michealsmitch/lithium-ion-cell-capacity-grading-process-curves

Lithium-ion Cell Capacity-Grading (FR) Process Curves

Channel-level process curves from the capacity-grading station of a cylindrical lithium-ion cell line. Every tester channel is sampled natively every 30 seconds for the whole process, giving the full voltage / current / capacity trajectory of each cell from the moment it is clamped.

This is the capacity-grading (FR) dataset. Pre-charge is a completely different process and is published separately; the two are deliberately never mixed.

• 75 runs, 17 machines, 2 products, 2 lots
• 23,808 tester channels / 21,499 loaded cells (90.3 % tray utilisation)
• 9 process steps per run, roughly 3–3.3 hours per run
• 9,344,663 samples

Process	FR capacity grading (分容)
Products	HJ and CK — two different products, not just two batches
Lots	HJ7PM06100 (43 runs), CK7PM230FX (32 runs)
Tester types	6A-256E (43 runs), 6A-400E (32 runs)
Lines	F1501, F1502, F1601, F1602
Date	2026-01-01 and 2026-01-08
Source language	Simplified Chinese and English (two header generations)

---

Files

Config	File	Rows	Cols	Size
curves	data/curves/part-*.parquet	9,344,663	29	105 MiB (75 shards)
runs	data/runs.parquet	75	16	13 KiB

This repository contains a single split (train) holding the entire dataset. No train/test partition is provided — see Suggested splitting below.

Quickstart

from datasets import load_dataset
import pandas as pd

runs = load_dataset("parquet", data_files="data/runs.parquet", split="train").to_pandas()
print(runs.groupby(["product", "lot_id", "tester"]).size())

curves = load_dataset("parquet", data_files="data/curves/*.parquet", split="train")
d = curves.to_pandas()
d = d[d.is_loaded & (d.step_no == 6)]               # low-rate finish = real capacity
print(d.groupby("lot_id").accumulated_capacity_mah.describe())

Process description

Cells arrive already formed, but at different states of charge. Each run then:

#	step_name	Purpose
1	rest	settle / read incoming OCV
2	cccv_charge	normalise the starting point by topping the cell up
3	rest	relax polarisation
4	cc_discharge	high-rate discharge, time-terminated for essentially every cell
5	rest	relax polarisation, let the cell cool
6	cc_discharge	low-rate finish down to the lower voltage limit
7	rest	settle
8	cccv_charge	recharge to the shipping state of charge
9	rest	settle / read final OCV

Observed outcomes: cumulative discharge capacity at step 6 ≈ 4.12 Ah for the HJ lot and ≈ 2.02 Ah for the CK lot; shipping OCV ≈ 3.55 V and 3.62 V; run time ≈ 3.0 h and 3.3 h. The two products also enter the station at different states of charge, so step-2 top-up capacity differs by more than an order of magnitude between them.

Set-points are not restated in this card. The currents, voltage limits and cut-off conditions are plainly visible as the plateaus and cut-offs in the data itself. Omitting the table here is a documentation choice by the data owner, not a technical safeguard — anyone holding the curves can read the set-points straight off them.

Schema

curves — one row per channel per sample

Vendor columns are translated from Simplified Chinese to English and keep the source order; the four leading columns are the only ones added, so that a sample can be tied back to a run and to a physical tray socket.

Column	Unit	Description
run_id		added join key → runs.run_id
power_channel		added tester channel, taken from the source file name
tray_position		added physical socket of the cell on the tray
is_loaded		added False for empty tray sockets — filter on this
curve_id		sample counter within the channel
cycle		cycle index; always 1 here
step_no		step number 1…9
test_time	mm:ss	time inside the current step
accumulated_time_min	min	cumulative process time
voltage_mv	mV	channel voltage
current_ma	mA	channel current
capacity_mah	mAh	capacity transferred within the step
energy_mwh	mWh	energy within the step
end_flag		non-zero only on the last sample of a step
temperature_c	°C	cell/channel temperature field — almost always 0
save_time		wall-clock timestamp
accumulated_capacity_mah	mAh	cumulative capacity for the run — the grading capacity
negative_pressure		all zero in this source
barcode		cell barcode, <model> <serial>; null for empty sockets
capacity_mah_bat	mAh	duplicate of capacity_mah in this source
ambient_temperature_c	°C	ambient temperature for this sample
step_name		canonical English step name (see controlled vocabularies)
stop_condition		why the step ended, on the last sample only
probe_voltage_mv	mV	all zero in this source
accumulated_energy_mwh	mWh	cumulative energy for the run
negative_electrode_temperature_c	°C	all zero in this source
cable_voltage_mv	mV	all zero in this source
contact_voltage_mv	mV	all zero in this source
contact_impedance		all zero in this source

