A joint system simulation verification method and apparatus based on multi-source data injection

Abstract

This invention provides a method and apparatus for joint system simulation verification based on multi-source data injection. Based on a preset task objective, the invention constructs a virtual training environment model in a simulation platform, including geographical environment, force units, task scenarios, and dynamic interference factors. Realistic simulation data is dynamically generated based on the virtual training environment model. False feature codes and dynamic intensity noise signals are generated, purified, and then injected into the real system under test to generate corresponding control commands. These control commands are fed back to the simulation platform, driving the entity state update in the virtual environment and evaluating performance indicators through a digital twin module. This invention can accurately reproduce the joint system operation state under complex real-world scenarios, comprehensively evaluate system functionality, real-time performance, anti-interference capabilities, and decision optimization, significantly reduce verification costs, and provide technical support for the stability and reliability of multi-system joint operation.

Description

Technical Field

[0001] This invention relates to the field of big data processing technology, and in particular to a joint system simulation verification method and apparatus based on multi-source data injection. Background Technology

[0002] In recent years, with the rapid development of various technologies, the scale of projects has been expanding. A large-scale project often includes three or more systems, and how to verify the stability and accuracy of the joint system operation has become a new challenge.

[0003] Existing technologies primarily employ digital simulation or hardware-in-the-loop simulation. However, traditional digital simulation relies on simplified system models and fixed scenario parameters, failing to realistically reproduce system interaction logic under complex environments. Insufficient model accuracy leads to low reliability of verification results, making it difficult to support practical application decisions. Hardware-in-the-loop simulation, on the other hand, requires the construction of complex physical interfaces and dedicated signal simulation equipment. This not only results in high upfront costs and maintenance difficulties but also poor scenario adaptability, making it unable to flexibly address the verification needs of different types and scales of joint systems. Furthermore, it leads to long verification cycles and significant human resource consumption.

[0004] Existing simulation verification methods generally lack the ability to simulate complex interference environments, and their anti-interference verification dimensions are singular, making it difficult to match the complex situations such as electromagnetic interference and data forgery in real application scenarios. Moreover, the effect evaluation is mostly focused on functional verification, lacking a comprehensive evaluation of the system's real-time performance, decision-making economy, and optimization potential, and thus failing to provide comprehensive data support for the iterative optimization of joint systems. Furthermore, joint systems in many key areas (such as military operational planning systems and aerospace information fusion systems) are limited by the safety, cost, and technical conditions of practical scenarios, and cannot be fully verified through actual flight tests, field exercises, etc., making it difficult to discover potential system problems in advance.

[0005] Therefore, there is an urgent need to provide a joint system simulation verification technology that combines high fidelity, low verification cost, strong scenario adaptability, and full-dimensional evaluation capabilities. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention provides a joint system simulation verification method and apparatus based on multi-source data injection, thereby resolving the problems existing in the background technology.

[0007] To achieve the above objectives, the present invention adopts the following technical solution: In a first aspect, the present invention provides a joint system simulation verification method based on multi-source data injection, comprising the following steps: S1) Based on the preset mission objectives, a virtual training environment model is constructed in the simulation platform, including the geographical environment, troop units, mission scenarios, and dynamic interference factors. S2) Dynamically generate realistic simulation data based on the virtual training environment model described above; S3) The generator of the deep convolutional generative adversarial network DCGAN generates fake feature codes with real feature distribution to simulate data forgery attacks. S4) Generate a basic noise signal according to the scenario requirements and dynamically adjust the intensity and frequency of the noise signal to obtain a dynamic intensity noise signal to simulate a complex electromagnetic interference environment. S5) Mix real simulated data, spurious feature codes, dynamic intensity noise signals, and original acquired data to form multi-source heterogeneous simulated data; S6) Use a Transformer-based deep learning classification model to clean up the multi-source heterogeneous simulation data to obtain cleaned multi-source heterogeneous simulation data; and inject the cleaned multi-source heterogeneous simulation data into the real system under test. S7) The system under test fuses the injected purified multi-source heterogeneous simulation data and generates corresponding control commands. S8) Feed the control commands back to the simulation platform, drive the entity state update in the virtual environment through the digital twin module, and evaluate the performance indicators of the functionality, real-time performance, anti-interference and decision optimization of each tested system based on the updated state.

