Index characterization evaluation method based on aerospace equipment model

Abstract

The invention discloses an index representation evaluation method based on an aerospace equipment model. The method comprises the following steps: S1, extracting a task index set and mapping the task index set to each subsystem; s2, generating task situation data by using a generative adversarial network; s3, extracting features through self-supervised learning, and generating a data set; s4, fusing data through multi-modal learning, and generating unified feature representation; s5, utilizing a graph neural network to construct an inter-subsystem index relation graph; s6, performing time sequence modeling and task progress prediction by using the time sequence convolutional network; s7, generating adversarial self-supervised learning enhanced data and performing label self-calibration; and S8, performing unsupervised learning analysis on the data and dynamically adjusting an evaluation standard. Through combination of the generative adversarial network, the self-supervised learning, the graph neural network and the time sequence convolutional network, diversified data generation, subsystem relation modeling, time sequence prediction and task dynamic optimization of aerospace tasks are realized.

Description

Technical Field

[0001] The present invention relates to the field of aerospace technology, and in particular to an index characterization and evaluation method based on an aerospace equipment model. Background Art

[0002] With the complexity of space missions and the diversification of execution environments, how to improve the efficiency of space equipment and evaluate abnormal situations that may occur during mission execution has become an important issue that needs to be solved. Traditional space mission evaluation methods usually rely on expert experience and static models, lack the ability to adjust dynamically, especially during the mission execution, they cannot reflect the mutual influence between subsystems and their impact on the overall mission in real time. As missions become increasingly complex, the dependencies between mission indicators and subsystems involved in mission execution become more complex, so traditional rule-based evaluation methods are often unable to cope with these complex changes and potential mission risks.

[0003] Existing aerospace mission assessment systems generally rely on static mathematical models or expert-designed rule sets. These methods often lack accuracy and adaptability when faced with complex and changing mission scenarios. Especially in a multi-task, multi-subsystem collaborative execution environment, the mission assessment system cannot respond to changes in mission status in a timely manner, resulting in the inability to provide early warnings of potential risks during mission execution. In addition, existing assessment methods are usually based on a single data source, such as mission execution logs or sensor data, while ignoring the effective fusion of multimodal data from multiple sensors, subsystems, etc. This makes the traditional method of mission assessment lack comprehensiveness and accuracy, and cannot effectively capture dynamic changes in mission execution.

[0004] At the same time, with the development of artificial intelligence technology, it has become a trend to use machine learning and deep learning methods to improve the evaluation and optimization of space missions. Current research has begun to explore the use of deep learning methods to automatically extract key features of mission data and make predictions, but these methods still have some defects. First, when dealing with high-dimensional and multimodal data, existing machine learning methods often face problems such as difficulty in data fusion and complex feature selection, resulting in insufficient accuracy and real-time performance of mission evaluation results. Secondly, existing deep learning models usually rely on a large amount of labeled data for training, while the labeling of mission data is usually difficult and costly, especially during the execution of space missions, when much data is unlabeled. This makes the existing self-supervised learning methods not yet fully applied to the processing of space mission data.

[0005] Therefore, how to provide an indicator characterization and evaluation method based on aerospace equipment models is an urgent problem that technicians in this field need to solve. Summary of the invention

[0006] One purpose of the present invention is to propose an indicator characterization and evaluation method based on aerospace equipment models. The present invention proposes a new task evaluation method by combining advanced technologies such as generative adversarial networks, self-supervised learning, graph neural networks, and temporal convolutional networks. It can reduce dependence on manually labeled data by generating diverse task simulation data and extracting effective features from them, thereby improving the diversity and quality of task data sets. At the same time, by using graph neural networks to establish an indicator relationship diagram between the subsystems of aerospace equipment, it is possible to accurately capture the dependencies and synergies between the subsystems, thereby optimizing the overall evaluation of the task. By modeling the time series data in task execution through a temporal convolutional network, it is possible to predict the task progress and the state changes of each subsystem in real time, and dynamically adjust the task evaluation criteria and execution strategies according to the actual progress of the task.

