Unmanned aerial vehicle group digital twinning real-time enhancement method based on incremental data

Abstract

The invention relates to the technical field of digital twinning, and discloses an unmanned aerial vehicle group digital twinning real-time enhancement method based on incremental data, comprising the following steps: S1, receiving multi-source time sequence data of an unmanned aerial vehicle group; s2, dividing the data into observed scene data and new scene data according to scene features; s3, calculating the maximum processable data volume based on the real-time demand; s4, screening high-value data through a self-adaptive data selector, and constructing an incremental training data set; and S5, performing real-time incremental training on the digital twin by using the incremental training data set, and updating model parameters. Real-time incremental training is executed at a ground control station, an updated model can be deployed to an unmanned aerial vehicle terminal in real time, historical data of an observed scene is reserved, new scene data is converted into observed data through data migration in training, subsequent screening is optimized, the accuracy of real-time state prediction in a dynamic environment is improved, and the real-time state prediction efficiency is improved. And the requirements of a networked air-ground cooperative system are met.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of digital twinning, in particular to a method for real-time enhancement of a digital twin of a UAV group based on incremental data. BACKGROUND

[0002] Digital twinning technology realizes monitoring and simulation by connecting virtual and physical systems, and is a key technology for improving intelligent level in networked air-ground collaborative UAV groups and other dynamic systems.

[0003] Existing digital twinning model enhancement mostly adopts a learning-driven method, which updates the model using new data through incremental learning, but the traditional method is mostly based on initial limited data construction, and relies on full data retraining or simple incremental training when updating.

[0004] However, the application inventors have found that the above-mentioned technology is prone to performance degradation when facing new scene data in a dynamic environment, and cannot balance the adaptation to new data and the preservation of historical knowledge, resulting in insufficient real-time prediction accuracy. SUMMARY

[0005] To overcome the deficiencies of the prior art, the application provides a method for real-time enhancement of a digital twin of a UAV group based on incremental data, which solves the problem of insufficient real-time prediction accuracy caused by performance degradation when facing new scene data and the inability to balance the adaptation to new data and the preservation of historical knowledge.

[0006] To achieve the above-mentioned purposes, the application is implemented by the following technical scheme: a method for real-time enhancement of a digital twin of a UAV group based on incremental data, comprising the following steps:

[0007] S1, receiving multi-source time series data of the UAV group;

[0008] S2, dividing the data into observed scene data and new scene data according to scene characteristics;

[0009] S3, calculating the maximum processable data amount based on real-time requirements;

[0010] S4, filtering high-value data through an adaptive data selector to construct an incremental training data set;

[0011] S5, using the incremental training data set to perform real-time incremental training on the digital twin and update the model parameters.

[0012] By adopting the technical scheme, the multi-source time series data of the unmanned aerial vehicle group is received first, and is divided into observed and new scene data according to scene characteristics, then the maximum processable data amount is determined based on real-time requirements, high-value data is selected through an adaptive selector to construct an incremental training data set, and finally the model is updated through real-time incremental training. This process can process new scene data specifically to avoid performance degradation, and balance the adaptation to new data and the preservation of historical knowledge through incremental training, effectively improving the accuracy of real-time prediction in a dynamic environment.

[0013] Preferably, the filtering process of the adaptive data selector includes:

[0014] extracting multi-dimensional features of the time series data;

[0015] calculating data value weights based on feature interaction;

[0016] selecting high-value data according to the value weights and real-time requirements.

[0017] Preferably, the calculation of the data value weights adopts a reinforcement learning strategy, and a reward function is used to balance the model performance and data processing efficiency, and the reward function is defined as:

[0018]

[0019] wherein, Ψ (a d ) is a model performance evaluation item, and φ (a d ) is the volume of the incremental training data set.

[0020] Preferably, the calculation formula of the model performance evaluation item Ψ (a d ) is:

[0021] Ψ (a d ) = αΨ n (a d ) + βΨ o (a d ) ;

[0022] wherein, Ψ n (a d ) represents a new scene performance reward, Ψ o (a d ) represents an observed scene performance reward, and α and β are adjustable weight coefficients.