[0008] Preferably, in step S1), the task scenario includes the task flow, target parameters, and constraints; the dynamic interference factor is generated through a stochastic process model and includes random electromagnetic interference, sudden failure scenarios, and sudden changes in environmental parameters.

[0009] Preferably, in step S2), the real simulation data includes motion data, perception data, attribute data, pattern data, and environmental interference signals; wherein, the motion data includes target position, target velocity, and target acceleration. The sensing data includes radar echoes, communication signals, and sensor detection signals; The attribute data includes platform type, identification code, and performance parameters; The aforementioned pattern data includes movement trajectory and operation mode.

[0010] Preferably, in step S6), a Transformer-based deep learning classification model is used to perform deep feature matching between the multi-source heterogeneous simulation data and a preset target feature database. Key feature extraction is enhanced through an attention mechanism, and a softmax classifier is used to determine the authenticity and correctness of the data. This automatically identifies and filters false, erroneous, and interfering data, outputting high-fidelity, clean multi-source heterogeneous simulation data. Specifically: S61) The multi-source heterogeneous simulation data is mapped into a high-dimensional feature vector through the embedding layer, and positional encoding is injected to obtain the embedded features; S62) Based on the embedded features, they are mapped to query Q, key K, and value V matrices respectively, and split into multiple attention heads for parallel computation. The weight distribution of each attention head is calculated by scaling the dot product attention formula to strengthen the weight of key features. S63) The attention output is superimposed on the original embedded features through residual connections, and combined with layer normalization to stabilize the training process, so as to enhance the key features and preserve the original features. S64) The enhanced key features are nonlinearly transformed by a two-layer fully connected network to obtain the feedforward network mapped features; S65) The feedforward network is mapped to the predicted probability of real and valid data, fake data, erroneous or interfering data through the classification layer, and the probability distribution is normalized by the softmax function. S66) The difference between the predicted probability and the true label is calculated based on the cross-entropy loss function, and the model parameters are updated through the AdamW optimizer, with weight decay added to suppress overfitting; S67) Set a confidence threshold, retain heterogeneous simulated data that are predicted to be true and valid data and whose probability is greater than the confidence threshold, filter out false data, erroneous data and interference data, and output high-fidelity clean data.

[0011] Preferably, in step S7), the tested system uses Kalman filtering to fuse the injected purified multi-source heterogeneous simulation data, and realizes inter-system collaborative interaction and generates control commands through a distributed collaborative decision-making algorithm.

[0012] Preferably, in step S8), the motion trajectory and attribute parameters of entities in the virtual environment are corrected through the digital twin module, specifically including the following steps: S81) Based on the Bayesian optimization algorithm, a dynamic motion trajectory containing key path points is generated. After the initial position data is injected, the simulation platform automatically turns and moves according to the predetermined motion program. During the simulation, the platform's speed, height, and attitude attributes are dynamically adjusted through an adaptive PID control algorithm according to real-time environmental parameters and task requirements. (S82) If the target is lost, the digital twin module predicts and generates the target's continuous motion trajectory based on the target's motion characteristics before the loss, using the extended Kalman filter algorithm, to ensure the continuity and consistency of the simulation process.

[0013] Preferably, in step S8), the real-time performance evaluation is as follows: by statistically analyzing the response time, instruction transmission delay, and status update cycle of each tested system after receiving data, weights are assigned according to the degree of influence of each indicator on real-time performance, and a comprehensive real-time performance score is calculated; at the same time, the comprehensive real-time performance score is used to evaluate real-time performance.

[0014] Preferably, in step S8), the anti-interference assessment is as follows: The accuracy, recall, and F1 score of data identification are calculated using a confusion matrix to comprehensively quantify the matching accuracy between purified effective simulated data and real-world data. At the same time, an interference intensity attenuation coefficient is introduced to construct an anti-interference performance evaluation model.

[0015] Preferably, in step S8), the functional evaluation is as follows: The weights of task completion, target identification accuracy, instruction execution success rate, and system coordination efficiency indicators were determined using the analytic hierarchy process (AHP). A functional evaluation matrix was constructed, and the comprehensive functional score was calculated by weighted summation.

[0016] Preferably, in step S8), the decision optimization evaluation is as follows: Calculate the cost-benefit ratio of the decision based on resource consumption and task benefits, assess the feasibility and economy of the decision, and output optimization suggestions based on the assessment structure.