[0007] According to an embodiment of the present invention, an index characterization and evaluation method based on aerospace equipment model includes the following steps:

[0008] S1. Through mission requirement analysis, identify and decompose various performance indicators of aerospace missions, map these indicators to different subsystems of aerospace equipment, and generate a mission indicator set;

[0009] S2. Generate task scenario data based on the task indicator set based on the generative adversarial network to form task simulation data;

[0010] S3, using self-supervised learning to automatically extract effective features from task simulation data and generate a data set for training;

[0011] S4, combined with multimodal learning methods, the dataset is fused with data from different sensors to generate a unified feature representation;

[0012] S5. Based on the generated unified feature representation, the indicator relationship diagram between the subsystems of aerospace equipment is constructed through the graph neural network to optimize the interdependence between the subsystems;

[0013] S6. Apply the time series convolutional network to perform time series modeling on the indicator relationship graph, generate a time series model, and analyze the mission progress, spacecraft status, and environmental change data;

[0014] S7. Based on task simulation data, we use generative adversarial self-supervised learning to generate diverse data, and use self-supervised learning to perform label self-calibration to provide diverse data for training the time series model.

[0015] S8. Use unsupervised learning methods to analyze diverse data, automatically identify potential patterns and anomalies, and dynamically adjust evaluation criteria based on different stages of the task and environmental changes.

[0016] Optionally, the S2 specifically includes:

[0017] S21, based on the task indicator set, extract the core indicators of task execution from the task requirement analysis, and use the generator in the generative adversarial network to generate task context data;

[0018] S22. Build a generative adversarial network discriminator to evaluate the authenticity of task context data and optimize the generator to make the task context data more consistent with the task's indicator set requirements;

[0019] S23, in each training iteration, the generator generates task simulation data based on the task indicator set, inputs a random noise vector and a high-dimensional feature representation;

[0020] S24. Back-propagation optimization of the generator based on the loss function output by the discriminator:

[0021]

[0022] Among them, L is the quantitative formula, D real is the real task data, D fake is the generated task simulation data, D(·) is the authenticity score of the task situation output by the discriminator;

[0023] S25. During the iteration process, the generator continuously adjusts parameters to generate task scenario data with a set of indicators that better matches the task.

[0024] Optionally, the S3 specifically includes:

[0025] S31, extract key features from the generated task simulation data set, use a self-supervised learning algorithm to perform unsupervised feature learning on the data, and generate a preliminary feature representation;

[0026] S32. Use convolutional neural network to process the temporal features of task simulation data, generate high-dimensional feature representation, and capture the temporal dependency and dynamic change information during task execution;

[0027] S33, integrating the features extracted by self-supervised learning and convolutional neural network to generate a data set;

[0028] S34, performing cluster analysis on the data set, classifying the data according to the task execution context, and generating labels for each type of task context;

[0029] S35. Use a deep autoencoder to reduce the dimension of the data set to generate low-dimensional task data;

[0030] S36. Standardize low-dimensional task data:

[0031]

[0032] Where D standardized For standardization, D reduced is the low-dimensional task data, μ is the mean of the data set, σ is the standard deviation, and the final data set is generated.

[0033] Optionally, the S4 specifically includes:

[0034] S41. Based on the generated dataset and multimodal datasets from different sensors, align the task simulation data with the sensor data, build a unified input feature matrix, and synchronize different data sources.

[0035] S42, extracting features from the multimodal data set through a convolutional neural network to obtain high-dimensional features of each modality;

[0036] S43, using multimodal learning algorithms to fuse data sets and high-dimensional features to generate a unified feature representation, and using weighted fusion strategies to uniformly represent multi-source data as a multi-dimensional vector;

[0037] S44. Weighted fusion of multi-dimensional vectors:

[0038]

[0039] Among them, F fused is the fusion feature matrix, n represents multi-dimensionality, w i is the weight coefficient of the modal data feature, It is a high-dimensional feature;

[0040] S45. Batch normalize the fused feature matrix:

[0041]

[0042] Among them, F fused is the fusion feature matrix, F normalized is the normalized feature matrix, μ is the mean of the feature matrix, and σ is the standard deviation;

[0043] S46. Generate a final unified feature representation based on the normalized feature matrix.

[0044] Optionally, the S5 specifically includes:

[0045] S51, extracting the task index of each subsystem from the task simulation data according to the generated unified feature representation, and generating a subsystem feature vector;

[0046] S52, constructing an indicator relationship diagram between the subsystems of aerospace equipment by using the generated subsystem feature vectors and the indicator dependency relationship of the task;

[0047] S53, using the indicator relationship graph and the subsystem feature vector through the graph neural network, updating the features of each subsystem, capturing the dependencies between subsystems, and generating the updated feature representation of each subsystem;

[0048] S54, through the information transmission mechanism, information is transmitted and aggregated between the subsystem nodes to generate the final subsystem feature representation;

[0049] S55. Based on the final subsystem feature representation, calculate the overall evaluation index of the task to provide a comprehensive evaluation result for task optimization.

