Robust and Model-Free SoC Estimation for UAV Batteries via Feature-Selection-Augmented Deep Learning

Abstract

This study presents a model-free, feature-selection-based deep-learning approach for accurate and robust State of Charge (SoC) estimation of lithium-polymer (Li-Po) batteries in mission- and safety-critical unmanned aerial vehicle (UAV) applications. SoC estimation is challenging because it is not directly measurable and battery dynamics are nonlinear, time-varying, and strongly affected by the changing load profiles of real flight missions. Although learning-based estimators avoid the dependence of model-based filters on accurate battery models and tuned noise covariances, their performance depends strongly on the selected input features and network architecture. In the proposed framework, seventeen engineered features derived from voltage and current measurements are ranked using Pearson correlation, Spearman correlation, mutual information, partial correlation, and time-lagged cross-correlation. The resulting feature subsets are evaluated using a Deep Feedforward Neural Network (DFNN) and a Long Short-Term Memory (LSTM) network, with the LSTM sequence length treated as an additional design variable. The framework is assessed using real UAV flight data, together with feature-ablation and sensor-noise robustness analyses conducted at 1%, 3%, and 5% current and voltage noise levels using repeated independent noise realizations. The results reveal an architecture-dependent interaction: the feature-selection criterion best suited to the memoryless DFNN differs from that favoring the recurrent LSTM. They also show that configurations minimizing average error do not necessarily minimize worst-case error or provide the highest robustness under noisy measurements. The ablation and reduced-input analyses identify cumulative energy as the strongest predictive feature within the evaluated dataset, while also showing that the LSTM retains substantial estimation capability when this feature is excluded. The best-performing four-input LSTM configuration, combining voltage, cumulative energy, current, and instantaneous power with a sequence length of 500, achieves an RMSE of 0.1266%, representing a 62.4% reduction relative to the best DFNN configuration. Among the evaluated configurations and under the investigated operating conditions, this four-input LSTM estimator provides the lowest RMSE while retaining a compact input structure for UAV battery-management applications.