[0017] Secondly, the present invention provides a joint system simulation verification device based on multi-source data injection, comprising: The simulation environment construction module, based on preset task objectives, constructs a virtual training environment model in the simulation platform, including geographical environment, troop units, task scenarios, and dynamic interference factors. The heterogeneous simulation data generation module dynamically generates heterogeneous simulation data based on the aforementioned virtual training environment model; The interference information generation module is used to generate false feature codes and dynamic intensity interference signals; The preprocessing module uses a Transformer-based deep learning classification model to clean up multi-source heterogeneous simulation data; and then injects the cleaned multi-source heterogeneous simulation data into the real system under test to drive the system under test to run. The control command generation module is used to fuse the injected purified multi-source heterogeneous simulation data and generate corresponding control commands. The simulation and evaluation module uses control commands to drive the update of entity states in the virtual environment, and evaluates the performance indicators of each tested system in terms of functionality, real-time performance, anti-interference ability, and decision optimization based on the updated states.

[0018] The beneficial effects of this invention are as follows: 1. This invention includes a virtual training environment for geographical environment, force units, mission scenarios and dynamic interference factors. Combined with confrontation training to simulate complex interference scenarios, it can restore the real operating state of the joint system without actual flight or field exercise. It solves the technical problem that joint systems in key areas are difficult to verify in practice. It is applicable to the verification needs of joint systems in multiple fields such as military and aerospace. 2. This invention comprehensively evaluates simulation performance through real-time performance, anti-interference ability, functionality, and decision optimization, and verifies that the results are more consistent with real-world application scenarios; 3. The digital twin system of the present invention dynamically adjusts the platform trajectory and target attributes through Bayesian optimization and EKF algorithm, solves the simulation continuity problem in the target loss scenario, and provides support for the stability of multi-system joint operation; 4. The simulation data of this invention covers multiple dimensions such as target location, radar echo, communication signal, and platform attributes. Combined with false feature coding and dynamic intensity interference signal, it restores the complex real operating environment. 5. This invention uses the Transformer deep learning model to clean up data, filtering out false, erroneous, and interfering data, ensuring the high fidelity of the injected data, and improving the credibility of the verification results. Attached Figure Description

[0019] Figure 1 This is a flowchart illustrating the method of Embodiment 1 of the present invention; Figure 2 This is a schematic diagram of the system framework of Embodiment 2 of the present invention. Detailed Implementation

[0020] The specific embodiments of the present invention will be further described below with reference to the accompanying drawings: Example 1 like Figure 1 As shown, this embodiment provides a joint system simulation verification method based on multi-source data injection, including the following steps: S1) Based on the preset mission objectives, a virtual training environment model is constructed in the simulation platform, including geographical environment, troop units, mission scenarios and dynamic interference factors. In this embodiment, the preset mission objective can be reconnaissance, identification, strike, or collaborative operation; In this embodiment, the geographical environment includes topography, weather, electromagnetic environment, etc. In this embodiment, the task scenario includes the task flow, target parameters, and constraints. The dynamic interference factor is generated through a stochastic process model, including random electromagnetic interference, sudden failure scenarios, and sudden changes in environmental parameters.

[0021] S2) Dynamically generate heterogeneous simulation data based on the virtual training environment model described above; In this embodiment, the heterogeneous simulation data includes motion data, sensing data, attribute data, pattern data, and environmental interference signals; wherein: The motion data includes target position, target velocity, and target acceleration; The sensing data includes radar echoes, communication signals, and sensor detection signals; The attribute data includes platform type, identification code, and performance parameters; The aforementioned pattern data includes movement trajectory and operation mode; The environmental interference signals mentioned include electromagnetic interference, noise interference, etc.

[0022] S3) The generator of the deep convolutional generative adversarial network DCGAN generates fake feature codes with real feature distribution to simulate data forgery attacks. In this embodiment, the generator of the Deep Convolutional Generative Adversarial Network (DCGAN) generates fake feature codes with a true feature distribution. The representation is: In the formula, This is a generator mapping function; The mean value encoded as the true feature; The standard deviation of the true feature encoding; for 3D random noise vector; The generator mapping function is expressed as follows: In the formula, The LeakyReLU function is used as the activation function. This is a transpose convolution operation; The weight matrix for each layer; - These are the bias vectors for each layer; This represents the number of layers in the generator network.

