Low earth orbit satellite communication link switching method based on deep learning

Abstract

The invention discloses a low-orbit satellite communication link switching method based on deep learning, relates to the technical field of satellite communication link switching, and aims to solve the problem that the switching effect is poor when a low-orbit satellite communication link is switched. According to the method, switching strategy reasoning is carried out by utilizing the powerful computing and learning capabilities of the neural network model, different link environments can be automatically adapted, the accuracy and flexibility of switching decision making are improved, a switching strategy comprises link selection, switching opportunity and switching conditions, and the switching strategy is expressed through a probability value, a classification label or a numerical value, so that the accuracy and the flexibility of switching decision making are improved. According to the technical scheme, various possible switching schemes can be provided, if the operation process or performance of link switching does not reach the expected standard, link switching is carried out again according to the switching strategy, the link switching can be ensured to be in the optimal state all the time through the continuous optimization and adjustment process, and the overall performance and stability of a communication system are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of satellite communication link switching, and specifically to a low-orbit satellite communication link switching method based on deep learning. Background Art

[0002] Satellite communication link switching refers to the process of transferring from one communication link to another in a satellite communication system according to certain conditions and strategies.

[0003] The Chinese patent with publication number CN116113004A discloses a satellite link switching method based on electronic fences. The main method is to arbitrarily partition the coverage area within the same satellite coverage area, and to define the satellite coverage area in multiple ways according to the rules of the electronic fence, so as to realize different service areas under the same satellite, so as to improve the utilization of satellite bandwidth and the service level of the satellite within the coverage area. In addition, different application strategies and communication strategies can be implemented within the coverage area of ​​the same satellite or within the coverage area of ​​multiple satellites to meet the needs of satellite communication in different scenarios. Although the above patent solves the problem of link switching, the following problems still exist in actual operation: 1. The acquired low-orbit satellite communication link data was not effectively processed and data feature extracted, resulting in the unclear internal structure and characteristics of the data.

[0004] 2. Failure to build a more specific model based on the requirements of low-orbit satellite communication links results in an inability to adapt to unexpected situations when performing link switching.

[0005] 3. The link switching mode and requirements are not adjusted according to the actual situation, resulting in poor link switching effect. Summary of the invention

[0006] The purpose of the present invention is to provide a low-orbit satellite communication link switching method based on deep learning, which uses the powerful computing and learning capabilities of a neural network model to reason about switching strategies, can automatically adapt to different link environments, and improve the accuracy and flexibility of switching decisions. The switching strategy includes link selection, switching timing, and switching conditions, which are expressed by probability values, classification labels, or numerical values, and can provide a variety of possible switching schemes. If the operation process or performance of the link switching does not meet the expected standards, the link switching is performed again according to the switching strategy. The process of continuous optimization and adjustment can ensure that the link switching is always in the best state, improve the overall performance and stability of the communication system, and solve the problems in the prior art.

[0007] To achieve the above object, the present invention provides the following technical solutions: A low-orbit satellite communication link switching method based on deep learning includes: S1: Satellite data collection and processing: collect the initial data of the low-orbit satellite communication link, and perform data preprocessing after the collection is completed. After the data preprocessing, the target link data is obtained; S2: Extraction of collected data features: Extract key feature data from the target link data and mark the extracted key feature data as model output parameter data; S3: Feature data model construction: A deep learning model is constructed according to the requirements of the low-orbit satellite communication link. After the deep learning model is constructed, a neural network model is obtained; S4: Model training and optimization: Use the model output parameter data to train the neural network model, adjust the model parameters of the neural network model according to the model training situation, and obtain the link neural network model after the adjustment is completed; S5: Real-time link switching execution: The link neural network model is transmitted to the control terminal, and the real-time link switching is implemented. According to the switching strategy output by the model, the link switching is automatically executed, and the real-time communication link data in the low-orbit satellite communication system is collected; S6: Switching link performance evaluation: Evaluate the collected real-time communication link data and the performance of the link switching method, and optimize and adjust according to the evaluation results.

[0008] Preferably, the initial data of the low-orbit satellite communication link is collected in S1, and after the collection is completed, the data is preprocessed, including: Retrieve initial data of low-orbit satellite communication links from the database; The initial data of the low-orbit satellite communication link includes satellite status data, communication link data, environmental factor data, ground station data, user terminal data and historical communication data; After the initial data collection is completed, data preprocessing is performed, which includes data cleaning, data standardization, data conversion and data enhancement; After data preprocessing is completed, the target link data is obtained.

[0009] Preferably, the initial data of the low-orbit satellite communication link is collected in S1, and data preprocessing is performed after the collection is completed, which also includes: The target link data is subjected to quality determination, and a data abnormality alarm is issued when the quality is abnormal, wherein the data amount of each type of initial data contained in the target link data is extracted; the initial data includes satellite status data, communication link data, environmental factor data, ground station data, user terminal data and historical communication data; Extracting the data amount of each type of initial data included in the target link data as the first data amount; Extracting the data volume of each type of initial data before data preprocessing as the second data volume; Obtaining a data anomaly coefficient of each type of initial data using the first data amount and the second data amount; The data anomaly coefficient of each type of initial data is obtained by the following formula: ; Among them, J represents the data anomaly coefficient of each initial data; C 01 and C 02 represents the first data amount and the second data amount respectively; n represents the number of data contained in each type of initial data before data preprocessing; m represents the number of data contained in each type of initial data corresponding to the target link data; C hi represents the data volume of the i-th data contained in each corresponding initial data in the target link data; C xi represents the data volume of the i-th data contained in each type of initial data corresponding to the target link data before data preprocessing; Compare the data anomaly coefficient of each initial data with a preset coefficient threshold; When the number of types of initial data whose data anomaly coefficient exceeds a preset coefficient threshold exceeds a preset number threshold, the initial data whose data anomaly coefficient exceeds the preset coefficient threshold is subjected to an abnormality degree determination and abnormality risk determination.

