BATeCHAIN Digital Twin Data

P. Matorras Cuevas, M. A. Melgarejo

This repository contains the trained models as well as the datasets used for such trainings These datasets contain preprocessed, cycle-level representations and trained models derived from three publicly available NMC/graphite lithium-ion battery aging studies: the KIT 228-cell dataset (Luh & Blank, Scientific Data 2024; DOI: 10.1038/s41597-024-03831-x), the RWTH Mannheim 48-cell dataset (Sanyo/Panasonic UR18650E), and the UConn 44-cell aging dataset (Panasonic UR18650AA; https://digitalcommons.lib.uconn.edu/reil_datasets/2). Together, the three sources cover 320 NMC/graphite 18650 cells across a wide range of cycling conditions: constant-current cycling at 25 °C, variable-rate cycling at different depths of discharge, and real-world driving-cycle profiles with thermal variation, spanning 300–2 800 cycles per cell. For each source, the raw voltage/current/temperature timeseries are segmented into full charge–discharge cycles. Per-cycle summaries include discharge and charge capacities, coulombic efficiency, mean and peak temperatures, and sparse State-of-Health (SoH) labels derived from reference performance tests. SoH labels for regular cycles are assigned by PCHIP monotone interpolation between consecutive check-up measurements. An additional set of derived features is provided for each cycle: ten scalar health indicators (internal resistance, energy efficiency, charge duration, constant-voltage tail fraction, temperature statistics, and charge/discharge capacity) and RevIN-normalised discharge waveforms resampled to 30 s resolution (300 timesteps × 3 channels: voltage, current, temperature). Three trained SoH estimation models are also included, all trained jointly on the combined three-dataset corpus. The first is a cross-cycle LSTM operating on scalar health indicators only (no waveform input). The second and third are two-stage waveform encoder models: a per-cycle 1D CNN encoder followed by a cross-cycle LSTM, and the same cross-cycle LSTM paired instead with a frozen pretrained Chronos time-series backbone as the per-cycle encoder. These representations and models are intended for machine learning research on SoH estimation and remaining-useful-life prediction from field-available operational signals, without reliance on specialised reference tests or impedance measurements.