A unified deep learning framework for high-performance RUL prediction

All cells were tested using a battery cycler (CRC 05-10, Wonik PNE, South Korea) within a temperature-controlled chamber (IL-11, JEIO TECH, South Korea) maintained at 25 °C. Each cell underwent cycles that alternated between fast charging (3C) and slow charging (0.5C), followed by a standardized 1C CC discharge and a 10-minute rest period between charging and discharging. For full-range fast charging, a 3C CC-CV protocol was employed in a SOC range of 2.5−4.2 V with a 20-minute time limit.

Dataset 2 employed the same protocol but within a voltage range of 3.2−4.2 V and a shortened time limit of 17 minutes. Similarly, full-range slow charging used a 0.5C CC-CV protocol within 2.5−4.2 V with a 120-minute limit, while Dataset 2 applied in 3.2−4.2 V, limiting the time to 100 minutes. To constrain the SOC range during charging to the MID–HIGH region in Dataset 2, the discharge cut-off voltage was set at 3.2 V, which corresponds to approximately 16.6% SOC under 1C discharge.

A diagnostic cycle at a 0.2C (dis)charging was inserted every 30 cycles from BOL for all datasets. Voltage, current, and capacity data were recorded every 10 seconds during each cycle.