[0010] Preferably, when the number of types of initial data whose data abnormality coefficient exceeds a preset coefficient threshold exceeds a preset number threshold, the initial data whose data abnormality coefficient exceeds the preset coefficient threshold is subjected to abnormality degree determination and abnormality risk alarm, including: When the number of types of initial data whose data abnormality coefficient exceeds the preset coefficient threshold for each type of initial data exceeds the preset number threshold, the initial data whose data abnormality coefficient exceeds the preset coefficient threshold for each type of initial data is retrieved as the target initial data; Extracting the amount of invalid data and the amount of abnormal data in the target initial data during the data cleaning process; Extract the weight value corresponding to each invalid data and the weight value corresponding to the abnormal data; The weight evaluation coefficient corresponding to the target initial data is obtained by using the weight value corresponding to the invalid data and the weight value corresponding to the abnormal data in combination with the data abnormality coefficient of each initial data; The weight evaluation coefficient corresponding to the target initial data is obtained by the following formula: ; Among them, G represents the weight evaluation coefficient corresponding to the target initial data; x represents the number of invalid data; y represents the number of abnormal data; f xi Indicates the weight value corresponding to the i-th invalid data; f yi represents the weight value corresponding to the i-th abnormal data; f xz represents the middle value of the weight corresponding to x invalid data; f yz represents the middle value of the weight corresponding to y abnormal data; f xm Indicates the weight value corresponding to the maximum value of the data volume of each data contained in the invalid data; f ym Indicates the weight value corresponding to the maximum value of the data volume of each data contained in the abnormal data; f xmax Indicates the maximum value of the weight of invalid data; f ymax Indicates the maximum value of the weight when abnormal data appears; ε indicates the preset minimum value, which is used to prevent the numerator from being 0; J indicates the data abnormality coefficient of each initial data; Comparing the weight evaluation coefficient corresponding to the target initial data with a preset weight evaluation coefficient threshold; When the weight evaluation coefficient corresponding to the target initial data exceeds the preset weight evaluation coefficient threshold, it is determined that the abnormality degree corresponding to the target initial data is serious, and an abnormal risk alarm for the target initial data is issued; When the weight evaluation coefficient corresponding to the target initial data does not exceed the preset weight evaluation coefficient threshold, the amount of invalid data and abnormal data corresponding to the target initial data is used in combination with the weight evaluation coefficient corresponding to the target initial data to determine the degree of abnormality and issue an abnormal risk alarm.

[0011] Preferably, when the weight evaluation coefficient corresponding to the target initial data does not exceed the preset weight evaluation coefficient threshold, the amount of invalid data and abnormal data corresponding to the target initial data is combined with the weight evaluation coefficient corresponding to the target initial data to perform abnormal degree determination and abnormal risk alarm, including: When the weight evaluation coefficient corresponding to the target initial data does not exceed the preset weight evaluation coefficient threshold, calling the coefficient difference between the weight evaluation coefficient corresponding to the target initial data and the preset weight evaluation coefficient threshold; Retrieve the amount of invalid data corresponding to the target initial data; Retrieving the data volume of abnormal data corresponding to the target initial data; Retrieving the weight evaluation coefficient corresponding to the target initial data; Obtaining an abnormal risk evaluation coefficient corresponding to the target initial data by using the data volume of invalid data and the data volume of abnormal data corresponding to the target initial data, the coefficient difference between the weight evaluation coefficient corresponding to the target initial data and the preset weight evaluation coefficient threshold, and the weight evaluation coefficient corresponding to the target initial data; The abnormal risk assessment coefficient corresponding to the target initial data is obtained by the following formula: ; Among them, Y represents the abnormal risk assessment coefficient corresponding to the target initial data; C r01 and C r02 represents the first data volume and the second data volume corresponding to the target initial data; C s and C w represents the amount of invalid data and abnormal data contained in the target initial data; G c represents the coefficient difference between the weight evaluation coefficient corresponding to the target initial data and the preset weight evaluation coefficient threshold; G represents the weight evaluation coefficient corresponding to the target initial data; G y Indicates the preset weight evaluation coefficient threshold; f s and f w Indicates the weighted sum of invalid data and abnormal data contained in the target initial data; Comparing the abnormal risk assessment coefficient corresponding to the target initial data with a preset abnormal risk assessment coefficient threshold; When the abnormal risk assessment coefficient corresponding to the target initial data exceeds a preset abnormal risk assessment coefficient threshold, it is determined that the abnormal degree corresponding to the target initial data is serious, and an abnormal risk alarm corresponding to the target initial data is issued.

[0012] Preferably, extracting key feature data from the target link data in S2 includes: The chi-square test method is used to confirm the feature types in the target link data, which include time features, spatial features and signal features; Among them, the characteristic types of satellite status data are time characteristics and space characteristics; the characteristic types of communication link data are time characteristics and signal characteristics; the characteristic types of environmental factor data are time characteristics and space characteristics; the characteristic types of ground station data are time characteristics and space characteristics; the characteristic types of user terminal data are time characteristics and signal characteristics; the characteristic types of historical communication data are time characteristics and signal characteristics; Perform principal component analysis based on the characteristic types of target link data; After the principal component analysis is completed, efficient features are automatically generated through the autoencoder in deep learning; The generated efficient features are annotated as model output parameter data.

[0013] Preferably, a deep learning model is built in S3 according to the requirements of the low-orbit satellite communication link, including: The low-orbit satellite communication link requirements are obtained from the database, and the performance indicator data in the low-orbit satellite communication link requirements are confirmed; Performance indicator data include throughput, delay data, jitter data, bit error rate, signal-to-noise ratio, link availability, switching time, packet loss rate and power consumption; Confirm the deep learning architecture based on the performance indicator data, including feedforward neural network, convolutional neural network and recurrent neural network; Among them, the feedforward neural network regresses and classifies the performance indicator data; the convolutional neural network processes the spatial characteristics in the performance indicator data; and the recurrent neural network processes the time series data in the performance indicator data. Design the network layer structure of the confirmed deep learning architecture, including the dimension of input features, number of layers, neuron data of each layer and activation function; After the network layer structure design is completed, the neural network model is obtained.

[0014] Preferably, in S4, the neural network model is trained using the model output parameter data, and the model parameters of the neural network model are adjusted according to the model training situation, including: The model output parameter data is divided into data sets, which are training set, validation set and test set; After the data set is divided, model training is performed. The model training process is as follows: First, forward propagation is performed. Forward propagation is to input the training set of model output parameter data into the neural network model. Each layer in the neural network model calculates the training set and generates a predicted output after calculation. The cross entropy loss calculation model is used to calculate the error between the predicted output and the actual output. The actual output is the key feature output data in the model output parameter data. After the calculation is completed, the loss function between the predicted output and the actual output is obtained; Calculate the gradient of the loss function relative to the neural network model through the back-propagation algorithm; And use the gradient descent method to adjust the parameters of the gradient of the neural network model; After each training cycle, the performance of the model is evaluated using the validation set; When the number of model training rounds or the loss function of the validation set evaluation results does not decrease, the model training is completed; According to the model performance evaluated by the validation set, the model whose performance index best meets the preset index is labeled as the link neural network model.

[0015] Preferably, in S5, the link neural network model is transmitted to the control terminal, and real-time link switching is implemented, link switching is automatically performed according to the switching strategy output by the model, and real-time communication link data in the low-orbit satellite communication system is collected, including: Real-time acquisition of current low-orbit satellite communication link data, including satellite status, ground station status, environmental factors and user terminal status; Input the current low-orbit satellite communication link data into the link neural network model and perform switching strategy reasoning; The switching strategy includes link selection, switching timing and switching conditions. The output of the switching strategy is expressed by probability value, classification label or numerical value. Determine whether to execute the switching strategy according to the switching conditions, where the switching conditions include signal strength, delay, and bit error rate; When the switching conditions meet the link switching criteria, the switching strategy is executed. At the same time, key indicators are monitored in real time during the link switching process. If the operation process or performance of the link switching does not meet the expected standards, the link switching is performed again according to the switching strategy.