[0023] Preferably, the calculation formula of the maximum processable data amount is:

[0024] N c = n b × s b ;

[0025] wherein, n bs is the maximum number of training batches b s is the single batch data volume.

[0026] Preferably, the scene division process comprises:

[0027] Maintain a global scene dictionary and dynamically add new scene entries;

[0028] Manage observed scene data and new scene data based on scene labels.

[0029] Preferably, it also includes:

[0030] After the incremental training is completed, the new scene data is migrated to the observed scene data management module, and the data usage frequency label is updated to optimize subsequent screening.

[0031] Preferably, the method is applied to a networked air-ground collaborative unmanned aerial vehicle swarm system, wherein:

[0032] The incremental training process is performed at the ground control station, and the updated digital twin model is deployed to the unmanned aerial vehicle terminal in real time.

[0033] Preferably, the high-value data preferentially meets the data whose prediction error exceeds a threshold in the new scene data or the data close to the decision boundary in the observed scene.

[0034] Preferably, the method supports collaborative data enhancement of multiple unmanned aerial vehicle swarms, improves model generalization ability by integrating swarm data, and is suitable for real-time state prediction in dynamic environments.

[0035] The present application provides a real-time enhancement method for unmanned aerial vehicle swarm digital twin based on incremental data. It has the following advantages:

[0036] 1. The present application is executed at the ground control station through real-time incremental training, and the updated model can be deployed to the unmanned aerial vehicle terminal in real time. The data is divided into observed and new scene data through scene division. During incremental training, data whose prediction error exceeds a threshold in the new scene is preferentially screened, and the historical data of the observed scene is also retained. During training, new scene data is converted to observed data through data migration, optimizing subsequent screening. It not only adapts to new scenes but also retains historical knowledge, improves the accuracy of real-time state prediction in dynamic environments, and meets the needs of networked air-ground collaborative systems.

[0037] 2. The present application selects high-value data through an adaptive data selector, balances model performance and data processing efficiency by combining reinforcement learning reward functions, optimizes training through maximum processable data volume calculation, and improves efficiency.

[0038] 3. The present application supports collaborative data enhancement of multiple unmanned aerial vehicle swarms, integrates swarm data to improve model generalization ability, performs incremental training at the ground control station, and deploys the updated model to the terminal in real time, enhancing practicality. Attached Figure Description

[0039] Figure 1 This is a flowchart of the real-time augmentation method for UAV swarm digital twins based on incremental data according to the present invention.

[0040] Figure 2 This is a module connection diagram of the real-time enhancement method for UAV swarm digital twins based on incremental data according to the present invention. Detailed Implementation

[0041] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0042] Please see the appendix Figure 1 and Figure 2 This invention provides a real-time augmentation method for unmanned aerial vehicle (UAV) swarm digital twins based on incremental data, comprising the following steps:

[0043] S1. Receive multi-source time series data from the drone swarm;

[0044] S2. Based on scene characteristics, the data is divided into observed scene data and new scene data;

[0045] S3. Calculate the maximum amount of data that can be processed based on real-time demand;

[0046] S4. Use an adaptive data selector to filter high-value data and build an incremental training dataset;

[0047] S5. Use the incremental training dataset to perform real-time incremental training on the digital twin and update the model parameters.

[0048] Specifically, in S1, real flight scenarios are used as input, and then high-value time series data are continuously filtered through S2-S4. In S5, an incremental learning strategy is combined to realize online updates of model weights and to perform online incremental updates of the UAV energy consumption digital twin model.