[0050] Optionally, the S6 specifically includes:

[0051] S61. Based on the final subsystem feature representation, the feature of each subsystem is used as the input of the time series data to construct a time series input matrix;

[0052] S62, applying a temporal convolutional network to process the temporal input matrix, extracting temporal features through multi-layer convolution operations, and generating a deep feature representation of the temporal data;

[0053] S63, through the causal convolution operation in the convolutional neural network, generating the time series feature representation after the causal convolution, and further optimizing the time series modeling;

[0054] H causal =σ(R causal ·X time ·W causal );

[0055] Among them, H causal is the temporal feature representation of the output of the causal convolution operation, R causal is the causal mask in the causal convolution, X time is the time series input matrix, W causal is the weight matrix of causal convolution, σ is the activation function;

[0056] S64. Use the skip connection technology in the convolutional neural network to merge the convolution outputs of each layer to generate the final temporal feature representation;

[0057] S65. Based on the final time series feature representation, predict the future state of task execution and generate a time series prediction result of the task;

[0058] S66. Based on the time series prediction results, combined with the task objectives and constraints, dynamically adjust and optimize the task progress, reasonably allocate task resources, and finally output the optimized task execution plan.

[0059] Optionally, the S8 specifically includes:

[0060] S81. Based on the generated diverse data, perform cluster analysis on the diverse data through unsupervised learning algorithms to identify the patterns of task situations;

[0061] S82. Use K-means clustering algorithm to cluster the diverse data and generate cluster labels c for each data point. i , and measure the clustering effect of different task situations by calculating the density of points within the cluster;

[0062] S83. Based on the clustering effect, calculate the density distribution of each task context cluster, and use the density value to identify abnormal situations of the diversified data;

[0063] S84, analyzing the clustering effect and abnormal data, automatically adjusting the task evaluation standard, and generating the dynamically adjusted task evaluation standard according to different stages of task execution and environmental changes by dynamically adjusting various performance indicators of the task;

[0064] S85. Optimize task progress, subsystem status and resource allocation based on dynamically adjusted task evaluation criteria:

[0065]

[0066] Among them, I final is the final task evaluation result, p is the dimension of the task evaluation standard, w i is the weight coefficient of the evaluation criteria, evaluation criteria for adjusted tasks;

[0067] S86. Monitor task execution through a real-time feedback mechanism, and adjust task evaluation criteria in real time by combining smart wearable devices and sensor data.

[0068] The beneficial effects of the present invention are:

[0069] The present invention successfully solves many problems existing in existing space mission assessment methods by combining advanced technologies such as generative adversarial networks, self-supervised learning, graph neural networks, and temporal convolutional networks, and significantly improves the accuracy, real-time, and robustness of space mission assessment. First, by using generative adversarial networks, the present invention is able to generate diverse mission simulation data, greatly enriching the mission data set and reducing reliance on manually labeled data. This not only increases the diversity of the data, but also ensures that more mission execution scenarios can be covered during the assessment process, including normal tasks and potential failure modes, thereby providing more comprehensive and accurate data support for mission assessment.

[0070] Through self-supervised learning, the present invention further improves the quality of task data sets, automatically extracts effective features from generated task simulation data, reduces reliance on labeled data, and reduces the cost and time of data labeling. In particular, in the process of generating task data, the present invention combines generative adversarial self-supervised learning, and the generated task data is not only diversified, but also capable of self-calibration labels to ensure the authenticity and validity of the data. This innovative design provides more reliable training data for subsequent task evaluations, further improving the accuracy of the evaluation system.

[0071] The present invention also successfully solves the problem of complex dependency modeling between subsystems through graph neural networks. The indicator relationship diagram constructed using GCN can capture the interdependence and synergy between the subsystems of aerospace equipment, thereby accurately reflecting the impact of different subsystems in mission execution. This modeling method based on graph neural networks allows task evaluation to not only consider the performance of a single subsystem, but also comprehensively evaluate the collaborative work effect of all subsystems, ensuring the overall performance of the task and the optimization of various task objectives.