[0016] Preferably, the real-time communication link data collected in S6 and the performance of the link switching method are evaluated, and optimization and adjustment are performed according to the evaluation results, including: Collect real-time communication link data during and after link switching and confirm performance indicators; Compare the performance indicators during and after the link switching process; Determine whether the performance indicators after link switching have been improved based on the indicator comparison results; The performance indicators that have not been improved are marked as abnormal, and the abnormal problems are identified according to the degree of abnormality.

[0017] Compared with the prior art, the present invention has the following beneficial effects: 1. The low-orbit satellite communication link switching method based on deep learning provided by the present invention can reasonably allocate satellite resources, improve resource utilization, reduce power consumption and cost by optimizing the switching strategy. The link status can be monitored and predicted in real time through the deep learning model, and the optimal switching time and target satellite can be automatically selected, thereby avoiding unnecessary switching and packet loss, and improving the stability and reliability of the communication system.

[0018] 2. The low-orbit satellite communication link switching method based on deep learning provided by the present invention calculates the gradient of the loss function to the neural network model through the back-propagation algorithm, and uses the gradient descent method to adjust the parameters, thereby ensuring that the model parameters of the neural network model can be gradually optimized to the optimal solution. The deep learning model has strong learning and adaptability, and can therefore adapt well to various complex situations that may arise in the switching of low-orbit satellite communication links.

[0019] 3. The low-orbit satellite communication link switching method based on deep learning provided by the present invention uses the powerful computing and learning capabilities of the neural network model to reason about the switching strategy, which can automatically adapt to different link environments and improve the accuracy and flexibility of switching decisions. The switching strategy includes link selection, switching timing and switching conditions, which are expressed by probability values, classification labels or numerical values, and can provide a variety of possible switching schemes. If the operation process or performance of the link switching does not meet the expected standards, the link switching is performed again according to the switching strategy. The process of continuous optimization and adjustment can ensure that the link switching is always in the best state, thereby improving the overall performance and stability of the communication system. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 A schematic diagram of the switching steps of a low-orbit satellite communication link of the present invention; Figure 2 It is a schematic diagram of the low-orbit satellite communication link switching process of the present invention. DETAILED DESCRIPTION

[0021] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0022] In order to solve the problem in the prior art that the acquired low-orbit satellite communication link data is not effectively processed and data feature extracted, resulting in the lack of obvious internal structure and features of the data, please refer to Figure 1 and Figure 2 , this embodiment provides the following technical solutions: A low-orbit satellite communication link switching method based on deep learning includes: S1: Satellite data collection and processing: collect the initial data of the low-orbit satellite communication link, and perform data preprocessing after the collection is completed. After the data preprocessing, the target link data is obtained; Among them, the deep learning model can use a large amount of historical data and real-time data to learn and predict the dynamic changes of low-orbit satellite communication links. Through this learning method, the model can more accurately identify potential link switching points; S2: Extraction of collected data features: Extract key feature data from the target link data and mark the extracted key feature data as model output parameter data; Among them, the automatic feature generation method can more accurately capture the inherent laws of the data and improve the generalization ability of the model; S3: Feature data model construction: A deep learning model is constructed according to the requirements of the low-orbit satellite communication link. After the deep learning model is constructed, a neural network model is obtained; Among them, by optimizing the switching strategy, satellite resources can be reasonably allocated, resource utilization can be improved, and power consumption and costs can be reduced; S4: Model training and optimization: Use the model output parameter data to train the neural network model, adjust the model parameters of the neural network model according to the model training situation, and obtain the link neural network model after the adjustment is completed; Among them, the deep learning model has strong learning and adaptability, so it can adapt well to various complex situations that may arise in the switching of low-orbit satellite communication links; S5: Real-time link switching execution: The link neural network model is transmitted to the control terminal, and the real-time link switching is implemented. According to the switching strategy output by the model, the link switching is automatically executed, and the real-time communication link data in the low-orbit satellite communication system is collected; Among them, the powerful computing and learning capabilities of the neural network model are used to reason about the switching strategy, which can automatically adapt to different link environments and improve the accuracy and flexibility of switching decisions; S6: Switching link performance evaluation: Evaluate the collected real-time communication link data and the performance of the link switching method, and optimize and adjust according to the evaluation results; Among them, by collecting and analyzing link data in real time, the deep learning model can learn the characteristics of these complex environments and make more reasonable switching decisions based on them, which helps to improve the stability and reliability of the network in complex environments.

[0023] The initial data of the low-orbit satellite communication link in S1 will be collected, and the data will be pre-processed after the collection is completed, including: Retrieve initial data of low-orbit satellite communication links from the database; The initial data of the low-orbit satellite communication link includes satellite status data, communication link data, environmental factor data, ground station data, user terminal data and historical communication data; After the initial data collection is completed, data preprocessing is performed, which includes data cleaning, data standardization, data conversion and data enhancement; After data preprocessing is completed, the target link data is obtained.

[0024] Specifically, satellite status data includes satellite orbit parameters, satellite attitude data, and satellite operating speed and acceleration; communication link data includes signal strength, signal quality, signal propagation delay, link capacity, and bandwidth utilization; environmental factor data includes geographic location information, weather conditions, and ionospheric conditions; ground station data includes the location, status, receiving and sending capabilities, and coverage of the ground station; user terminal data includes the location information, moving speed, moving direction, and communication demand data of the user terminal; historical communication data includes previous link switching record business history data, etc. The initial data covers multiple aspects such as satellite status data, communication link data, environmental factor data, ground station data, user terminal data, and historical communication data, ensuring the comprehensiveness of the data. By quickly retrieving the initial data from the database and combining it with an efficient data preprocessing process, it can ensure that the deep learning model can process and analyze the data in real time, which helps to provide timely and accurate suggestions in link switching decisions, thereby optimizing the performance and stability of satellite communication links. The deep learning model can use a large amount of historical data and real-time data to learn and predict the dynamic changes of low-orbit satellite communication links. Through this learning method, the model can more accurately identify potential link switching points and provide users with the best switching strategy.

