Real-Time Digital Twin Enhancement Platform Addresses Critical UAV Swarm Adaptability

This figure illustrates the EnFlexiTwin architecture - a flexible digital twin framework for drone swarm operations that bridges physical drone operations with digital simulation and learning systems.

Architecture Overview

The system operates across three domains:

Physical Spaces (Top)

Shows actual drones operating in the real world, sending sensor data (S1) to the system.

Virtual-Physical Interaction Layer (Middle - Pink/Red Section)

This is the core processing engine with several interconnected modules:

1. Data Receiving Module (DR): Ingests real-time sensor data (S2) from physical drones

2. Scenario Judgement Module (SJ): Analyzes incoming data to determine if the scenario is:

   • Previously observed (existing scenario)
   • Novel (new scenario requiring learning)

3. Data Management:

   • Observed-scenario Data Management Module (ODM): Handles known scenarios (S3)
   • New-scenario Data Management Module (NDM): Processes novel situations (S3)

4. Incremental Learning Components:

   • ΔAdvisor: Selects appropriate datasets for incremental learning (S4)
   • DVE (Dataset Volume Estimation): Estimates required training data volume (S5.1, S5.2)
   • Both feed into the learning process with time constraints

5. Bidirectional Communication:

   • S7 flows show data exchange between observed and new scenario modules
   • S8 enables virtual-physical interaction
   • S6 connects to digital twins

Digital Spaces (Bottom - Green Section)

Contains Digital Twins that run simulations including:

• Flying energy management models: Optimize drone power consumption
• Pose tracking models: Monitor drone position and orientation

The dotted lines (S5.1) show how digital twin simulations inform the learning process.

Key Innovation

The architecture dynamically adapts by:

• Distinguishing between known and unknown scenarios in real-time
• Incrementally learning from new scenarios without retraining entire models
• Using digital twins to reduce physical testing requirements
• Estimating how much new data is needed for effective learning

This creates a flexible system that can handle both routine operations (using existing models) and novel situations (through adaptive learning), making it particularly valuable for autonomous drone swarms operating in unpredictable environments.

Real-Time Digital Twin Enhancement Platform Addresses Critical UAV Swarm Adaptability Challenge

BLUF (Bottom Line Up Front)

Researchers at Northwestern Polytechnical University have developed EnFlexiTwin, a digital twin enhancement platform that enables UAV swarms to adapt their operational models in real-time using incremental learning, addressing a fundamental challenge in autonomous aerial transportation systems where pre-trained models degrade when encountering new flight scenarios.

Background and Problem Statement

Unmanned aerial vehicles are rapidly emerging as critical components of future intelligent transportation systems, with applications spanning urban air mobility, emergency response, and low-altitude logistics. However, current digital twin modeling approaches face significant limitations when UAVs encounter operational conditions not present in their initial training datasets.

Research published in IEEE Transactions on Intelligent Transportation Systems demonstrates that digital twin models for UAV power consumption can experience substantial performance degradation—up to 40% error increase—when new flight scenarios emerge that differ from original training conditions. This degradation occurs because comprehensive pre-deployment data collection covering all possible operational states is impractical in dynamic transportation environments.

"The initial need to collect comprehensive data that reflects all possible states, environments, and mission scenarios" represents a critical bottleneck, according to the research team led by Mengjie Lee and Yining Zhu. "In dynamic transportation contexts, such exhaustive data acquisition is often impractical, leading to models built on limited observations."

Technical Solution: EnFlexiTwin Architecture

The EnFlexiTwin platform addresses these challenges through a modular architecture comprising six key components integrated between physical UAV swarms and their digital counterparts:

Data Management Infrastructure: The system employs a Data Receiving module that synchronizes time-series data from multiple flying UAVs through centralized buffered data pools, implementing conflict resolution policies to handle transmission delays and out-of-order packets. A Scenario Judgment module classifies flight data into specific operational profiles based on altitude, payload, and speed characteristics.

Intelligent Data Selection: At the core of EnFlexiTwin is AdaSor (Adaptive Data Selector), a lightweight attention-based neural network that autonomously identifies high-value time-series data for model enhancement. AdaSor formulates data selection as a single-step reinforcement learning problem, learning adaptive policies that balance performance on historical scenarios with rapid adaptation to emerging conditions.

