameters; so that the next round of training can be targeted to strengthen the weak dimensions, forming a closed-loop iteration of training-evaluation-optimization until the comprehensive training score reaches the target training score.

[0106] S443, Perform parameter update processing on the parameter adjustment data information set and the training parameter configuration data information to obtain optimized training parameter configuration data information.

[0107] It should be noted that the parameter update process refers to applying the parameter adjustment values ​​from the parameter adjustment data set to the corresponding training parameters in the training parameter configuration data to generate the optimized training parameter configuration data. The optimized training parameter configuration data can be used as input for the next round of training. The dynamic scene generation engine generates a new virtual training scene based on the optimized parameters, forming a closed-loop iteration of training-evaluation-optimization until the comprehensive training score reaches the target training score.

[0108] As can be seen, the virtual-real joint simulation training method for unmanned intelligent systems described in the embodiments of the present invention automatically calculates the adjustment amount of each training parameter through the closed-loop optimization calculation model of training parameters, and focuses on optimizing weak dimensions by combining the dimensional deviation amplification mechanism, thereby realizing the automated closed-loop optimization of training parameters and avoiding the inefficiency and subjectivity of manual experience-based parameter tuning.

[0109] In an optional embodiment, during the real-time data interaction and synchronization processing of the initial training state data information, the data interaction layer may employ a transmission scheme that integrates UDP and TCP dual protocols for data transmission. It should be noted that the aforementioned dual-protocol fusion transmission scheme refers to using UDP protocol for data with high real-time requirements to ensure low latency, and using TCP protocol for data with high reliability requirements to ensure data integrity. Specifically, real-time status data such as position data, velocity data, and attitude angle data are transmitted using UDP protocol, with an end-to-end latency of no more than 100ms; data with high reliability requirements such as control command data, configuration parameter data, and evaluation result data are transmitted using TCP protocol.

[0110] It should be noted that the data interaction layer also includes data priority scheduling processing, which sets different transmission priorities for different types of data. The transmission priority of status data is higher than that of configuration data, ensuring the priority transmission of real-time status data.

[0111] As can be seen, the virtual-real joint simulation training method for unmanned intelligent systems described in this embodiment of the invention, through the fusion of UDP and TCP dual protocols for transmission and data priority scheduling, takes into account both the real-time performance and reliability of data transmission, ensuring the high efficiency and stability of data interaction in virtual-real joint training.

[0112] Example 2 Please see Figure 3 , Figure 3 This is a schematic diagram of the structure of a virtual-real joint simulation training device for an unmanned intelligent system disclosed in an embodiment of the present invention. Figure 3 The described virtual-real co-simulation training device for unmanned intelligent systems is applied in unmanned intelligent system simulation training systems, such as local servers or cloud servers used for unmanned intelligent system training, etc., and the embodiments of the present invention are not limited thereto. Figure 3 As shown, the virtual-real joint simulation training device for the unmanned intelligent system includes: The acquisition module 201 is used to acquire physical entity device configuration data information set, training task configuration data information and training parameter configuration data information; The modeling and scene generation module 202 is used to perform modeling and scene generation processing on the physical entity equipment configuration data information set, the training task configuration data information and the training parameter configuration data information to obtain the digital equipment model data information set and the virtual training scene data information. The virtual-real joint training module 203 is used to perform virtual-real joint training processing on the digital equipment model data information set and the virtual training scene data information to obtain a training process data information set. The evaluation and closed-loop optimization module 204 is used to evaluate and perform closed-loop optimization on the training process data information set and the training parameter configuration data information to obtain target training evaluation report information and optimized training parameter configuration data information.

[0113] As can be seen, the virtual-real joint simulation training device for unmanned intelligent systems described in this embodiment of the invention acquires physical entity equipment configuration, training task configuration, and training parameter configuration data through the acquisition module, constructs digital equipment models and generates virtual training scenarios through the modeling and scene generation module, realizes collaborative training between the virtual environment and physical entities through the virtual-real joint training module, and realizes multi-dimensional quantitative evaluation of training effects and automatic optimization of training parameters through the evaluation and closed-loop optimization module. This solves the technical problems of low virtual-real synchronization accuracy, poor scene adaptability, single evaluation dimension, and insufficient cross-device compatibility in existing virtual-real joint simulation training.

[0114] Example 3 Please see Figure 4 , Figure 4 This is a schematic diagram of another unmanned intelligent system virtual-real joint simulation training device disclosed in an embodiment of the present invention. Figure 4 The described virtual-real co-simulation training device for unmanned intelligent systems is applied in unmanned intelligent system simulation training systems, such as local servers or cloud servers used for unmanned intelligent system training, etc., and the embodiments of the present invention are not limited thereto. Figure 4 As shown, the virtual-real joint simulation training device for the unmanned intelligent system includes: Processor 301; A memory 302 containing executable program code is coupled to the processor 301; The processor 301 calls the executable program code stored in the memory 302 to execute some or all of the steps of the virtual-real joint simulation training method for an unmanned intelligent system according to Embodiment 1.