[0025] Specifically, the initial data of the low-orbit satellite communication link in S1 is collected, and the data is pre-processed after the collection is completed, which also includes The target link data is subjected to quality determination, and a data abnormality alarm is issued when the quality is abnormal, wherein the data amount of each type of initial data contained in the target link data is extracted; the initial data includes satellite status data, communication link data, environmental factor data, ground station data, user terminal data and historical communication data; Extracting the data amount of each type of initial data included in the target link data as the first data amount; Extracting the data volume of each type of initial data before data preprocessing as the second data volume; Obtaining a data anomaly coefficient of each type of initial data using the first data amount and the second data amount; The data anomaly coefficient of each type of initial data is obtained by the following formula: ; Among them, J represents the data anomaly coefficient of each initial data; C 01 and C 02represents the first data amount and the second data amount respectively; n represents the number of data contained in each type of initial data before data preprocessing; m represents the number of data contained in each type of initial data corresponding to the target link data; C hi represents the data volume of the i-th data contained in each corresponding initial data in the target link data; C xi represents the data volume of the i-th data contained in each type of initial data corresponding to the target link data before data preprocessing; Compare the data anomaly coefficient of each initial data with a preset coefficient threshold; When the number of types of initial data whose data anomaly coefficient exceeds a preset coefficient threshold exceeds a preset number threshold, the initial data whose data anomaly coefficient exceeds the preset coefficient threshold is subjected to an abnormality degree determination and abnormality risk determination.

[0026] The technical effect of the above technical solution is: by extracting the data volume of each initial data in the target link data and comparing it with the data volume before data preprocessing, the data anomaly coefficient of each initial data can be accurately calculated. This calculation method not only takes into account the change in data volume, but also combines the change in the number of data before and after data preprocessing, thereby providing a more comprehensive data quality assessment. When the data anomaly coefficient exceeds the preset coefficient threshold, and the number of initial data types of the anomaly coefficient exceeds the preset number threshold, the system will trigger an abnormal alarm. This mechanism ensures the timely discovery and processing of abnormal situations and reduces the potential risks caused by data quality problems. After detecting data anomalies, the system will determine the degree of abnormality and analyze the abnormal risk of abnormal data. This processing method not only helps to quickly locate the problem, but also takes corresponding processing measures according to the degree of abnormality and risk level, thereby improving the efficiency and effect of abnormal processing. By continuously monitoring the quality of the target link data and timely discovering and processing data anomalies, the technical solution can significantly improve the reliability and stability of the system. This helps to ensure the normal operation of the communication link and reduce communication interruptions or performance degradation caused by data quality problems. Through accurate data quality determination and exception processing, the technical solution can avoid unnecessary waste of resources. For example, data of poor quality can be cleaned or replaced in a timely manner to avoid the use of invalid or erroneous data in subsequent processing and improve resource utilization efficiency.

[0027] In summary, the technical effects of this technical solution in terms of performance indicators are mainly reflected in accurate data quality judgment, timely abnormal alarm, efficient abnormality handling, enhanced system reliability and optimized resource utilization, etc. These effects jointly improve the overall performance of the system and user experience.

[0028] Specifically, when the number of types of initial data whose data anomaly coefficient exceeds a preset coefficient threshold for each type of initial data exceeds a preset number threshold, an abnormality degree determination and an abnormality risk alarm are performed on the initial data whose data anomaly coefficient exceeds the preset coefficient threshold for each type of initial data, including: When the number of types of initial data whose data abnormality coefficient exceeds the preset coefficient threshold for each type of initial data exceeds the preset number threshold, the initial data whose data abnormality coefficient exceeds the preset coefficient threshold for each type of initial data is retrieved as the target initial data; Extracting the amount of invalid data and the amount of abnormal data in the target initial data during the data cleaning process; Extract the weight value corresponding to each invalid data and the weight value corresponding to the abnormal data; The weight evaluation coefficient corresponding to the target initial data is obtained by using the weight value corresponding to the invalid data and the weight value corresponding to the abnormal data in combination with the data abnormality coefficient of each initial data; The weight evaluation coefficient corresponding to the target initial data is obtained by the following formula: ; Among them, G represents the weight evaluation coefficient corresponding to the target initial data; x represents the number of invalid data; y represents the number of abnormal data; f xi Indicates the weight value corresponding to the i-th invalid data; f yi represents the weight value corresponding to the i-th abnormal data; f xz represents the middle value of the weight corresponding to x invalid data; f yz represents the middle value of the weight corresponding to y abnormal data; f xm Indicates the weight value corresponding to the maximum value of the data volume of each data contained in the invalid data; f ym Indicates the weight value corresponding to the maximum value of the data volume of each data contained in the abnormal data; f xmax Indicates the maximum value of the weight of invalid data; f ymax Indicates the maximum value of the weight when abnormal data appears; ε indicates the preset minimum value, which is used to prevent the numerator from being 0; J indicates the data abnormality coefficient of each initial data; Comparing the weight evaluation coefficient corresponding to the target initial data with a preset weight evaluation coefficient threshold; When the weight evaluation coefficient corresponding to the target initial data exceeds the preset weight evaluation coefficient threshold, it is determined that the abnormality degree corresponding to the target initial data is serious, and an abnormal risk alarm for the target initial data is issued; When the weight evaluation coefficient corresponding to the target initial data does not exceed the preset weight evaluation coefficient threshold, the amount of invalid data and abnormal data corresponding to the target initial data is used in combination with the weight evaluation coefficient corresponding to the target initial data to determine the degree of abnormality and issue an abnormal risk alarm.

[0029] The technical effect of the above technical solution is: by introducing the weight values ​​of invalid data and abnormal data, and combining the data abnormality coefficient, the weight evaluation coefficient of the target initial data is calculated. This calculation method not only takes into account the amount of data, but also combines the weight and characteristics of the data, making the abnormality assessment more refined and accurate. After detecting the data abnormality, the technical solution not only makes a preliminary judgment on the data abnormality coefficient, but also further subdivides the degree of abnormality according to the weight evaluation coefficient. This multi-level processing method helps to more accurately identify and process abnormal data of different degrees, and improve the responsiveness and processing efficiency of the system. According to the comparison result between the weight evaluation coefficient and the preset weight evaluation coefficient threshold, the system can flexibly trigger an abnormal alarm. When the weight evaluation coefficient exceeds the threshold, the system will determine that the degree of abnormality is serious and alarm, which helps to timely discover and process serious abnormal data and reduce potential risks. The technical solution not only takes into account the number of invalid data and abnormal data, but also combines their weights, the maximum amount of data, and the maximum value of the weight value, and conducts a comprehensive abnormality analysis. This comprehensive analysis method helps to have a deeper understanding of the characteristics and impact of abnormal data, and provides a more valuable reference for subsequent abnormal processing. Through refined anomaly assessment and multi-level anomaly handling, the technical solution can avoid unnecessary waste of resources. This differentiated processing method helps optimize resource utilization and improve the overall performance of the system. By continuously monitoring the quality of data and promptly discovering and handling abnormal data, the technical solution can enhance the stability of the system. This helps ensure the normal operation of the system and the accuracy of the data, and reduces potential risks caused by data quality issues.