[0072] In addition, the present invention models the time series data in task execution through a time series convolutional network, which solves the shortcomings of existing methods in task progress prediction and dynamic adjustment. The time series convolutional network can capture the time series dependencies in the task execution process, accurately predict key indicators such as task progress, task success probability, resource consumption, and dynamically adjust the evaluation criteria and execution strategy according to the actual progress of the task. This dynamic evaluation method based on time series modeling ensures that the task can respond to changes in task status in real time, further improving the flexibility and robustness of task execution.

[0073] Through multimodal learning, the present invention solves the problem of difficulty in multimodal data fusion in the prior art. After effectively fusing data from different sensors, a unified feature representation is generated, providing more comprehensive and accurate data support for task evaluation. This data fusion technology greatly improves the data utilization efficiency in the task evaluation process and ensures the accuracy and stability of the evaluation system when processing complex tasks and multimodal data. BRIEF DESCRIPTION OF THE DRAWINGS

[0074] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0075] Figure 1 A flowchart of an index characterization and evaluation method based on aerospace equipment model proposed by the present invention;

[0076] Figure 2This is a specific structural diagram of the time series modeling and dynamic prediction of the indicator characterization evaluation method based on the aerospace equipment model proposed by the present invention. DETAILED DESCRIPTION

[0077] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, which only illustrate the basic structure of the present invention in a schematic manner, and therefore only show the components related to the present invention.

[0078] refer to Figure 1-2 , an index characterization and evaluation method based on aerospace equipment model, comprising the following steps:

[0079] S1. Through mission requirement analysis, identify and decompose various performance indicators of aerospace missions, map these indicators to different subsystems of aerospace equipment, and generate a mission indicator set;

[0080] S2. Generate task scenario data based on the task indicator set based on the generative adversarial network to form task simulation data;

[0081] S3, using self-supervised learning to automatically extract effective features from task simulation data and generate a data set for training;

[0082] S4, combined with multimodal learning methods, the dataset is fused with data from different sensors to generate a unified feature representation;

[0083] S5. Based on the generated unified feature representation, the indicator relationship diagram between the subsystems of aerospace equipment is constructed through the graph neural network to optimize the interdependence between the subsystems;

[0084] S6. Apply the time series convolutional network to perform time series modeling on the indicator relationship graph, generate a time series model, and analyze the mission progress, spacecraft status, and environmental change data;

[0085] S7. Based on task simulation data, we use generative adversarial self-supervised learning to generate diverse data, and use self-supervised learning to perform label self-calibration to provide diverse data for training the time series model.

[0086] S8. Use unsupervised learning methods to analyze diverse data, automatically identify potential patterns and anomalies, and dynamically adjust evaluation criteria based on different stages of the task and environmental changes.

[0087] In this implementation, S2 specifically includes:

[0088] S21, based on the task indicator set, extract the core indicators of task execution from the task requirement analysis, and use the generator in the generative adversarial network to generate task context data;

[0089] S22. Build a generative adversarial network discriminator to evaluate the authenticity of task context data and optimize the generator to make the task context data more consistent with the task's indicator set requirements;

[0090] S23, in each training iteration, the generator generates task simulation data based on the task indicator set, inputs a random noise vector and a high-dimensional feature representation;

[0091] S24. Back-propagation optimization of the generator based on the loss function output by the discriminator:

[0092]

[0093] Among them, L is the quantitative formula, D real is the real task data, D fake is the generated task simulation data, D(·) is the authenticity score of the task situation output by the discriminator;

[0094] S25. During the iteration process, the generator continuously adjusts parameters to generate task scenario data with a set of indicators that better matches the task.

[0095] In this implementation, S3 specifically includes:

[0096] S31, extract key features from the generated task simulation data set, use a self-supervised learning algorithm to perform unsupervised feature learning on the data, and generate a preliminary feature representation;

[0097] S32. Use convolutional neural network to process the temporal features of task simulation data, generate high-dimensional feature representation, and capture the temporal dependency and dynamic change information during task execution;

[0098] S33, integrating the features extracted by self-supervised learning and convolutional neural network to generate a data set;

[0099] S34, performing cluster analysis on the data set, classifying the data according to the task execution context, and generating labels for each type of task context;

[0100] S35. Use a deep autoencoder to reduce the dimension of the data set to generate low-dimensional task data;

[0101] S36. Standardize low-dimensional task data:

[0102]

[0103] Where D standardized For standardization, D reduced is the low-dimensional task data, μ is the mean of the data set, σ is the standard deviation, and the final data set is generated.

