A Comparative Study of Machine Learning and Deep Learning Models for State of Charge and Remaining Useful Life Estimation on a Rotary-Wing UAV Battery

Abstract

Battery state estimation is the main safety constraint for electric rotary-wing unmanned aerial vehicles (UAVs): mission decisions depend on both the instantaneous State of Charge (SOC) and the Remaining Useful Life (RUL). The present study compares seven machine learning and deep learning models (LR, SVM, k-NN, GBT, EL, LSTM, and a simplified RWKV) on real flight data from a rotary-wing helicopter testbed with a Pixhawk autopilot and an NVIDIA Jetson Nano mission computer. The dataset has 1310 samples (∼262 s) of nine on-board sensor signals. Mission-based RUL is defined as the projected time until SOC reaches a 20% safe-landing threshold. All models use an 80/20 random split, five regression metrics (RMSE, MAE, R2, MSE, PRMSE), and five random seeds. GBT wins on SOC with R2=0.9943±0.0014, MAE =0.25%, and 3.8μs per-sample inference on a workstation CPU; this latency leaves headroom for on-board mission planning. Battery temperature and voltage together carry over 90% of the predictive signal. GBT wins again on RUL (R2=0.596±0.042, MAE =583 s). The same ordering (tree ensemble ≻ recurrent ≻ linear) holds for both tasks; the remaining RUL gap reflects the single-flight dataset. The SOC labels originate from the on-board autopilot’s Coulomb-counting-based fuel-gauge estimator, so the SOC numbers should be read as a reproduction of that on-board trace at sub-microsecond inference latency rather than as independent accuracy; the calibration-free Coulomb-counting baseline reaches a marginally higher R2 (0.9950, MAE =0.35%) on the same task.