[0030] In summary, the technical effects of this technical solution in terms of performance indicators are mainly reflected in refined anomaly assessment, multi-level anomaly handling, flexible anomaly alarm mechanism, comprehensive anomaly analysis, optimized resource utilization and enhanced system stability, etc. These effects jointly improve the overall performance of the system and user experience.

[0031] Specifically, when the weight evaluation coefficient corresponding to the target initial data does not exceed the preset weight evaluation coefficient threshold, the amount of invalid data and abnormal data corresponding to the target initial data is combined with the weight evaluation coefficient corresponding to the target initial data to perform abnormal degree determination and abnormal risk alarm, including: When the weight evaluation coefficient corresponding to the target initial data does not exceed the preset weight evaluation coefficient threshold, calling the coefficient difference between the weight evaluation coefficient corresponding to the target initial data and the preset weight evaluation coefficient threshold; Retrieve the amount of invalid data corresponding to the target initial data; Retrieving the data volume of abnormal data corresponding to the target initial data; Retrieving the weight evaluation coefficient corresponding to the target initial data; Obtaining an abnormal risk evaluation coefficient corresponding to the target initial data by using the data volume of invalid data and the data volume of abnormal data corresponding to the target initial data, the coefficient difference between the weight evaluation coefficient corresponding to the target initial data and the preset weight evaluation coefficient threshold, and the weight evaluation coefficient corresponding to the target initial data; The abnormal risk assessment coefficient corresponding to the target initial data is obtained by the following formula: ; Among them, Y represents the abnormal risk assessment coefficient corresponding to the target initial data; C r01 and C r02 represents the first data volume and the second data volume corresponding to the target initial data; C s and C w represents the amount of invalid data and abnormal data contained in the target initial data; G c represents the coefficient difference between the weight evaluation coefficient corresponding to the target initial data and the preset weight evaluation coefficient threshold; G represents the weight evaluation coefficient corresponding to the target initial data; G y Indicates the preset weight evaluation coefficient threshold; f s and f w Indicates the weighted sum of invalid data and abnormal data contained in the target initial data; Comparing the abnormal risk assessment coefficient corresponding to the target initial data with a preset abnormal risk assessment coefficient threshold; When the abnormal risk assessment coefficient corresponding to the target initial data exceeds a preset abnormal risk assessment coefficient threshold, it is determined that the abnormal degree corresponding to the target initial data is serious, and an abnormal risk alarm corresponding to the target initial data is issued.

[0032] The technical effect of the above technical solution is: by introducing the abnormal risk evaluation coefficient, the technical solution can further divide the abnormal degree of the target initial data in detail when the weight evaluation coefficient does not exceed the preset threshold. This division helps to more accurately identify and handle those situations where the weight evaluation coefficient is not high but the amount of abnormal data and the total weight value are large, thereby improving the response accuracy of the system. The calculation of the abnormal risk evaluation coefficient takes into account a variety of factors, including the amount of invalid data and abnormal data, the coefficient difference between the weight evaluation coefficient and the preset threshold, the total weight value, etc. This comprehensive evaluation method can more comprehensively reflect the abnormal risk of the target initial data and provide a more accurate basis for subsequent abnormal processing. The technical solution flexibly triggers abnormal alarms based on the comparison results between the abnormal risk evaluation coefficient and the preset threshold. When the abnormal risk evaluation coefficient exceeds the preset threshold, the system will determine that the abnormal degree is serious and alarm, which helps to timely discover and handle potential abnormal risks and ensure the stable operation of the system. Through detailed abnormal degree division and comprehensive abnormal risk assessment, the technical solution can avoid unnecessary waste of resources. For data with a relatively mild abnormal degree, relatively simple processing measures can be taken; while for data with a serious abnormal degree, more complex processing measures are required. This differentiated processing method helps optimize resource utilization and improve the overall performance of the system. By continuously monitoring the quality of data and promptly discovering and processing abnormal data, the technical solution can enhance the reliability of the system. This helps ensure the accuracy and integrity of the data, reduce potential risks caused by data quality issues, and improve the stability and availability of the system. The formulas and parameters in the technical solution can be adjusted and optimized according to actual needs to adapt to different scenarios and data characteristics. This scalability and customizability enable the technical solution to be more flexibly applied to various fields and scenarios, improving its practicality and applicability.

[0033] In summary, the technical effects of this technical solution on performance indicators are mainly reflected in more detailed abnormality level division, comprehensive abnormality risk assessment, flexible abnormality alarm mechanism, optimized resource utilization, enhanced system reliability, scalability and customizability, etc. These effects jointly improve the overall performance and user experience of the system, and provide strong support for data quality monitoring and abnormality handling.

[0034] Extract key feature data from the target link data in S2, including: The chi-square test method is used to confirm the feature types in the target link data, which include time features, spatial features and signal features; Among them, the characteristic types of satellite status data are time characteristics and space characteristics; the characteristic types of communication link data are time characteristics and signal characteristics; the characteristic types of environmental factor data are time characteristics and space characteristics; the characteristic types of ground station data are time characteristics and space characteristics; the characteristic types of user terminal data are time characteristics and signal characteristics; the characteristic types of historical communication data are time characteristics and signal characteristics; Perform principal component analysis based on the characteristic types of target link data; After the principal component analysis is completed, efficient features are automatically generated through the autoencoder in deep learning; The generated efficient features are annotated as model output parameter data.

[0035] Specifically, the chi-square test method can be used to accurately confirm the types of various features in the target link data, including time features, spatial features, and signal features, and ensure the consistency of the selected features with the actual data distribution, thereby improving the accuracy and reliability of subsequent analysis. Clearly dividing the data feature types into time features, spatial features, and signal features will help with subsequent principal component analysis and feature extraction. Through principal component analysis, the most representative features can be extracted from the original data. The autoencoder in deep learning can automatically generate efficient features, avoiding the tediousness and subjectivity of manual feature selection. The automatic feature generation method can more accurately capture the inherent laws of the data and improve the generalization ability of the model. The autoencoder can generate more efficient and representative features by continuously optimizing the weight parameters. These features can better reflect the inherent structure and characteristics of the data, thereby improving the prediction accuracy and stability of the model.

[0036] In order to solve the problem that the existing technology does not build a more specific model according to the requirements of low-orbit satellite communication links, resulting in the inability to adapt to unexpected situations when performing link switching, please refer to Figure 1 and Figure 2 , this embodiment provides the following technical solutions: The deep learning model is built in S3 according to the requirements of low-orbit satellite communication links, including: The low-orbit satellite communication link requirements are obtained from the database, and the performance indicator data in the low-orbit satellite communication link requirements are confirmed; Performance indicator data include throughput, delay data, jitter data, bit error rate, signal-to-noise ratio, link availability, switching time, packet loss rate and power consumption; Confirm the deep learning architecture based on the performance indicator data, including feedforward neural network, convolutional neural network and recurrent neural network; Among them, the feedforward neural network regresses and classifies the performance indicator data; the convolutional neural network processes the spatial characteristics in the performance indicator data; and the recurrent neural network processes the time series data in the performance indicator data. Design the network layer structure of the confirmed deep learning architecture, including the dimension of input features, number of layers, neuron data of each layer and activation function; After the network layer structure design is completed, the neural network model is obtained.