[0049] First, the data receiving module DR in S1 receives multi-source time series data from the UAV swarm. The data format follows:

[0050]

[0051]

[0052] Table I. Definitions of Variables Related to UAV Flight Status and Energy Consumption

[0053] Subsequently, the scene judgment module SJ constructs a global scene dictionary according to the rule of "keeping the same group of flight parameters for 2 minutes", and if the current data is completely consistent with any record in the history, it is classified into the observed scene data module ODM, otherwise it is classified into the new scene data module NDM and a new entry is added to the dictionary. The scene judgment module SJ divides the data into two categories of observed and new scenes according to the rule that the flight parameters remain unchanged for 2 minutes in succession, and maintains a global scene dictionary. The observed scene data module ODM is used to store the data of the historical scenes that have been observed, and records the number of times each data is selected for training. The new scene data module NDM is used to store the new scene data that appears for the first time, and each data is labeled with a prediction error label.

[0054] After classification, the data volume estimation module DVE calculates the maximum processable data volume Nc according to the real-time demand Tc and the training device computing power, and the formula is

[0055] N c =n b ×s b ;

[0056] Where n b =(T c -T d )÷t d , T d is the AdaSor decision time, t d is the average time consumption of a single batch training, and s b is the batch size.

[0057] AdaSor is a lightweight adaptive data selector that selects high-value samples from historical and new scene data under the constraint of N c . The core is a multi-head attention mechanism: first, the 20x7-dimensional time series input is encoded into 128-dimensional features hi by the recurrent neural network RNN, then the related vectors p i , q i of the average feature y n and the boundary feature u n of each candidate data are calculated by the multi-head attention MHA, and finally the selected probability δ i , η i is output by softmax, where (δ i +η i =1). In the training stage, AdaSor optimizes the reward function with reinforcement learning strategy πθ:

[0058]

[0059] Where a d represents the data selection action, and φ(a d) to augment the volume of the training dataset, Ψ(a d ) = αΨ n (a d ) + βΨ o (a d ), where Ψ n (a d ) and Ψ o (a d ) represent the mean squared error on the new scenario and the observed scenario validation set respectively, and α, β are set according to whether the new scenario appears: for example, α = 0.1, β = 1 when the new scenario appears, and α = 1, β = 0 otherwise, and AdaSor generates the incremental training dataset G i according to the threshold δ c > 0.5 and the limit N θ , whose volume is where the recurrent neural network RNN first encodes the 20 x 7 time series segment into a 128-dimensional feature vector to capture the time correlation, and then uses the multi-head attention MHA to calculate the relevance of each candidate data to the average feature or boundary feature of the scenario, and finally gives the selection probability.

[0060] After G θ is sent to the training device, the TCN updates the weight ω in an incremental manner, and the loss function remains the mean squared error:

[0061]

[0062] After training is completed, the new scenario data is migrated to the observed scenario data management module ODM, the usage frequency is recorded, and the scenario dictionary is updated, and the next round of cycle is repeated. The whole process is implemented in the air-ground collaborative unmanned aerial vehicle system, and the unmanned aerial vehicle sends the flight record to the ground control station in real time through the wireless link.

[0063] The screening process of the adaptive data selector includes:

[0064] extracting multi-dimensional features of the time series data;

[0065] calculating data value weights based on feature interaction;

[0066] selecting high-value data according to the value weights and real-time demand.

[0067] The calculation of the data value weight adopts a reinforcement learning strategy, and the reward function is used to balance the model performance and data processing efficiency, and the reward function is defined as:

[0068]

[0069] where Ψ(a d ) is a model performance evaluation item, and φ(a d) is the volume of the augmented training dataset.

[0070] The model performance evaluation item Ψ(a d ) is calculated as follows:

[0071] Ψ(a d ) = αΨ n (a d ) + βΨ o (a d );

[0072] wherein Ψ n (a d ) represents a new scene performance reward, Ψ o (a d ) represents an observed scene performance reward, and α and β are adjustable weight coefficients.

[0073] The calculation formula of the maximum processable data amount is as follows:

[0074] N c = n b × s b ;

[0075] wherein n b is the maximum training batch number, and s b is the single batch data amount.