[0104] In this implementation, S4 specifically includes:

[0105] S41. Based on the generated dataset and multimodal datasets from different sensors, align the task simulation data with the sensor data, build a unified input feature matrix, and synchronize different data sources.

[0106] S42, extracting features from the multimodal data set through a convolutional neural network to obtain high-dimensional features of each modality;

[0107] S43, using multimodal learning algorithms to fuse data sets and high-dimensional features to generate a unified feature representation, and using weighted fusion strategies to uniformly represent multi-source data as a multi-dimensional vector;

[0108] S44. Weighted fusion of multi-dimensional vectors:

[0109]

[0110] Among them, F fused is the fusion feature matrix, n represents multi-dimensionality, w i is the weight coefficient of the modal data feature, It is a high-dimensional feature;

[0111] S45. Batch normalize the fused feature matrix:

[0112]

[0113] Among them, F fused is the fusion feature matrix, F normalized is the normalized feature matrix, μ is the mean of the feature matrix, and σ is the standard deviation;

[0114] S46. Generate a final unified feature representation based on the normalized feature matrix.

[0115] In this implementation manner, S5 specifically includes:

[0116] S51, extracting the task index of each subsystem from the task simulation data according to the generated unified feature representation, and generating a subsystem feature vector;

[0117] S52, constructing an indicator relationship diagram between the subsystems of aerospace equipment by using the generated subsystem feature vectors and the indicator dependency relationship of the task;

[0118] S53, using the indicator relationship graph and the subsystem feature vector through the graph neural network, updating the features of each subsystem, capturing the dependencies between subsystems, and generating the updated feature representation of each subsystem;

[0119] S54, through the information transmission mechanism, information is transmitted and aggregated between the subsystem nodes to generate the final subsystem feature representation;

[0120] S55. Based on the final subsystem feature representation, calculate the overall evaluation index of the task to provide a comprehensive evaluation result for task optimization.

[0121] In this implementation manner, S6 specifically includes:

[0122] S61. Based on the final subsystem feature representation, the feature of each subsystem is used as the input of the time series data to construct a time series input matrix;

[0123] S62, applying a temporal convolutional network to process the temporal input matrix, extracting temporal features through multi-layer convolution operations, and generating a deep feature representation of the temporal data;

[0124] S63, through the causal convolution operation in the convolutional neural network, generating the time series feature representation after the causal convolution, and further optimizing the time series modeling;

[0125] H causal =σ(R causal ·X time ·W causal );

[0126] Among them, H causal is the temporal feature representation of the output of the causal convolution operation, R causal is the causal mask in the causal convolution, X time is the time series input matrix, W causal is the weight matrix of causal convolution, σ is the activation function;

[0127] S64. Use the skip connection technology in the convolutional neural network to merge the convolution outputs of each layer to generate the final temporal feature representation;

[0128] S65. Based on the final time series feature representation, predict the future state of task execution and generate a time series prediction result of the task;

[0129] S66. Based on the time series prediction results, combined with the task objectives and constraints, dynamically adjust and optimize the task progress, reasonably allocate task resources, and finally output the optimized task execution plan.

[0130] In this implementation manner, S8 specifically includes:

[0131] S81. Based on the generated diverse data, perform cluster analysis on the diverse data through unsupervised learning algorithms to identify the patterns of task situations;

[0132] S82. Use K-means clustering algorithm to cluster the diverse data and generate cluster labels c for each data point. i , and measure the clustering effect of different task situations by calculating the density of points within the cluster;

[0133] S83. Based on the clustering effect, calculate the density distribution of each task context cluster, and use the density value to identify abnormal situations of the diversified data;

[0134] S84, analyzing the clustering effect and abnormal data, automatically adjusting the task evaluation standard, and generating the dynamically adjusted task evaluation standard according to different stages of task execution and environmental changes by dynamically adjusting various performance indicators of the task;

[0135] S85. Optimize task progress, subsystem status and resource allocation based on dynamically adjusted task evaluation criteria:

[0136]

[0137] Among them, I final is the final task evaluation result, p is the dimension of the task evaluation standard, w i is the weight coefficient of the evaluation criteria, evaluation criteria for adjusted tasks;

[0138] S86. Monitor task execution through a real-time feedback mechanism, and adjust task evaluation criteria in real time by combining smart wearable devices and sensor data.