[0037] Specifically, the performance indicator data of low-orbit satellite communication links are obtained from the database, including key indicators such as throughput, delay data, jitter data, bit error rate, signal-to-noise ratio, link availability, switching time, packet loss rate and power consumption. The comprehensiveness and accuracy of these data provide a solid foundation for the subsequent deep learning model training, thereby ensuring the reliability and accuracy of the model in practical applications. According to the performance indicator data, the scheme confirms the appropriate deep learning architecture, including feedforward neural network, convolutional neural network and recurrent neural network. The selection of these architectures is targeted and can handle different characteristics in the performance indicator data respectively. Feedforward neural network: regresses and classifies the performance indicator data, can handle simple input-output relationships, and is suitable for preliminary judgment of link status. Convolutional neural network: processes the spatial characteristics in the performance indicator data, can capture the spatial characteristics in the data, and is suitable for analyzing the spatial distribution of the link status. Recurrent neural network: processes the time series data in the performance indicator data, can capture the time dependency in the data, and is suitable for predicting the change of link status over time. The network layer structure design is also carried out, including the dimension of the input feature, the number of layers, the number of neurons in each layer, and the activation function and other key parameters. The reasonable selection of these parameters enables the neural network model to better fit the performance indicator data, improve the generalization ability and robustness of the model, and the deep learning model can be dynamically adjusted according to the changes in the link status, thereby effectively reducing the switching failure rate. At the same time, by optimizing the switching strategy, satellite resources can be reasonably allocated, resource utilization can be improved, power consumption and cost can be reduced, and the real-time monitoring and prediction of the link status by the deep learning model can automatically select the best switching time and target satellite, thereby avoiding unnecessary switching and packet loss, and improving the stability and reliability of the communication system.

[0038] In S4, the neural network model is trained using the model output parameter data, and the model parameters of the neural network model are adjusted according to the model training situation, including: The model output parameter data is divided into data sets, which are training set, validation set and test set; After the data set is divided, model training is performed. The model training process is as follows: First, forward propagation is performed. Forward propagation is to input the training set of model output parameter data into the neural network model. Each layer in the neural network model calculates the training set and generates a predicted output after calculation. The cross entropy loss calculation model is used to calculate the error between the predicted output and the actual output. The actual output is the key feature output data in the model output parameter data. After the calculation is completed, the loss function between the predicted output and the actual output is obtained.

[0039] Calculate the gradient of the loss function relative to the neural network model through the back-propagation algorithm; And use the gradient descent method to adjust the parameters of the gradient of the neural network model; After each training cycle, the performance of the model is evaluated using the validation set; When the number of model training rounds or the loss function of the validation set evaluation results does not decrease, the model training is completed; According to the model performance evaluated by the validation set, the model whose performance index best meets the preset index is labeled as the link neural network model.

[0040] Specifically, by dividing the model output parameter data into training set, validation set and test set, the effective use of data is ensured. Each data set has its specific purpose. The training set is used for model learning, the validation set is used to adjust and optimize the model parameters, and the test set is used to finally evaluate the model performance. This division method helps to avoid overfitting and improve the generalization ability of the model. During the forward propagation process, the training set data is input into the neural network model, and the predicted output is generated through the calculation of each layer. This step ensures that the model can fully learn the characteristics of the data. The cross entropy loss function is used to calculate the error between the predicted output and the actual output. This loss function is particularly effective in the low-orbit satellite communication link switching problem because it can accurately measure the accuracy of the model prediction. The gradient of the loss function to the neural network model is calculated through the back-propagation algorithm, and the gradient descent method is used to adjust the parameters, ensuring that the model parameters of the neural network model can be gradually optimized to the optimal solution. This method can significantly improve the performance of the model, especially when dealing with complex communication link switching problems. After each training cycle, the performance of the model is evaluated using the validation set, which helps to promptly discover problems in the model training process, such as overfitting or underfitting. The model performance can be further optimized by adjusting the training parameters or model structure. When the number of model training rounds reaches the preset value or the loss function of the validation set evaluation result no longer decreases, the model training is completed. This training stop condition ensures that the model will not be over-trained and at the same time ensures the controllability of the training process. The deep learning model has strong learning and adaptability, so it can adapt well to various complex situations that may occur in the switching of low-orbit satellite communication links, such as link quality fluctuations, satellite position changes, etc.

[0041] In order to solve the problem that the link switching mode and requirements are not adjusted according to the actual situation in the prior art, resulting in poor link switching effect, please refer to Figure 1 and Figure 2 , this embodiment provides the following technical solutions: In S5, the link neural network model is transmitted to the control terminal, and real-time link switching is implemented. According to the switching strategy output by the model, link switching is automatically performed, and real-time communication link data in the low-orbit satellite communication system is collected, including: Real-time acquisition of current low-orbit satellite communication link data, including satellite status, ground station status, environmental factors and user terminal status; Input the current low-orbit satellite communication link data into the link neural network model and perform switching strategy reasoning; The switching strategy includes link selection, switching timing and switching conditions. The output of the switching strategy is expressed by probability value, classification label or numerical value. Determine whether to execute the switching strategy according to the switching conditions, where the switching conditions include signal strength, delay, and bit error rate; When the switching conditions meet the link switching criteria, the switching strategy is executed. At the same time, key indicators are monitored in real time during the link switching process. If the operation process or performance of the link switching does not meet the expected standards, the link switching is performed again according to the switching strategy.

[0042] Specifically, by obtaining the current low-orbit satellite communication link data in real time, including satellite status, ground station status, environmental factors and user terminal status, the real-time and accuracy of link switching decisions can be ensured. During the link switching process, key indicators such as signal strength, delay and bit error rate are monitored in real time, and potential problems can be discovered and solved in a timely manner to ensure the smooth progress of link switching. The current low-orbit satellite communication link data is input into the link neural network model, and the switching strategy reasoning is performed using the powerful computing and learning capabilities of the neural network model. It can automatically adapt to different link environments and improve the accuracy and flexibility of switching decisions. The switching strategy includes link selection, switching timing and switching conditions, which are expressed by probability values, classification labels or numerical values, and can provide a variety of possible switching solutions. This helps to find the optimal switching path and timing in a complex link environment. If the operation process or performance of the link switching does not meet the expected standards, the link switching is performed again according to the switching strategy. The continuous optimization and adjustment process can ensure that the link switching is always in the best state and improve the overall performance and stability of the communication system.