[0023] in, Given a normal distribution with a mean of 0 and a covariance matrix of identity matrix I, this formula is used to generate initial random noise that conforms to statistical characteristics.

[0024] In this embodiment, the loss function of the Deep Convolutional Generative Adversarial Network (DCGAN) is... for: in, This represents the expectation operation; To predict the probability of the discriminator's assessment of the authenticity of the generated data, the loss function is minimized using the Adam optimizer. Update the weights and biases of each layer of the generator.

[0025] S4) Generate a basic noise signal according to the scenario requirements and dynamically adjust the intensity and frequency of the noise signal to obtain a dynamic intensity noise signal to simulate a complex electromagnetic interference environment. In this embodiment, the basic Gaussian noise signal The generation is as follows: In the formula, The mean is The covariance matrix is The normal distribution; The intensity and frequency of the noise signal are dynamically adjusted, specifically as follows: ; ; In the formula, for The dynamically adjusted noise signal that is ultimately output at any given moment; The basic Gaussian noise signal; This is the time-related intensity adjustment coefficient; for A frequency-modulated continuous wave (FM) signal at a given time; for The intensity of the noise signal at any given moment; For reference noise intensity.

[0026] S5) Mix real simulated data, spurious feature codes, dynamic intensity noise signals, and original acquired data to form multi-source heterogeneous simulated data; S6) Use a Transformer-based deep learning classification model to clean up the multi-source heterogeneous simulation data to obtain cleaned multi-source heterogeneous simulation data; and inject the cleaned multi-source heterogeneous simulation data into the real system under test. In this embodiment, a Transformer-based deep learning classification model is used to perform deep feature matching between multi-source heterogeneous simulation data and a pre-defined target feature database. An attention mechanism is used to enhance key feature extraction, and a softmax classifier is employed to determine the authenticity and correctness of the data. This automatically identifies and filters false, erroneous, and interfering data, outputting high-fidelity, clean heterogeneous simulation data. Specifically: S61) Multi-source heterogeneous simulation data is embedded through an embedding layer. The mapping is performed on the high-dimensional feature vector, and positional encoding is injected to obtain the embedded features. ,Right now: In the formula, For embedding layer mapping functions; This is a positional encoding function used to inject temporal features into the data; Multi-source heterogeneous simulation data Quantity; S62), based on embedded features Mapped to queries respectively ,key ,value The matrix is ​​split into multiple attention heads for parallel computation. The weight distribution of each attention head is calculated by scaling the dot product attention formula to strengthen the weight of key features. In the formula, This indicates the attention output and attention fusion features; Indicates a fusion operation; Represents the global weight moments of the multi-head attention mechanism; For the first The feature vector output by each head; For the number of heads; , , These are the weights of the query, key, and value matrices, respectively. This indicates scaling dot product attention; They represent the first The query, key, and value matrix of each attention head; For single-head attention feature dimensions; S63) By superimposing the attention output with the original embedded features through residual connections, and combining this with layer normalization to stabilize the training process, key features are enhanced while the original features are preserved; that is: In the formula, These are the enhanced key features; Presentation layer normalization operation; S24) By using a two-layer fully connected network to perform nonlinear transformations on the enhanced key features, the model's ability to fit complex features is improved; that is: In the formula, , Here is the weight matrix of the feedforward network; , For the feedforward network bias; Features are mapped to feedforward networks; S65) The processed features are mapped to the predicted probabilities of real and valid data, fake data, and erroneous or interfering data through the classification layer, and the probability distribution is normalized by the softmax function. In the formula, To predict the probability distribution; , These are the classification layer weight matrix and the bias, respectively. S66) The difference between the predicted probability and the true label is calculated based on the cross-entropy loss function, and the model parameters are updated through the AdamW optimizer, with weight decay added to suppress overfitting; that is: In the formula, The cross-entropy loss function; For heterogeneous simulation data Quantity; The number of categories; corresponding to real and valid data, fake data, and erroneous or interfering data, respectively. One-hot encoding of the true labels; minimizing the cross-entropy loss function using the AdamW optimizer. ; S67) Set confidence threshold Predictions that are considered true and valid data with a probability greater than the confidence threshold are retained. Heterogeneous simulation data is used to filter out false, erroneous, and interfering data, and output high-fidelity clean data. ,Right now: In the formula, Let this be an indicator function, representing the probability of a prediction being true and valid data greater than [a certain value]. The data should be retained if it is valid, otherwise it should be filtered.