[0139] Embodiment 1:

[0140] In order to verify the feasibility of the present invention in implementation, the present invention is applied to simulated deep space exploration. The goal of the mission is to simulate the spacecraft to perform a series of complex operations during the exploration process, including planetary orbit adjustment, data collection, and energy management. The spacecraft used in this mission includes multiple subsystems, such as propulsion systems, energy management systems, thermal control systems, and communication systems. Each subsystem plays a vital role in the execution of the mission. Because the mission will be affected by various environmental factors for a long time, and there are complex dependencies between the subsystems, traditional mission evaluation methods often cannot effectively predict mission progress, resource consumption, and possible failures.

[0141] In this simulation task, the method of the present invention is applied to solve the problems of existing evaluation methods, such as the inability to respond to dynamic changes in task execution in real time, the scarcity of task data, and the insufficient modeling of dependencies between subsystems. By generating diverse task simulation data through generative adversarial networks, the task evaluation system can simulate the task execution process under different scenarios, such as normal tasks, propulsion system failures, and energy shortages, greatly enhancing the diversity and coverage of task data.

[0142] At the beginning of the mission, 50,000 mission simulation data were generated based on the generative adversarial network, including normal execution scenarios, propulsion system failures, communication interruptions, abnormal energy consumption, etc. These data not only cover the normal situations that may occur during the mission execution, but also simulate sudden failures and mission execution scenarios in complex environments. Through self-supervised learning, the system automatically extracts key features from these generated mission data, such as mission success rate, propulsion efficiency, communication delay, temperature fluctuations, etc., providing rich feature data for subsequent mission evaluation.

[0143] Subsequently, a graph neural network was used to construct an indicator relationship diagram between the subsystems of the spacecraft, capturing the interdependence between the subsystems. Through the graph neural network model, the system can adjust the operating status of each subsystem in real time to ensure the smooth execution of the mission. When the performance of the propulsion system degrades, the graph neural network can promptly identify and prompt the energy management system to reallocate resources, thereby avoiding energy shortages that may occur in the middle of the mission.

[0144] During the mission execution, the time series convolutional network is used to model and predict the time series data of the mission execution. By performing time series modeling on the real-time data of each subsystem of the spacecraft during the mission, the time series convolutional network can accurately predict the progress of the mission and the state changes of each subsystem. At the 100th hour of the mission execution, the time series convolutional network predicted that the fuel consumption of the propulsion system was about to exceed the predetermined value, which might cause the mission to fail to be completed on time. Based on this prediction, the mission evaluation system adjusted the energy allocation strategy to prioritize the energy needs of the propulsion system, successfully avoiding the problem of fuel shortage and ensuring the smooth progress of the mission in the subsequent stages.

[0145] After the task was completed, the system compared the performance of the traditional task evaluation method and the method of the present invention. The traditional method was based on static models and manually set rules for evaluation, and the task success rate was 85%. The method of the present invention achieved a final task success rate of 95% through real-time prediction and dynamic adjustment of task evaluation criteria. In terms of fault prediction, the accuracy of the traditional method was 70%, but through the dynamic evaluation of generative adversarial networks and graph neural networks combined with temporal convolutional networks, the method of the present invention improved the fault prediction accuracy to 92%. In terms of resource utilization, the resource utilization rate of the traditional method was 80%, but the method of the present invention increased the resource utilization rate to 90% through dynamic task adjustment and resource optimization.

[0146] Table 1 Statistics of task simulation datasets

[0147]

[0148] Table 2 Comparison of task execution performance

[0149]

[0150] Table 1 shows the basic situation of the mission simulation data set generated during the implementation of the present invention. The data set contains a total of 50,000 mission simulation data, covering normal scenarios, fault scenarios and special scenario mission data in mission execution, ensuring the comprehensiveness and diversity of mission evaluation. The generated mission data covers more than 95% of mission execution scenarios, including normal mission execution, propulsion system failure, communication interruption, abnormal energy consumption and other scenarios. As can be seen from the table, normal mission data accounts for 60% of the mission simulation data. This type of data reflects the mission execution of the spacecraft under normal conditions, including the standard operating status of each subsystem such as the propulsion system, energy system, and communication system. Fault mission data and special scenario mission data account for 30% and 10% respectively. These data simulate possible failures or abnormal situations during the mission, such as propulsion system failure, energy system abnormality, communication interruption, etc. These failure and abnormal scenario data are crucial for evaluating the risks and potential problems that may occur during the mission, and can help the mission evaluation model better cope with the complex and changeable mission scenarios in reality.