[0043] Evaluate the real-time communication link data collected in S6 and the performance of the link switching method, and make optimizations and adjustments based on the evaluation results, including: Collect real-time communication link data during and after link switching and confirm performance indicators; Compare the performance indicators during and after the link switching process; Determine whether the performance indicators after link switching have been improved based on the indicator comparison results; The performance indicators that have not been improved are marked as abnormal, and the abnormal problems are identified according to the degree of abnormality.

[0044] Specifically, by collecting communication link data during and after link switching in real time, the most accurate and timely network status information can be obtained. This helps the deep learning model to understand the network dynamics more accurately, so as to make more reasonable switching decisions. By comparing the performance indicators before and after the switching, the effect of link switching can be intuitively reflected. This comparison not only helps to verify the correctness of the switching decision, but also provides data support for subsequent optimization. By comparing the performance indicators before and after the switching, it can be clearly judged whether the link switching has brought performance improvement. This helps to ensure that each switch is beneficial and avoid the waste of resources caused by unnecessary switching operations. For performance indicators that have not improved performance, abnormal annotations are made, and problems are identified according to the degree of abnormality, so that potential problems in the network can be discovered and solved in a timely manner. This helps to maintain the stability and reliability of the network and improve the user experience. In low-orbit satellite communications, the network topology is complex and changeable. By collecting and analyzing link data in real time, the deep learning model can learn the characteristics of these complex environments and make more reasonable switching decisions based on them, which helps to improve the stability and reliability of the network in complex environments.

[0045] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device.

[0046] While the embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that many changes, modifications, substitutions and variations can be made to the embodiments without departing from the principles and spirit of the invention.

Claims

1. A low-orbit satellite communication link switching method based on deep learning, characterized in that: include: S1: Satellite data collection and processing: collect the initial data of the low-orbit satellite communication link, and perform data preprocessing after the collection is completed. After the data preprocessing, the target link data is obtained; S2: Extraction of collected data features: Extract key feature data from the target link data and mark the extracted key feature data as model output parameter data; S3: Feature data model construction: A deep learning model is constructed according to the requirements of the low-orbit satellite communication link. After the deep learning model is constructed, a neural network model is obtained; S4: Model training and optimization: Use the model output parameter data to train the neural network model, adjust the model parameters of the neural network model according to the model training situation, and obtain the link neural network model after the adjustment is completed; S5: Real-time link switching execution: The link neural network model is transmitted to the control terminal, and the real-time link switching is implemented. According to the switching strategy output by the model, the link switching is automatically executed, and the real-time communication link data in the low-orbit satellite communication system is collected; S6: Switching link performance evaluation: Evaluate the collected real-time communication link data and the performance of the link switching method, and optimize and adjust according to the evaluation results.

2. The method for switching low-orbit satellite communication links based on deep learning according to claim 1, characterized in that: The initial data of the low-orbit satellite communication link will be collected in S1, and the data will be pre-processed after the collection is completed, including: Retrieve initial data of low-orbit satellite communication links from the database; The initial data of the low-orbit satellite communication link includes satellite status data, communication link data, environmental factor data, ground station data, user terminal data and historical communication data; After the initial data collection is completed, data preprocessing is performed, which includes data cleaning, data standardization, data conversion and data enhancement; After data preprocessing is completed, the target link data is obtained.

3. The method for switching low-orbit satellite communication links based on deep learning according to claim 2, characterized in that: The initial data of the low-orbit satellite communication link will be collected in S1, and the data will be pre-processed after the collection is completed, including: The target link data is subjected to quality determination, and a data abnormality alarm is issued when the quality is abnormal, wherein the data amount of each type of initial data contained in the target link data is extracted; the initial data includes satellite status data, communication link data, environmental factor data, ground station data, user terminal data and historical communication data; Extracting the data amount of each type of initial data included in the target link data as the first data amount; Extracting the data volume of each type of initial data before data preprocessing as the second data volume; Obtaining a data anomaly coefficient of each type of initial data using the first data amount and the second data amount; The data anomaly coefficient of each type of initial data is obtained by the following formula: ； Among them, J represents the data anomaly coefficient of each initial data; C 01 and C 02 represents the first data amount and the second data amount respectively; n represents the number of data contained in each type of initial data before data preprocessing; m represents the number of data contained in each type of initial data corresponding to the target link data; C hi represents the data volume of the i-th data contained in each corresponding initial data in the target link data; C xi represents the data volume of the i-th data contained in each type of initial data corresponding to the target link data before data preprocessing; Compare the data anomaly coefficient of each initial data with a preset coefficient threshold; When the number of types of initial data whose data anomaly coefficient exceeds a preset coefficient threshold exceeds a preset number threshold, the initial data whose data anomaly coefficient exceeds the preset coefficient threshold is subjected to an abnormality degree determination and abnormality risk determination.

4. The method for switching low-orbit satellite communication links based on deep learning according to claim 3, characterized in that: When the number of types of initial data whose data abnormality coefficient exceeds the preset coefficient threshold exceeds the preset number threshold, the initial data whose data abnormality coefficient exceeds the preset coefficient threshold is subjected to abnormality degree determination and abnormality risk alarm, including: When the number of types of initial data whose data abnormality coefficient exceeds the preset coefficient threshold for each type of initial data exceeds the preset number threshold, the initial data whose data abnormality coefficient exceeds the preset coefficient threshold for each type of initial data is retrieved as the target initial data; Extracting the amount of invalid data and the amount of abnormal data in the target initial data during the data cleaning process; Extract the weight value corresponding to each invalid data and the weight value corresponding to the abnormal data; The weight evaluation coefficient corresponding to the target initial data is obtained by using the weight value corresponding to the invalid data and the weight value corresponding to the abnormal data in combination with the data abnormality coefficient of each initial data; The weight evaluation coefficient corresponding to the target initial data is obtained by the following formula: ； Among them, G represents the weight evaluation coefficient corresponding to the target initial data; x represents the number of invalid data; y represents the number of abnormal data; f xi Indicates the weight value corresponding to the i-th invalid data; f yi represents the weight value corresponding to the i-th abnormal data; f xz represents the middle value of the weight corresponding to x invalid data; f yz represents the middle value of the weight corresponding to y abnormal data; f xm Indicates the weight value corresponding to the maximum value of the data volume of each data contained in the invalid data; f ym Indicates the weight value corresponding to the maximum value of the data volume of each data contained in the abnormal data; f xmax Indicates the maximum value of the weight of invalid data; f ymax Indicates the maximum value of the weight when abnormal data appears; ε indicates the preset minimum value, which is used to prevent the numerator from being 0; J indicates the data abnormality coefficient of each initial data; Comparing the weight evaluation coefficient corresponding to the target initial data with a preset weight evaluation coefficient threshold; When the weight evaluation coefficient corresponding to the target initial data exceeds the preset weight evaluation coefficient threshold, it is determined that the abnormality degree corresponding to the target initial data is serious, and an abnormal risk alarm for the target initial data is issued; When the weight evaluation coefficient corresponding to the target initial data does not exceed the preset weight evaluation coefficient threshold, the amount of invalid data and abnormal data corresponding to the target initial data is used in combination with the weight evaluation coefficient corresponding to the target initial data to determine the degree of abnormality and issue an abnormal risk alarm.