[0027] S7) The system under test fuses the injected purified multi-source heterogeneous simulation data and generates corresponding control commands. In this embodiment, the system under test employs Kalman filtering to fuse the injected, purified multi-source heterogeneous simulation data. A distributed collaborative decision-making algorithm is used to achieve inter-system collaborative interaction and generate initial control commands, specifically including the following steps: S71) Set each tested system participating in the collaboration as a distributed decision node. , Configure the number of nodes; configure each node with a local decision-making model, data interaction interface, and task priority weights. ; S72), each node is based on local data after multi-source data fusion. A candidate set of local control commands is generated through a local decision-making model. ,in, Given the number of instructions; and calculate the local utility value for each candidate instruction. ,Right now: ; In the formula, For utility evaluation functions; For task target parameters; S73) Information synchronization between nodes is achieved through the Real-time Data Transmission Protocol (RTPS), and each node broadcasts its local candidate instruction set to other nodes. Local utility value and its own task priority weight Simultaneously, it receives broadcast information from other nodes, forming a global information set. ,Right now: ; S74) Each node calculates the global utility value of each candidate instruction based on the global information set using a weighted summation algorithm. Filter out the candidate instructions with the highest global utility value ,Right now: S75) If there are multiple candidate instructions whose global utility values ​​differ by less than a threshold To trigger conflict negotiation, the optimal instruction after final negotiation is determined using the Nash bargaining model in game theory. The Nash bargaining model is expressed as follows: ; In the formula, The retention utility value for each node; S76) The globally optimal instruction or the negotiated instruction is used as the initial control instruction and issued to the corresponding execution unit.

[0028] S8) Feed the control commands back to the simulation platform, drive the entity state update in the virtual environment through the digital twin module, and evaluate the performance indicators of the functionality, real-time performance, anti-interference performance and decision optimization of each tested system based on the updated state. In this embodiment, the digital twin module drives the update of entity states in the virtual environment, specifically including the following steps: S81) Based on the Bayesian optimization algorithm, a dynamic motion trajectory containing key path points is generated. After the initial position data is injected, the simulation platform automatically turns and moves according to the predetermined motion program. During the simulation process, the speed, height, and attitude attributes of the simulation platform are dynamically adjusted through the adaptive PID control algorithm according to the real-time environmental parameters and task requirements. (S82) If the target is lost, the digital twin module predicts and generates the target's continuous motion trajectory based on the target's motion characteristics before the loss, using the extended Kalman filter algorithm, to ensure the continuity and consistency of the simulation process.

[0029] In this embodiment, the real-time performance evaluation is as follows: by statistically analyzing the response time, instruction transmission delay, and status update cycle of each tested system after receiving data, weights are assigned according to the degree of influence of each indicator on real-time performance, and a comprehensive real-time performance score is calculated; at the same time, the comprehensive real-time performance score is used to evaluate real-time performance.

[0030] The anti-interference assessment is as follows: the precision, recall, and F1 score of data identification are calculated using a confusion matrix to comprehensively quantify the matching accuracy between the purified effective simulated data and real scene data. At the same time, an interference intensity attenuation coefficient is introduced to construct an anti-interference performance assessment model.

[0031] The aforementioned functional evaluation is as follows: the weights of the indicators of task completion, target recognition accuracy, instruction execution success rate, and system coordination efficiency are determined by using the analytic hierarchy process (AHP), a functional evaluation matrix is ​​constructed, and the comprehensive functional score is calculated by weighted summation.

[0032] The aforementioned decision optimization assessment involves calculating the cost-benefit ratio of a decision based on resource consumption and task benefits, evaluating the feasibility and economy of the decision, and outputting optimization suggestions based on the assessment structure.

[0033] Example 2 like Figure 2 As shown, this embodiment provides a joint system simulation verification device based on multi-source data injection, including: The simulation environment construction module, based on preset task objectives, constructs a virtual training environment model in the simulation platform, including geographical environment, troop units, task scenarios, and dynamic interference factors. The heterogeneous simulation data generation module dynamically generates heterogeneous simulation data based on the aforementioned virtual training environment model; The interference information generation module is used to generate false feature codes and dynamic intensity interference signals; The preprocessing module uses a Transformer-based deep learning classification model to clean up multi-source heterogeneous simulation data; and then injects the cleaned multi-source heterogeneous simulation data into the real system under test to drive the system under test to run. The control command generation module is used to fuse the injected purified multi-source heterogeneous simulation data and generate corresponding control commands. The simulation and evaluation module uses control commands to drive the update of entity states in the virtual environment, and evaluates the performance indicators of each tested system in terms of functionality, real-time performance, anti-interference ability, and decision optimization based on the updated states.

