Planning-Compatible Data-Driven Voltage and SoC Prediction From Operational Multirotor Flight Logs With Independent-Flight Validation
IEEE Access, cilt.14, ss.93771-93799, 2026 (SCI-Expanded, Scopus)
- Yayın Türü: Makale / Tam Makale
- Cilt numarası: 14
- Basım Tarihi: 2026
- Doi Numarası: 10.1109/access.2026.3703556
- Dergi Adı: IEEE Access
- Derginin Tarandığı İndeksler: Science Citation Index Expanded (SCI-EXPANDED), Scopus, Compendex, INSPEC, Directory of Open Access Journals
- Sayfa Sayıları: ss.93771-93799
- Anahtar Kelimeler: battery state of charge (SoC), battery voltage prediction, design of experiments, Gaussian process regression (GPR), operational flight logs, UAV
- Kayseri Üniversitesi Adresli: Evet
Özet
Accurate preflight prediction of battery terminal voltage and state of charge (SoC) is essential for endurance and safety-margin planning in multirotor UAV missions.We propose a reproducible segment-level workflow that turns operational ArduPilot logs into planning-compatible predictors of end-of-segment voltage and SoC, validated on independent flights. Here, planning-compatible means that the predictor can be queried before execution using only variables available at planning time (plan geometry/nominal speed) together with an initial battery state, without requiring in-flight telemetry streams. Our automatic labeling pipeline aligns executed QGroundControl plan files with log traces, leveraging Mission Item State Estimate (MISE) mission-progress messages and using reached command events. With planning-time inputs only (start voltage/SoC, horizontal progression, altitude change, and nominal speed), we estimate segment duration and then predict terminal voltage and SoC, and we evaluate against single-stage baselines using leakage-safe, flight-grouped cross-validation. To explicitly capture boundary effects induced by stabilization and yaw-alignment maneuvers, we additionally propose an intersegment state-update model that predicts post-boundary voltage and SoC from pre-boundary voltage/SoC, planned intersegment wait time (treated as a known input), and yaw-change magnitude. Training data are collected efficiently using a Taguchi L25 design over progression, altitude change, and speed, combined with combinatorial segment ordering across nine missions. Independent-flight validation comprises: (i) an out-of-design-distance study targeting within-segment generalization, (ii) an intersegment stress test spanning wait time and yaw change, and (iii) mission-level open-loop rollouts on free-flight missions. All three blocks use external flights that are fully disjoint from the training flights used in cross-validation and model selection, so the reported results reflect true flight-to-flight generalization rather than temporally correlated splits. Results show that the selected two-stage GPR-to-GPR model achieves strong leakage-safe accuracy under flight-grouped cross-validation, with an MAE of about 1.24 s for segment duration, 0.018 V for end-voltage, and 0.32 percentage points for end-SoC. External ablation validation confirms stable generalization with bounded errors (overall MAE of 0.89 s, 0.024 V, and 0.16 pp, respectively), while the intersegment state-update remains robust across dwell-time and yaw-change conditions (RMSE of about 0.023 V and 0.14 pp), supporting practical plan-only offline mission feasibility and safety-margin assessment.