The system constructs selection policies through multi-head attention mechanisms that capture relationships between new and historical data. For each candidate sample, AdaSor generates features including predicted values, prediction errors, and usage frequency—then calculates selection probabilities that maximize a composite reward function balancing accuracy and data efficiency.

Real-Time Constraint Management: A Data Volume Estimation module dynamically calculates the maximum training dataset size permissible within specified time constraints, accounting for inference latency, sampling overhead, and computational capacity. This ensures model updates complete within operational deadlines—critical for maintaining continuous UAV service.

Experimental Validation

The research team conducted extensive validation using real-world flight data from 209 flights of a DJI Matrice M100 quadrotor performing small-package delivery missions, totaling 10 hours 45 minutes of flight time across 65 kilometers. The dataset encompasses varied operational parameters including commanded ground speeds, payload masses, and cruise altitudes collected from April through October 2019.

Testing employed a temporal convolutional network (TCN) as the baseline digital twin model for power consumption prediction. The TCN processes sensor data including GPS, IMU, voltage/current measurements, and anemometer readings to estimate real-time power consumption.

Simulation Results: In time-varying simulations with 8-UAV swarms over 20-minute trajectories, EnFlexiTwin achieved win rates of 70-95% compared to nine baseline methods across varying real-time constraints (30-120 seconds). The platform maintained zero version delay—meaning models always reflected the most current enhancement—while baseline approaches often fell behind due to excessive training data volumes.

When real-time requirements increased from 30 to 120 seconds, EnFlexiTwin demonstrated consistent mean squared error (MSE) reduction from approximately 100 to 60 for global performance, and from 140 to 80 for new scenario adaptation. Competing methods showed substantially higher errors and version delays exceeding 15 updates.

Swarm Scalability: Testing across UAV swarm scales from 3 to 20 vehicles revealed EnFlexiTwin's ability to leverage multi-platform data effectively. As swarm size increased, the platform's relative performance improved significantly—particularly for adaptation to new scenarios—while maintaining zero version delay. Baseline methods experienced growing version delays as data volumes increased with larger swarms.

Physical UAV Validation: The research team deployed EnFlexiTwin on a practical networked air-ground cooperative system comprising three AMOVLAB P200-A2-TX2 quadrotors equipped with Jetson Nano edge computing modules, connected wirelessly to an NVIDIA Tesla V100-equipped ground station. Physical flight experiments with 30-90 second real-time constraints confirmed simulation findings, with EnFlexiTwin achieving 20-47% lower MSE on global performance and 14-46% lower MSE on new scenario adaptation compared to baseline approaches.

Technical Innovations and Contributions

Attention-Based Data Selection: Ablation studies confirmed that AdaSor's multi-head attention mechanism substantially outperforms multi-layer perceptron alternatives of comparable capacity, reducing error fluctuations and improving stability. The attention architecture better captures temporal dependencies and contextual relationships in sequential flight data.

Catastrophic Forgetting Mitigation: Analysis of the first 10 minutes of testing demonstrates EnFlexiTwin's ability to dynamically balance new scenario data with selectively sampled historical data. The system consistently prioritized recently emerged scenario data while retaining representative historical samples, achieving both rapid adaptation and long-term knowledge preservation—a key advantage over naive incremental learning approaches.

Reward Function Design: The reinforcement learning reward formulation balances generalization to new environments (weighted by α) with retention of historical performance (weighted by β), plus a logarithmic penalty for excessive data selection. Testing across weight configurations confirmed that balanced weighting (α=1, β=1) achieves the most stable and lowest prediction errors across all scenarios.

Broader Research Context and Related Work

The EnFlexiTwin research addresses limitations in both mechanism-driven and learning-driven approaches to digital twin model enhancement. Mechanism-driven methods—which integrate predefined mathematical models based on physics or engineering principles—provide accuracy but require extensive domain expertise and struggle with unforeseen scenarios. Learning-driven approaches offer flexibility but existing implementations either retrain models from scratch (computationally prohibitive) or suffer from catastrophic forgetting when adapting to new conditions.

Recent parallel research in related domains has explored similar challenges:

Industrial Digital Twins: Costa et al. (2024) proposed adaptive digital twins for pressure swing adsorption systems integrating feedback tracking, online learning, and uncertainty assessment. However, their approach retrains complete models when performance degrades below thresholds—a time-intensive process unsuitable for UAV real-time constraints.