[0034] The embodiments and descriptions above are merely illustrative of the principles and preferred embodiments of the present invention. Various changes and modifications may be made to the present invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed.

Claims

1. A joint system simulation verification method based on multi-source data injection, characterized in that, Includes the following steps: S1) Based on the preset mission objectives, a virtual training environment model is constructed in the simulation platform, including the geographical environment, troop units, mission scenarios, and dynamic interference factors. S2) Dynamically generate realistic simulation data based on the virtual training environment model described above; S3) The generator of the deep convolutional generative adversarial network DCGAN generates fake feature codes with real feature distribution to simulate data forgery attacks. S4) Generate a basic noise signal according to the scenario requirements and dynamically adjust the intensity and frequency of the noise signal to obtain a dynamic intensity noise signal to simulate a complex electromagnetic interference environment. S5) Mix real simulated data, spurious feature codes, dynamic intensity noise signals, and original acquired data to form multi-source heterogeneous simulated data; S6) Use a Transformer-based deep learning classification model to clean up the multi-source heterogeneous simulation data to obtain cleaned multi-source heterogeneous simulation data; and inject the cleaned multi-source heterogeneous simulation data into the real system under test. S7) The system under test fuses the injected purified multi-source heterogeneous simulation data and generates corresponding control commands. S8) Feed the control commands back to the simulation platform, drive the entity state update in the virtual environment through the digital twin module, and evaluate the performance indicators of the functionality, real-time performance, anti-interference and decision optimization of each tested system based on the updated state.

2. The joint system simulation verification method based on multi-source data injection according to claim 1, characterized in that: In step S2), the real simulation data includes motion data, perception data, attribute data, pattern data, and environmental interference signals; wherein, the motion data includes target position, target velocity, and target acceleration. The sensing data includes radar echoes, communication signals, and sensor detection signals; The attribute data includes platform type, identification code, and performance parameters; The aforementioned pattern data includes movement trajectory and operation mode.

3. The joint system simulation verification method based on multi-source data injection according to claim 2, characterized in that: In step S3), spurious feature encoding with a true feature distribution is used. The representation is: In the formula, This is a generator mapping function; The mean value encoded as the true feature; The standard deviation of the true feature encoding; for A dimensional random noise vector.

4. The joint system simulation verification method based on multi-source data injection according to claim 3, characterized in that: In step S3), the loss function of the deep convolutional generative adversarial network DCGAN for: in, This represents the expectation operation; To predict the probability of the discriminator's assessment of the authenticity of the generated data, the loss function is minimized using the Adam optimizer. Update the weights and biases of each layer of the generator.

5. The joint system simulation verification method based on multi-source data injection according to claim 4, characterized in that: In step S4), the fundamental Gaussian noise signal The generation is as follows: In the formula, The mean is The covariance matrix is The normal distribution; The intensity and frequency of the noise signal are dynamically adjusted, specifically as follows: ； ； In the formula, for The dynamically adjusted noise signal that is ultimately output at any given moment; The basic Gaussian noise signal; This is the time-related intensity adjustment coefficient; for A frequency-modulated continuous wave (FM) signal at a given time; for The intensity of the noise signal at any given moment; For reference noise intensity.

6. The joint system simulation verification method based on multi-source data injection according to claim 5, characterized in that: In step S6), a Transformer-based deep learning classification model is used to perform deep feature matching between multi-source heterogeneous simulation data and a preset target feature database. Key feature extraction is enhanced through an attention mechanism. A softmax classifier is used to determine the authenticity and correctness of the data, automatically identifying and filtering false data, erroneous data and interference data, and outputting high-fidelity and clean multi-source heterogeneous simulation data.