5. The method for switching low-orbit satellite communication links based on deep learning according to claim 4, characterized in that: When the weight evaluation coefficient corresponding to the target initial data does not exceed the preset weight evaluation coefficient threshold, the amount of invalid data and abnormal data corresponding to the target initial data is combined with the weight evaluation coefficient corresponding to the target initial data to perform abnormal degree determination and abnormal risk alarm, including: When the weight evaluation coefficient corresponding to the target initial data does not exceed the preset weight evaluation coefficient threshold, calling the coefficient difference between the weight evaluation coefficient corresponding to the target initial data and the preset weight evaluation coefficient threshold; Retrieve the amount of invalid data corresponding to the target initial data; Retrieving the data volume of abnormal data corresponding to the target initial data; Retrieving the weight evaluation coefficient corresponding to the target initial data; Obtaining an abnormal risk evaluation coefficient corresponding to the target initial data by using the data volume of invalid data and the data volume of abnormal data corresponding to the target initial data, the coefficient difference between the weight evaluation coefficient corresponding to the target initial data and the preset weight evaluation coefficient threshold, and the weight evaluation coefficient corresponding to the target initial data; The abnormal risk assessment coefficient corresponding to the target initial data is obtained by the following formula: ； Among them, Y represents the abnormal risk assessment coefficient corresponding to the target initial data; C r01 and C r02 represents the first data volume and the second data volume corresponding to the target initial data; C s and C w represents the amount of invalid data and abnormal data contained in the target initial data; G c represents the coefficient difference between the weight evaluation coefficient corresponding to the target initial data and the preset weight evaluation coefficient threshold; G represents the weight evaluation coefficient corresponding to the target initial data; G y Indicates the preset weight evaluation coefficient threshold; f s and f w Indicates the weighted sum of invalid data and abnormal data contained in the target initial data; Comparing the abnormal risk assessment coefficient corresponding to the target initial data with a preset abnormal risk assessment coefficient threshold; When the abnormal risk assessment coefficient corresponding to the target initial data exceeds a preset abnormal risk assessment coefficient threshold, it is determined that the abnormal degree corresponding to the target initial data is serious, and an abnormal risk alarm corresponding to the target initial data is issued.

6. The method for switching low-orbit satellite communication links based on deep learning according to claim 2, characterized in that: Extract key feature data from the target link data in S2, including: The chi-square test method is used to confirm the feature types in the target link data, which include time features, spatial features and signal features; Among them, the characteristic types of satellite status data are time characteristics and space characteristics; the characteristic types of communication link data are time characteristics and signal characteristics; the characteristic types of environmental factor data are time characteristics and space characteristics; the characteristic types of ground station data are time characteristics and space characteristics; the characteristic types of user terminal data are time characteristics and signal characteristics; the characteristic types of historical communication data are time characteristics and signal characteristics; Perform principal component analysis based on the characteristic types of target link data; After the principal component analysis is completed, efficient features are automatically generated through the autoencoder in deep learning; The generated efficient features are annotated as model output parameter data.

7. The method for switching low-orbit satellite communication links based on deep learning according to claim 6, characterized in that: The deep learning model is built in S3 according to the requirements of low-orbit satellite communication links, including: The low-orbit satellite communication link requirements are obtained from the database, and the performance indicator data in the low-orbit satellite communication link requirements are confirmed; Performance indicator data include throughput, delay data, jitter data, bit error rate, signal-to-noise ratio, link availability, switching time, packet loss rate and power consumption; Confirm the deep learning architecture based on the performance indicator data, including feedforward neural network, convolutional neural network and recurrent neural network; Among them, the feedforward neural network regresses and classifies the performance indicator data; the convolutional neural network processes the spatial characteristics in the performance indicator data; and the recurrent neural network processes the time series data in the performance indicator data. Design the network layer structure of the confirmed deep learning architecture, including the dimension of input features, number of layers, neuron data of each layer and activation function; After the network layer structure design is completed, the neural network model is obtained.

8. The method for switching low-orbit satellite communication links based on deep learning according to claim 7, characterized in that: In S4, the neural network model is trained using the model output parameter data, and the model parameters of the neural network model are adjusted according to the model training situation, including: The model output parameter data is divided into data sets, which are training set, validation set and test set; After the data set is divided, model training is performed. The model training process is as follows: First, forward propagation is performed. Forward propagation is to input the training set of model output parameter data into the neural network model. Each layer in the neural network model calculates the training set and generates a predicted output after calculation. The cross entropy loss calculation model is used to calculate the error between the predicted output and the actual output. The actual output is the key feature output data in the model output parameter data. After the calculation is completed, the loss function between the predicted output and the actual output is obtained; Calculate the gradient of the loss function relative to the neural network model through the back-propagation algorithm; And use the gradient descent method to adjust the parameters of the gradient of the neural network model; After each training cycle, the performance of the model is evaluated using the validation set; When the number of model training rounds or the loss function of the validation set evaluation results does not decrease, the model training is completed; According to the model performance evaluated by the validation set, the model whose performance index best meets the preset index is labeled as the link neural network model.

9. The method for switching low-orbit satellite communication links based on deep learning according to claim 8, characterized in that: In S5, the link neural network model is transmitted to the control terminal, and real-time link switching is implemented. According to the switching strategy output by the model, link switching is automatically performed, and real-time communication link data in the low-orbit satellite communication system is collected, including: Real-time acquisition of current low-orbit satellite communication link data, including satellite status, ground station status, environmental factors and user terminal status; Input the current low-orbit satellite communication link data into the link neural network model and perform switching strategy reasoning; The switching strategy includes link selection, switching timing and switching conditions. The output of the switching strategy is expressed by probability value, classification label or numerical value. Determine whether to execute the switching strategy according to the switching conditions, where the switching conditions include signal strength, delay, and bit error rate; When the switching conditions meet the link switching criteria, the switching strategy is executed. At the same time, key indicators are monitored in real time during the link switching process. If the operation process or performance of the link switching does not meet the expected standards, the link switching is performed again according to the switching strategy.

10. The method for switching low-orbit satellite communication links based on deep learning according to claim 9, characterized in that: Evaluate the real-time communication link data collected in S6 and the performance of the link switching method, and make optimizations and adjustments based on the evaluation results, including: Collect real-time communication link data during and after link switching and confirm performance indicators; Compare the performance indicators during and after the link switching process; Determine whether the performance indicators after link switching have been improved based on the indicator comparison results; The performance indicators that have not been improved are marked as abnormal, and the abnormal problems are identified according to the degree of abnormality.

Images