Infrastructure Monitoring: Jiang et al. (2023) developed digital twin frameworks for fatigue lifecycle management of steel bridges by integrating probabilistic multi-scale deterioration models. While effective for slowly-evolving infrastructure systems, such mechanism-driven approaches lack the rapid adaptability required for dynamic aerial platforms.

Energy Systems: Wu et al. (2024) investigated online learning methods for self-updating digital twin models of power transformer temperature prediction, employing incremental training on data between update periods. Their approach, however, does not address data volume optimization or multi-source swarm data integration.

Incremental Learning Foundations

EnFlexiTwin builds upon advances in incremental learning (IL)—a machine learning paradigm where models learn continuously from new data without forgetting previously acquired knowledge. While IL research has matured in computer vision and natural language processing domains, application to time-varying digital twin enhancement represents relatively unexplored territory.

Traditional IL dataset construction approaches include:

• Random sampling: Uniform selection without considering data utility
• Mean feature sampling: Selecting samples closest to class averages
• Boundary-based sampling: Prioritizing data near decision boundaries
• Entropy-based sampling: Maintaining information entropy balance

EnFlexiTwin's AdaSor advances beyond these heuristics by learning data selection policies through experience with historical model behavior, automatically adapting to scenario-specific patterns without manual tuning.

Operational Implications for UAV Transportation

The research demonstrates critical capabilities for emerging UAV-based transportation applications:

Last-Mile Logistics: Small package delivery operations encounter highly variable conditions including diverse payload weights, delivery locations at different altitudes, and changing weather patterns. EnFlexiTwin enables delivery fleets to continuously refine energy management models as operational experience accumulates, improving route planning accuracy and reducing safety margins.

Urban Air Mobility: Passenger-carrying eVTOL aircraft will operate across varied urban microclimates, building wake turbulence, and traffic density patterns. Real-time digital twin enhancement allows vehicles to adapt performance models to local conditions while maintaining safety-critical accuracy.

Emergency Response: Disaster relief UAV operations face rapidly evolving scenarios with terrain damage, infrastructure changes, and dynamic mission parameters. The platform's ability to update models within 30-60 second windows enables responsive adaptation to unprecedented conditions.

Swarm Coordination: Multi-UAV cooperative missions benefit from EnFlexiTwin's native support for swarm data integration. The platform aggregates experience across fleet members, enabling faster collective learning than individual vehicle adaptation.

Future Research Directions

The research team identifies several priority areas for advancing digital twin enhancement capabilities:

Heterogeneous Swarm Support: Current implementation focuses on homogeneous UAV fleets where all vehicles share common digital twin architectures. Extension to mixed vehicle types—such as coordinated fixed-wing and rotorcraft operations—requires handling architectural diversity while maintaining efficient knowledge transfer.

Communication-Constrained Environments: Intermittent connectivity during beyond-visual-line-of-sight operations necessitates distributed learning approaches where individual UAVs perform local model updates, later synchronizing with ground-based enhancement infrastructure when links permit.

Non-Stationary Environment Adaptation: Long-duration missions may encounter gradual environmental drift (seasonal weather patterns, component wear) alongside abrupt scenario changes. Hybrid enhancement strategies combining continuous background adaptation with event-triggered rapid updates could improve robustness.

Multi-Modal Sensor Integration: Current implementation processes predominantly numerical time-series data. Incorporating visual data from onboard cameras and LiDAR for environment-aware model enhancement represents a promising extension, particularly for obstacle avoidance and terrain classification tasks.

Safety-Critical Verification: Deployment in certified transportation applications requires formal verification that enhanced models maintain safety properties. Research into provable bounds on model behavior after incremental updates could enable regulatory acceptance.

Industry and Policy Context

The research arrives as regulatory frameworks for advanced air mobility operations mature globally. The Federal Aviation Administration's beyond-visual-line-of-sight (BVLOS) waiver process increasingly emphasizes demonstration of robust system monitoring and risk mitigation capabilities—areas where digital twin technology provides significant advantages.

The European Union Aviation Safety Agency's Concepts of Operations for drones includes requirements for continuous performance monitoring and adaptive operational risk assessment. Real-time digital twin enhancement directly addresses these requirements by maintaining model fidelity as operational experience expands.