7. The joint system simulation verification method based on multi-source data injection according to claim 6, characterized in that: Step S6) specifically includes the following steps: S61) The multi-source heterogeneous simulation data is mapped into a high-dimensional feature vector through the embedding layer, and positional encoding is injected to obtain the embedded features; S62) Based on the embedded features, they are mapped to query Q, key K, and value V matrices respectively, and split into multiple attention heads for parallel computation. The weight distribution of each attention head is calculated by scaling the dot product attention formula to strengthen the weight of key features. S63) The attention output is superimposed on the original embedded features through residual connections, and combined with layer normalization to stabilize the training process, so as to enhance the key features and preserve the original features. S64) The enhanced key features are nonlinearly transformed by a two-layer fully connected network to obtain the feedforward network mapped features; S65) The feedforward network is mapped to the predicted probability of real and valid data, fake data, erroneous or interfering data through the classification layer, and the probability distribution is normalized by the softmax function. S66) The difference between the predicted probability and the true label is calculated based on the cross-entropy loss function, and the model parameters are updated through the AdamW optimizer, with weight decay added to suppress overfitting; S67) Set a confidence threshold, retain heterogeneous simulated data that are predicted to be true and valid data and whose probability is greater than the confidence threshold, filter out false data, erroneous data and interference data, and output high-fidelity clean data.

8. The joint system simulation verification method based on multi-source data injection according to claim 7, characterized in that: In step S7), the distributed collaborative decision-making algorithm is used to realize collaborative interaction between systems and generate control commands, specifically including the following steps: S71) Set each tested system participating in the collaboration as a distributed decision node. , Configure the number of nodes; configure each node with a local decision-making model, data interaction interface, and task priority weights. ; S72), each node is based on local data after multi-source data fusion. A candidate set of local control commands is generated through a local decision-making model. ,in, Given the number of instructions; and calculate the local utility value for each candidate instruction. ,Right now: ； In the formula, For utility evaluation functions; For task target parameters; S73) Information synchronization between nodes is achieved through the Real-time Data Transmission Protocol (RTPS), and each node broadcasts its local candidate instruction set to other nodes. Local utility value and its own task priority weight Simultaneously, it receives broadcast information from other nodes, forming a global information set. ,Right now: ； S74) Each node calculates the global utility value of each candidate instruction based on the global information set using a weighted summation algorithm. Filter out the candidate instructions with the highest global utility value ,Right now: S75) If there are multiple candidate instructions whose global utility values ​​differ by less than a threshold To trigger conflict negotiation, the optimal instruction after final negotiation is determined using the Nash bargaining model in game theory. The Nash bargaining model is expressed as follows: ； In the formula, The retention utility value for each node; S76) The globally optimal instruction or the negotiated instruction is used as the initial control instruction and issued to the corresponding execution unit.

9. The joint system simulation verification method based on multi-source data injection according to claim 8, characterized in that: In step S8), the motion trajectory and attribute parameters of entities in the virtual environment are corrected through the digital twin module, specifically including the following steps: S81) Based on the Bayesian optimization algorithm, a dynamic motion trajectory containing key path points is generated. After the initial position data is injected, the simulation platform automatically turns and moves according to the predetermined motion program. During the simulation, the platform's speed, height, and attitude attributes are dynamically adjusted through an adaptive PID control algorithm according to real-time environmental parameters and task requirements. (S82) If the target is lost, the digital twin module predicts and generates the target's continuous motion trajectory based on the target's motion characteristics before the loss, using the extended Kalman filter algorithm.

10. A joint system simulation and verification device based on multi-source data injection, characterized in that, include: The simulation environment construction module, based on preset task objectives, constructs a virtual training environment model in the simulation platform, including geographical environment, troop units, task scenarios, and dynamic interference factors. The heterogeneous simulation data generation module dynamically generates heterogeneous simulation data based on the aforementioned virtual training environment model; The interference information generation module is used to generate false feature codes and dynamic intensity interference signals, wherein the false feature codes are generated by the generator of a deep convolutional generative adversarial network (DCGAN). The preprocessing module uses a Transformer-based deep learning classification model to clean up multi-source heterogeneous simulation data; and then injects the cleaned multi-source heterogeneous simulation data into the real system under test to drive the system under test to run. The control command generation module is used to fuse the injected purified multi-source heterogeneous simulation data and generate corresponding control commands. The simulation and evaluation module uses control commands to drive the update of entity states in the virtual environment, and evaluates the performance indicators of each tested system in terms of functionality, real-time performance, anti-interference ability, and decision optimization based on the updated states.

Images