[0076] Specifically, AdaSor continuously performs feature extraction, value weight calculation, and high-value sample selection. The input is the multi-source time series uploaded by the unmanned aerial vehicle group in real time. First, the recurrent neural network (RNN) encodes the 20x7-dimensional time sequence segment into a 128-dimensional vector h i , and the non-time sequence label is mapped to the same dimension. After splicing, the two are sent to 6 heads of 512-dimensional MHA, and the relevant vector p i and the irrelevant vector q i are output. After softmax, the selection probability δ i and η i are obtained. Second, the reinforcement learning agent is in the state s d , and the reward function is directly adopted as follows:

[0077]

[0078] Since the UAV will encounter new flight height, speed, and load combinations during the task, AdaSor must identify new scenarios, label high-error samples, and complete incremental training under limited computing power. To this end, a complete flight profile is divided into 6 consecutive stages (see Table II), each of which triggers a scenario dictionary update, sample selection, TCN incremental training, model return, and a closed loop of this sequence, and controls whether to introduce new scenarios or multi-UAV cooperation through conditional labeling, thereby systematically evaluating the continuous learning ability of AdaSor. Among them, the time convolution network TCN is used as an example of the UAV energy consumption digital twin model to update the weight in an incremental manner.

[0079] Phase Flight Altitude Speed Load Conditions 0 8 25 4 0 \ 1 10 25 6 0 S&P 2 \ \ \ \ S / M&A 3 23 50 12 250 S&P 4 80 50 12 0 S&P 4 95 50 10 500 M&P 5 98 25 12 500 \ 5 100 50 12 500 \ 5 101 50 8 500 M&P

[0080] Table II: UAV flight phase parameters and test conditions

[0081] Conditional labeling:

[0082] S&P means "single UAV + new scenario appears";

[0083] M&P means "multi-UAV + new scenario appears";

[0084] S / M&A means "single or multi-UAV + no new scenario appears";

[0085] This dataset is used to train the AdaSor strategy network sequentially and evaluate its incremental learning effect when new scenarios are introduced step by step.

[0086] The scenario division process includes:

[0087] Maintain a global scenario dictionary and dynamically add new scenario entries;

[0088] Manage observed scenario data and new scenario data based on scenario labels.

[0089] Also includes:

[0090] After incremental training is complete, migrate new scenario data to the observed scenario data management module and update the data usage frequency label to optimize subsequent screening.

[0091] The method is applied to a networked air-ground collaborative UAV swarm system, wherein:

[0092] The incremental training process is performed on the ground control station, and the updated digital twin model is deployed to the UAV terminal in real time.

[0093] The high-value data prioritizes data with predicted errors exceeding the threshold in new scenario data or data close to the decision boundary in observed scenarios.

[0094] The method supports cooperative data enhancement of multiple UAV groups, improves model generalization ability by integrating group data, and is suitable for real-time state prediction in dynamic environments.

[0095] Specifically, scene division is performed by a scene judgment module SJ, and the operation steps are as follows:

[0096] A global scene dictionary is maintained, the key is a 10-tuple composed in Table I, and the value is the first occurrence timestamp. Every time a UAV record is received, the 10-tuple key is calculated. If the key already exists, the record is stored in the observed scene data management module ODM, and the usage frequency count corresponding to the key in the observed scene data management module ODM is incremented by 1. If the key does not exist, a new entry is added to the dictionary, the record is stored in the new scene data management module NDM, and the initial error label is recorded in the new scene data management module NDM. After incremental training is completed, AdaSor migrates all records in the new scene data management module NDM of this round to the observed scene data management module ODM. The migration updates the timestamp of each record to the current round number and resets the usage frequency count to 1, providing the next round of AdaSor with a frequency and error double standard for screening.

[0097] In this embodiment, the system is deployed in a networked air-ground collaborative UAV group system. Three UAVs transmit Table I format data at a frequency of 2Hz to the ground control station through a 5GHz wireless link. The ground station performs incremental training: AdaSor performs incremental training on the observed scene data management module ODM under the constraint of

[0098] N c =n b ×s b ;

[0099] The updated TCN weight file is returned to the onboard JetsonNano through the same link, replacing the old model and completing the real-time deployment loop. The digital twin model DTM refers to a digital model constructed in a virtual space that is real-time mapped and synchronized with the physical entity. This model continuously updates its state by receiving real-time data from the physical entity, achieving monitoring, prediction, and optimal control of the physical object.