The research also aligns with emerging U.S. Department of Defense initiatives in autonomous aerial systems, where contested environments demand rapid adaptation to unforeseen conditions without extensive retraining infrastructure. The ability to perform model updates on edge computing hardware within 30-90 second windows supports tactical responsiveness requirements.

Conclusion

EnFlexiTwin represents a significant advance in enabling adaptive autonomy for UAV-based transportation systems through intelligent, real-time digital twin enhancement. By automating the construction of incremental learning datasets using attention-based data selection, the platform addresses fundamental challenges in maintaining model accuracy under evolving operational conditions.

Validation across simulated and physical UAV swarms demonstrates practical viability under realistic time constraints and data volumes. The platform's ability to scale across swarm sizes while maintaining zero version delay distinguishes it from existing approaches that struggle with computational bottlenecks.

As UAV integration into transportation networks accelerates, technologies enabling continuous model adaptation without manual intervention will become increasingly critical. EnFlexiTwin provides a foundational architecture for maintaining digital twin fidelity throughout vehicle operational lifetimes, supporting the safety, efficiency, and reliability requirements of next-generation aerial mobility systems.

Sources and Citations

1. Lee, M., Zhu, Y., Hu, Y., Pan, Y., Chen, J., Yao, Y., Yang, G., & Zhou, X. (2025). Real-Time Enhancements of Digital Twins With Incremental Time Series Data in Networked Air-Ground Cooperative UAV Swarm Systems. IEEE Transactions on Intelligent Transportation Systems, 26(12), 22045-22060. https://doi.org/10.1109/TITS.2025.3618724

2. Rodrigues, T. A., Patrikar, J., Oliveira, N. L., Matthews, H. S., Scherer, S., & Samaras, C. (2021). In-flight positional and energy use data set of a DJI matrice 100 quadcopter for small package delivery. Scientific Data, 8(1), 155. https://doi.org/10.1038/s41597-021-00930-x

3. Choudhry, A., Moon, B., Patrikar, J., Samaras, C., & Scherer, S. (2021). CVaR-based flight energy risk assessment for multirotor UAVs using a deep energy model. 2021 IEEE International Conference on Robotics and Automation (ICRA), 262-268. https://doi.org/10.1109/ICRA48506.2021.9560787

4. Costa, E. A., Rebello, C. M., Schnitman, L., Loureiro, J. M., Ribeiro, A. M., & Nogueira, I. B. R. (2024). Adaptive digital twin for pressure swing adsorption systems: Integrating a novel feedback tracking system, online learning and uncertainty assessment for enhanced performance. Engineering Applications of Artificial Intelligence, 127, 107364. https://doi.org/10.1016/j.engappai.2023.107364

5. Jiang, F., Ding, Y., Song, Y., Geng, F., & Wang, Z. (2023). Digital twin-driven framework for fatigue lifecycle management of steel bridges. Structure and Infrastructure Engineering, 19(12), 1826-1846. https://doi.org/10.1080/15732479.2022.2052165

6. Wu, T., Yang, F., Farooq, U., Li, X., & Jiang, J. (2024). An online learning method for constructing self-update digital twin model of power transformer temperature prediction. Applied Thermal Engineering, 237, 121728. https://doi.org/10.1016/j.applthermaleng.2023.121728

7. Masana, M., Liu, X., Twardowski, B., Menta, M., Bagdanov, A. D., & van de Weijer, J. (2023). Class-incremental learning: Survey and performance evaluation on image classification. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(5), 5513-5533. https://doi.org/10.1109/TPAMI.2022.3213473

8. van de Ven, G. M., Tuytelaars, T., & Tolias, A. S. (2022). Three types of incremental learning. Nature Machine Intelligence, 4(12), 1185-1197. https://doi.org/10.1038/s42256-022-00568-3

9. Qin, Y., Wu, X., & Luo, J. (2022). Data-model combined driven digital twin of life-cycle rolling bearing. IEEE Transactions on Industrial Informatics, 18(3), 1530-1540. https://doi.org/10.1109/TII.2021.3089340

10. Attaran, M., & Celik, B. G. (2023). Digital twin: Benefits, use cases, challenges, and opportunities. Decision Analytics Journal, 6, 100165. https://doi.org/10.1016/j.dajour.2023.100165

Real-Time Enhancements of Digital Twins With Incremental Time Series Data in Networked Air-Ground Cooperative UAV Swarm Systems | IEEE Journals & Magazine | IEEE Xplore