A unified deep learning framework for high-performance RUL prediction
Description
Datasets 1 and 2 were constructed using NCA/Gr-Si based SONY 18650 VTC6 cells (A nominal capacity of 3 Ah and a nominal voltage of 3.6 V). Dataset 3 employed Samsung SDI INR18650 25R cells with NCA+NCM/Gr-Si (A nominal capacity of 2.5 Ah, and a nominal voltage of 3.6 V). Datasets 1, 2 and 3 consist of 90, 12 and 28 cells respectively. Each dataset consists of cycling data generated by assigning distinct operating profiles---each with a length of up to 1000 cycles and extracted from Markov matrices representing different user groups---to individual cells. During the generation of protocols from the configured Markov matrices, half of the initial cycles for each generated protocol were designated as fast-charging and the other half as slow-charging. These profiles, each consisting of a distinct charging sequence, emulate specific user group charging patterns by alternating between fast and slow charging based on derived probabilistic behavior. 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.
Files
Steps to reproduce
This dataset comprises three subsets: Dataset 1, Dataset 2, and Dataset 3. [Dataset 1] The Dataset 1 folder contains the following components: 1. feature_correlation.xlsx: An Excel file summarizing the Pearson correlation coefficients between the Remaining Useful Life (RUL) and 157 features extracted via feature engineering. 2. Fast_start_protocol_Dataset1 and Slow_start_protocol_Dataset1: These files contain 48 randomly generated protocols each, initiated with fast charging and slow charging, respectively. Each protocol spans 1000 cycles and was constructed by sampling from a Markov matrix. The first column indicates the protocol number, and each protocol consists of a randomized sequence of operations—Fast n (fast charging repeated n times), Slow m (slow charging repeated m times), and Capacity check 1 (a diagnostic 0.2C capacity check cycle)—based on pre-defined transition probabilities. Protocols shaded in gray indicate sequences that were excluded due to cell data errors occurring mid-cycle. 3. Raw cell cycling data: Each cell file is named using the format initial_unit_protocol number (e.g., Fast_01 corresponds to the first protocol in the Fast_start_protocol_Dataset1 file). The cell cycling data starts with a 0.2C rate performance test (RPT) cycle, followed by operational cycling based on the assigned Markov protocol. A 0.2C capacity check cycle is inserted every 30 cycles. Each check cycle consists of a 0.2C CC discharge (with a 2.5 V cut-off) followed by a full 0.2C charge-discharge cycle. The second discharge capacity from this check cycle is used to compute the SOH. [Datasets 2 and 3] The structure of Datasets 2 and 3 is identical to Dataset 1, except that they do not include a feature_correlation file. In Dataset 2, an additional sequence—a 0.2C charge followed by a 1C discharge with a 3.2 V cut-off—is inserted after each capacity check cycle. This step was added to reset the state-of-charge (SOC) before the next main cycling sequence. All RUL values were calculated by including the 0.2C capacity check cycles.
Institutions
- Seoul National University