[0151] Table 2 provides the comparison results of the traditional task evaluation method and the method of the present invention on different evaluation indicators. By comparing multiple indicators of task evaluation, it is demonstrated that the method of the present invention has significant advantages over the traditional method in terms of task success rate, fault prediction accuracy, resource utilization and task completion time. The traditional method is 85%, while the method of the present invention reaches 95%, an increase of 10%. The accuracy of the traditional method is 70%, while the accuracy of the method of the present invention is 92%, an increase of 22%. The resource utilization of the traditional method is 80%, while the resource utilization of the method of the present invention is 90%, an increase of 10%. The traditional method requires 120 hours to complete the task, while the method of the present invention can shorten the task completion time to 108 hours, reducing the time by 10%.

[0152] The data in Tables 1 and 2 show that the method of the present invention has significant advantages over traditional methods in many aspects. First, the diversified mission data generated by the generative adversarial network enables the mission evaluation to cover more mission scenarios, reducing the problems of data scarcity and insufficient coverage in traditional methods. Secondly, the graph neural network can accurately model the dependencies between the subsystems of the spacecraft, providing a more accurate prediction of the subsystem status for mission evaluation. The application of the temporal convolutional network enables real-time prediction and dynamic adjustment of key indicators such as mission progress and resource consumption, avoiding time delays and resource waste caused by inaccurate task execution in traditional methods.

[0153] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and inventive concept of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. An index characterization and evaluation method based on aerospace equipment model, characterized in that: The steps include: S1. Through mission requirement analysis, identify and decompose various performance indicators of aerospace missions, map various performance indicators to different subsystems of aerospace equipment, and generate a mission indicator set; S2. Generate task scenario data based on the task indicator set based on the generative adversarial network to form task simulation data; S3, using self-supervised learning to automatically extract effective features from task simulation data and generate a data set for training; S4, combined with multimodal learning methods, the dataset is fused with data from different sensors to generate a unified feature representation; S5. Based on the generated unified feature representation, the indicator relationship diagram between the subsystems of aerospace equipment is constructed through the graph neural network to optimize the interdependence between the subsystems; S6. Apply the time series convolutional network to perform time series modeling on the indicator relationship graph, generate a time series model, and analyze the mission progress, spacecraft status, and environmental change data; S7. Based on task simulation data, we use generative adversarial self-supervised learning to generate diverse data, and use self-supervised learning to perform label self-calibration to provide diverse data for training the time series model. S8. Use unsupervised learning methods to analyze diverse data, automatically identify potential patterns and anomalies, and dynamically adjust evaluation criteria based on different stages of the task and environmental changes.

2. The index characterization and evaluation method based on aerospace equipment model according to claim 1 is characterized in that: The S2 specifically includes: S21, based on the task indicator set, extract the core indicators of task execution from the task requirement analysis, and use the generator in the generative adversarial network to generate task context data; S22. Build a generative adversarial network discriminator to evaluate the authenticity of task context data and optimize the generator to make the task context data more consistent with the task's indicator set requirements; S23, in each training iteration, the generator generates task simulation data based on the task indicator set, inputs a random noise vector and a high-dimensional feature representation; S24. Back-propagation optimization of the generator based on the loss function output by the discriminator: Among them, L is the quantitative formula, D real is the real task data, D fake is the generated task simulation data, D(·) is the authenticity score of the task situation output by the discriminator; S25. During the iteration process, the generator continuously adjusts parameters to generate task scenario data with a set of indicators that better matches the task.

3. The index characterization and evaluation method based on aerospace equipment model according to claim 1 is characterized in that: The S3 specifically includes: S31, extract key features from the generated task simulation data set, use a self-supervised learning algorithm to perform unsupervised feature learning on the data, and generate a preliminary feature representation; S32. Use convolutional neural network to process the temporal features of task simulation data, generate high-dimensional feature representation, and capture the temporal dependency and dynamic change information during task execution; S33, integrating the features extracted by self-supervised learning and convolutional neural network to generate a data set; S34, performing cluster analysis on the data set, classifying the data according to the task execution context, and generating labels for each type of task context; S35. Use a deep autoencoder to reduce the dimension of the data set to generate low-dimensional task data; S36. Standardize low-dimensional task data: Where D standardized For standardization, D reduced is the low-dimensional task data, μ is the mean of the data set, σ is the standard deviation, and the final data set is generated.