[0115] As can be seen, the virtual-real joint simulation training device for unmanned intelligent systems described in the embodiments of the present invention realizes the automated execution and management of the training process by executing the virtual-real joint simulation training method by calling the program code in the memory through the processor.

[0116] Example 4 This invention also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements some or all of the steps of the virtual-real joint simulation training method for an unmanned intelligent system described in Embodiment 1.

[0117] Example 5 This invention also provides a computer program product that, when run on a computer, causes the computer to execute some or all of the steps of the virtual-real joint simulation training method for an unmanned intelligent system described in Embodiment 1.

[0118] The system embodiments described above are merely illustrative. The modules described as separate components may or may not be physically separate, and the components shown as modules may or may not be physical modules; that is, they may be located in one place or distributed across multiple network modules. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0119] Through the detailed description of the above embodiments, those skilled in the art can clearly understand that each implementation method can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, including read-only memory (ROM), random access memory (RAM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electronically erasable rewritable read-only memory (EEPROM), read-only optical disc (CD-ROM) or other optical disc storage, disk storage, magnetic tape storage, or any other computer-readable medium that can be used to carry or store data.

[0120] Finally, it should be noted that the virtual-real joint simulation training method and apparatus for unmanned intelligent systems disclosed in the embodiments of the present invention are merely preferred embodiments of the present invention, and are only used to illustrate the technical solutions of the present invention, not to limit it; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A virtual-real joint simulation training method for an unmanned intelligent system, characterized in that, The method includes: S1, obtain the physical entity device configuration data set, training task configuration data, and training parameter configuration data; S2, Modeling and scene generation processing are performed on the physical entity equipment configuration data information set, the training task configuration data information and the training parameter configuration data information to obtain the digital equipment model data information set and the virtual training scene data information; S3, perform virtual-real joint training processing on the digital equipment model data information set and the virtual training scene data information to obtain the training process data information set; S4, evaluate and perform closed-loop optimization on the training process data information set and the training parameter configuration data information to obtain target training evaluation report information and optimized training parameter configuration data information.

2. The virtual-real joint simulation training method for unmanned intelligent systems according to claim 1, characterized in that, The process of modeling and generating scenarios using the physical entity equipment configuration data set, the training task configuration data information, and the training parameter configuration data information to obtain a digital equipment model data set and virtual training scenario data information includes: S21, Perform digital equipment model construction processing on the physical entity equipment configuration data information set to obtain a digital equipment model data information set; the digital equipment model data information set includes several digital equipment model data information sets. S22, perform virtual training scene generation processing on the training task configuration data information and the training parameter configuration data information to obtain virtual training scene data information.

3. The virtual-real joint simulation training method for unmanned intelligent systems according to claim 2, characterized in that, The process of constructing a digital equipment model from the physical entity equipment configuration data set to obtain a digital equipment model data set includes: S211, Perform geometric mapping processing on the device geometric parameter data information in any of the physical entity device configuration data information in the physical entity device configuration data information set to obtain the geometric mapping model data information of the physical entity device configuration data information; S212, Perform physical mapping processing on the sensor parameter data information and actuator parameter data information in the configuration data information of the physical entity device to obtain the physical mapping model data information of the configuration data information of the physical entity device. S213, Perform control mapping processing on the device type information in the configuration data information of the physical entity device to obtain the control mapping model data information of the configuration data information of the physical entity device. S214, integrate the geometric mapping model data information, the physical mapping model data information, and the control mapping model data information of the physical entity equipment configuration data information to obtain the digital equipment model data information corresponding to the physical entity equipment configuration data information.

4. The virtual-real joint simulation training method for unmanned intelligent systems according to claim 2, characterized in that, The virtual training scenario generation process, which involves processing the training task configuration data information and the training parameter configuration data information to obtain virtual training scenario data information, includes: S221, Perform environmental parameter generation processing on the scene type information and initial difficulty level information in the training parameter configuration data information to obtain environmental parameter data information; S222, Perform task parameter generation processing on the operation level information in the training task configuration data information and the training parameter configuration data information to obtain task parameter data information; S223, perform interference parameter generation processing on the initial difficulty level information and historical training data information in the training parameter configuration data information to obtain interference parameter data information; S224, using a dynamic scene difficulty adaptive calculation model, calculates and processes the environmental parameter data, the task parameter data, and the interference parameter data to obtain the comprehensive scene difficulty value; The dynamic scene difficulty adaptive calculation model is as follows: ; ; In the formula, The overall difficulty value of the scenario. This represents the normalized difficulty value of the p-th environmental sub-parameter in the environmental parameter data, where P is the number of environmental sub-parameters. This represents the normalized difficulty value of the q-th task sub-parameter in the task parameter data, where Q is the number of task sub-parameters. R is the normalized difficulty value of the r-th interference sub-parameter in the interference parameter data information, where R is the number of interference sub-parameters. , and These are the weighting coefficients for environmental parameters, task parameters, and disturbance parameters, respectively. To adapt and adjust parameters based on historical training, The historical training difficulty deviation value in the historical training data information is used to adjust the scene difficulty based on historical training performance. S225, Perform scene rationality verification processing on the comprehensive difficulty value of the scene to obtain virtual training scene data information.