[0100] For records in the new scene data management module NDM, if the absolute value of its error label is greater than 5W, it is forced to be selected. For records in the observed scene data management module ODM, if the cosine distance between its features and the average features of the corresponding scene is greater than 0.15, it is considered to be of high value and is selected in descending order of delta. i

[0101] ​Multi-machine cooperation: each unmanned aerial vehicle independently collects data, all data are merged into a group data set on the ground station, AdaSor uniformly evaluates the value weight and constructs a single IL data set, realizes one-time training and simultaneously improves the generalization performance of the group digital twin model DTM, and is suitable for real-time state prediction of dynamic changes of wind speed, load and height.

[0102] While embodiments of the present application have been shown and described, it is to be understood that the embodiments described are merely divergences, modifications, replacements and variations of the embodiments, and the scope of the present application is defined by the appended claims and their equivalents.

Claims

1. A method for real-time enhancement of drone swarm digital twin based on incremental data, characterized in that, The method comprises the following steps: S1, receiving multi-source time series data of a UAV group; S2, dividing the data into observed scene data and new scene data according to scene characteristics; S3, calculating the maximum processable data amount based on real-time requirements; S4, screening high-value data through an adaptive data selector to build an incremental training data set; S5, using the incremental training data set to perform real-time incremental training on a digital twin and updating model parameters.

2. The incremental data based swarm of drones digital twin real-time augmentation method of claim 1, wherein, The screening process of the adaptive data selector comprises: extracting multi-dimensional features of the time series data; calculating data value weights based on feature interaction; selecting high-value data according to the value weights and real-time requirements.

3. The incremental data based swarm of drones digital twin real-time augmentation method of claim 2, wherein, The calculation of the data value weights adopts a reinforcement learning strategy, and a reward function is used to balance model performance and data processing efficiency, and the reward function is defined as: where Ψ(a d ) is a model performance evaluation term, and φ(a d ) is the volume of the incremental training dataset.

4. The incremental data based swarm of drones digital twin real-time augmentation method of claim 3, wherein, The model performance evaluation item Ψ(a d ) is calculated by the following formula: Ψ(a d ) = aΨ n (a d ) + bΨ o (a d ) where, Ψ n (a d ) represents the new scene performance reward, Ψ o (a d ) represents the observed scene performance reward, and α and β are adjustable weight coefficients.

5. The incremental data based swarm of drones digital twin real-time augmentation method of claim 2, wherein, The calculation formula of the maximum processable data amount is: N c = n b × s b ; Wherein, n b is the maximum training batch number, s b is the single batch data volume.

6. The incremental data based swarm of drones digital twin real-time augmentation method of claim 1, wherein, The scene division process comprises: maintaining a global scene dictionary and dynamically adding new scene entries; managing observed scene data and new scene data based on scene labels.

7. The incremental data based swarm of drones digital twin real-time augmentation method of claim 1, wherein, Further comprising: after the incremental training is completed, migrating the new scene data to the observed scene data management module and updating the data usage frequency label to optimize subsequent screening.

8. The incremental data based swarm of drones digital twin real-time augmentation method of claim 1, wherein, The method is applied to a networked air-ground collaborative UAV group system, wherein: the incremental training process is performed at a ground control station, and the updated digital twin model is deployed to a UAV terminal in real time.

9. The incremental data based swarm of drones digital twin real-time augmentation method of claim 1, wherein, The high-value data preferentially meets data whose prediction error exceeds a threshold in the new scene data or data close to a decision boundary in the observed scene.

10. The incremental data based swarm of drones digital twin real-time augmentation method of claim 1, wherein, The method supports collaborative data enhancement of multiple UAV groups, improves model generalization ability by integrating group data, and is suitable for real-time state prediction in dynamic environments.

Images