4. The index characterization and evaluation method based on aerospace equipment model according to claim 1 is characterized in that: The S4 specifically includes: S41. Based on the generated dataset and multimodal datasets from different sensors, align the task simulation data with the sensor data, build a unified input feature matrix, and synchronize different data sources. S42, extracting features from the multimodal data set through a convolutional neural network to obtain high-dimensional features of each modality; S43, using multimodal learning algorithms to fuse data sets and high-dimensional features to generate a unified feature representation, and using weighted fusion strategies to uniformly represent multi-source data as a multi-dimensional vector; S44. Weighted fusion of multi-dimensional vectors: Among them, F fused is the fusion feature matrix, n represents multi-dimensionality, w i is the weight coefficient of the modal data feature, f modali It is a high-dimensional feature; S45. Batch normalize the fused feature matrix: Among them, F fused is the fusion feature matrix, F normalized is the normalized feature matrix, μ is the mean of the feature matrix, and σ is the standard deviation; S46. Generate a final unified feature representation based on the normalized feature matrix.

5. The index characterization and evaluation method based on aerospace equipment model according to claim 1 is characterized in that: The S5 specifically includes: S51, extracting the task index of each subsystem from the task simulation data according to the generated unified feature representation, and generating a subsystem feature vector; S52, constructing an indicator relationship diagram between the subsystems of aerospace equipment by using the generated subsystem feature vectors and the indicator dependency relationship of the task; S53, using the indicator relationship graph and the subsystem feature vector through the graph neural network, updating the features of each subsystem, capturing the dependencies between subsystems, and generating the updated feature representation of each subsystem; S54, through the information transmission mechanism, information is transmitted and aggregated between the subsystem nodes to generate the final subsystem feature representation; S55. Based on the final subsystem feature representation, calculate the overall evaluation index of the task to provide a comprehensive evaluation result for task optimization.

6. The index characterization and evaluation method based on aerospace equipment model according to claim 1 is characterized in that: The S6 specifically includes: S61. Based on the final subsystem feature representation, the feature of each subsystem is used as the input of the time series data to construct a time series input matrix; S62, applying a temporal convolutional network to process the temporal input matrix, extracting temporal features through multi-layer convolution operations, and generating a deep feature representation of the temporal data; S63, through the causal convolution operation in the convolutional neural network, generating the time series feature representation after the causal convolution, and further optimizing the time series modeling; H causal =σ(R causal ·X time ·W causal ); Among them, H causal is the temporal feature representation of the output of the causal convolution operation, R causal is the causal mask in the causal convolution, X time is the time series input matrix, W causal is the weight matrix of causal convolution, σ is the activation function; S64. Use the skip connection technology in the convolutional neural network to merge the convolution outputs of each layer to generate the final temporal feature representation; S65. Based on the final time series feature representation, predict the future state of task execution and generate a time series prediction result of the task; S66. Based on the time series prediction results, combined with the task objectives and constraints, dynamically adjust and optimize the task progress, reasonably allocate task resources, and finally output the optimized task execution plan.

7. The index characterization and evaluation method based on aerospace equipment model according to claim 1 is characterized in that: The S8 specifically includes: S81. Based on the generated diverse data, perform cluster analysis on the diverse data through unsupervised learning algorithms to identify the patterns of task situations; S82. Use K-means clustering algorithm to cluster the diverse data and generate cluster labels c for each data point. i , and measure the clustering effect of different task situations by calculating the density of points within the cluster; S83. Based on the clustering effect, calculate the density distribution of each task context cluster, and use the density value to identify abnormal situations of the diversified data; S84, analyzing the clustering effect and abnormal data, automatically adjusting the task evaluation standard, and generating the dynamically adjusted task evaluation standard according to different stages of task execution and environmental changes by dynamically adjusting various performance indicators of the task; S85. Optimize task progress, subsystem status and resource allocation based on dynamically adjusted task evaluation criteria: Among them, I final is the final task evaluation result, p is the dimension of the task evaluation standard, w i is the weight coefficient of the evaluation criteria, evaluation criteria for adjusted tasks; S86. Monitor task execution through a real-time feedback mechanism, and adjust task evaluation criteria in real time by combining smart wearable devices and sensor data.

Images