5. The virtual-real joint simulation training method for unmanned intelligent systems according to claim 1, characterized in that, The evaluation and closed-loop optimization of the training process data set and the training parameter configuration data to obtain the target training evaluation report and optimized training parameter configuration data includes: S41, perform data preprocessing on the training process data information set to obtain a preprocessed training process data information set; S42, perform multi-dimensional index calculation processing on the preprocessed training process data information set to obtain a multi-dimensional evaluation index data information set; S43, perform comprehensive evaluation and calculation on the multi-dimensional evaluation index data information set and the training parameter configuration data information to obtain the target training evaluation report information; S44, perform closed-loop optimization processing on the target training evaluation report information and the training parameter configuration data information to obtain optimized training parameter configuration data information.

6. The virtual-real joint simulation training method for unmanned intelligent systems according to claim 5, characterized in that, The process of performing multi-dimensional index calculation on the preprocessed training process data information set to obtain a multi-dimensional evaluation index data information set includes: S421, Process the device status time-series data information in the preprocessed training process data information set to obtain device status index data information; S422, Process the timing data of the operation instructions in the preprocessed training process data information set to obtain operation behavior index data information; S423, Process the task execution result data information in the preprocessed training process data information set to obtain task performance index data information; S424, Process the virtual-real synchronization deviation data information and the operation instruction timing data information in the preprocessed training process data information set to obtain the cooperative index data information; S425, integrate and process the equipment status index data information, the operation behavior index data information, the task effect index data information, and the cooperation index data information to obtain a multi-dimensional evaluation index data information set.

7. The virtual-real joint simulation training method for unmanned intelligent systems according to claim 5, characterized in that, The process of comprehensively evaluating and calculating the multi-dimensional evaluation index data set and the training parameter configuration data to obtain the target training evaluation report information includes: S431, The multi-dimensional evaluation index data information set is subjected to index normalization processing to obtain a normalized evaluation index data information set; S432, using the hierarchical analysis calculation model, calculate and process the normalized evaluation index data information set and the evaluation index weight data information in the training parameter configuration data information to obtain the comprehensive training score value and the score data information set of each dimension. The hierarchical analysis calculation model is as follows: ; ; ; ; In the formula, The comprehensive training score is... Let be the dimension score of the i-th dimension, where i ranges from 1 to 4, corresponding to the device status dimension, operation behavior dimension, task effect dimension, and collaboration dimension, respectively. Let be the dimension weight coefficient of the i-th dimension; Let i be the number of evaluation metrics in the i-th dimension. The index weight coefficient of the j-th evaluation index in the i-th dimension; The normalized value of the j-th evaluation indicator in the i-th dimension of the normalized evaluation indicator data set; Let be the fuzzy membership value of the j-th evaluation indicator in the i-th dimension, representing the degree of membership of the evaluation level corresponding to the indicator value; is the kurtosis parameter of the fuzzy membership function. Let be the fuzzy membership center parameter for the i-th dimension and the j-th evaluation index; S433, integrate the comprehensive training score and the score data information set of each dimension to obtain the target training evaluation report information.

8. A virtual-real joint simulation training device for an unmanned intelligent system, characterized in that, The device includes: The acquisition module is used to acquire physical entity device configuration data information set, training task configuration data information, and training parameter configuration data information; The modeling and scene generation module is used to model and generate scenes from the physical entity equipment configuration data information set, the training task configuration data information, and the training parameter configuration data information to obtain a digital equipment model data information set and a virtual training scene data information. The virtual-real joint training module is used to perform virtual-real joint training processing on the digital equipment model data information set and the virtual training scene data information to obtain the training process data information set. The evaluation and closed-loop optimization module is used to evaluate and perform closed-loop optimization on the training process data information set and the training parameter configuration data information to obtain target training evaluation report information and optimized training parameter configuration data information.

9. A virtual-real joint simulation training device for an unmanned intelligent system, characterized in that, The device includes: processor; A memory containing executable program code is coupled to the processor; The processor calls the executable program code stored in the memory to execute the virtual-real joint simulation training method for unmanned intelligent systems as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions, which, when invoked, are used to execute the virtual-real joint simulation training method for unmanned intelligent systems as described in any one of claims 1-7.

Images