Driving behavior-based state of charge estimation for electric vehicles under sensor-degraded conditions

• A convolutional-long short-term memory (Conv-LSTM) based framework for battery-agnostic SoC estimation using driving behavior data. • Global false nearest neighbor (GFNN)-based sliding window selection to determine optimal temporal embedding for adaptive SoC sequence modeling. • Sensor degradation, including multi-level input noise and missing data, emulated to assess real-world robustness. • Validation on heterogeneous real-world EV fleet routes. • High estimation accuracy under both normal and degraded inputs (missing-data and multi-level noise conditions). Accurate estimation of the State of Charge (SoC) is essential for reliable and energy-efficient operation of electric vehicles (EVs), particularly under dynamic driving conditions and sensor degradation. Traditional SoC estimation approaches often rely on battery-internal measurements and electrochemical models, which limit their transferability across heterogeneous vehicles and restrict applicability in large-scale fleet deployments. This study proposes a battery-agnostic SoC estimation framework that integrates Global False Nearest Neighbor (GFNN)-based temporal embedding with a Convolutional Long Short-Term Memory (Conv-LSTM) network using real-world driving-behavior inputs (velocity, acceleration, and distance). Unlike many existing methods that rely on laboratory drive cycles or empirically selected window lengths, the proposed approach determines the sliding-window length systematically and is trained and validated using heterogeneous real-world driving data from BluSmart Mobility EV fleets collected across campus, market, city, and ring routes. To enhance practical relevance, the model is further evaluated under missing-data and multi-level noise scenarios to emulate realistic sensor degradation. Across four test trips, the proposed method achieves approximately 99.96%, 98.89%, 99.3%, 99.23%, and 98.72% accuracies under normal, missing, and low, medium, and high noise conditions, respectively, demonstrating reliable SoC estimation under real-world operating conditions.