runs — one row per tester run

Column	Description
run_id	<machine>/<lot>_<tray>_<tester-address>_<timestamp>
machine	MES machine number (153–170)
product	product family — HJ or CK; the lot code carries both a product and a batch
lot_id	production lot (HJ7PM06100 / CK7PM230FX)
tray_code	physical tray barcode
line	tester line prefix of the device address
device_address	pin-bed station, F<line>-<field>-<branch>
tester	tester type, 6A-256E or 6A-400E
procedure	FR for every row in this dataset
unit_number	tester unit / work-station index
start_time, end_time, duration_min	run envelope
n_positions, n_loaded, load_ratio	tray utilisation

Controlled vocabularies

step_name: rest, cc_charge, cccv_charge, cc_discharge, cccv_discharge, end

stop_condition: time, current, cv_point, voltage_lower, end_point, check_code_abnormal, voltage_median_offset, current_anomaly, voltage_anomaly, voltage_out_of_range, channel_disable

Any value not listed is passed through unchanged from the vendor file.

Both vendor header generations (Simplified Chinese and English) were mapped onto these values.

Indexing — tray socket vs tester channel vs tester unit

Three distinct numbers appear per run and they are not interchangeable:

Name	Meaning
tray_position	physical socket of the cell on the tray (托盘通道)
power_channel	tester power-supply channel (电源通道) — also the source file name
unit_number	which unit / work-station of the tester handled the tray (设备ID)

1. The tester records by power channel, not by tray socket. Joining the tester's per-channel record onto power_channel matches 256/256 or 400/400 cell barcodes; joining it onto tray_position matches 0.
2. The two numberings differ by a fixed wiring permutation, constant per tester. This dataset contains 4 permutations:

line	channels	first mappings
F1501	256	tray 1 → ch 241, ascending
F1502	256	tray 1 → ch 16, descending
F1601	400	tray 1 → ch 381, ascending
F1602	400	tray 1 → ch 20, descending

Both numbers are delivered per row, so (run_id, power_channel, tray_position) is complete; recover the permutation with d[["run_id","tray_position","power_channel"]].drop_duplicates(). 3. unit_number is a deterministic function of the tester address F<line>-<field>-<branch>:

Lines	Tester	Rule (exact on all runs)
F1501, F1502	256 ch	unit = 9 × ((field − 1) mod 4) + (branch − 1)
F1601, F1602	400 ch	unit = 8 × ((field − 1) mod 4) + (branch − 1)

Known issues and caveats

#	Issue	What to do
1	Step-4 capacity is time-truncated	Every cell runs the fixed high-rate discharge to the end of its time budget, so step-4 capacity ≈ I × t with almost no spread. Use the step-6 plateau of accumulated_capacity_mah as the grading capacity.
2	Unloaded tray sockets are present	2,309 of 23,808 channels never had a cell. They carry barcode = null, is_loaded = False and meaningless 0–few-mV curves. Filter on is_loaded.
3	HJ and CK are different products	Different capacity class, different discharge cut-off, different incoming state of charge. Never pool them without a product term.
4	Mixed cell models inside the CK lot	CK7PM230FX contains 12 different barcode prefixes, often 2–4 within one tray. The HJ lot is single-model. Group by the barcode prefix.
5	Tray utilisation varies a lot	Runs load between 10 and 400 cells; use runs.load_ratio / n_loaded.
6	temperature_c and several other columns are all zero	temperature_c is non-zero on only ~3 % of samples; probe_voltage_mv, negative_pressure, negative_electrode_temperature_c, cable_voltage_mv, contact_voltage_mv and contact_impedance are entirely zero. There is no per-cell temperature; use ambient_temperature_c.
7	capacity_mah_bat duplicates capacity_mah	The vendor file carries two columns with the same Chinese header; both are kept for format fidelity.
8	30 s cadence slips slightly	Use accumulated_time_min or save_time, not a row index, for time alignment.

What is deliberately not included

Excluded	Why
RecipeInfo.csv	process set-points
WorkInfo.xml	contains <FlowData>, the complete parameter block — i.e. the recipe itself
EigenData.mdb	the tester's own database
UnitN.csv	per-step summary; only its tray↔channel mapping is used, denormalised onto the curves
Temperature*.csv, BoardTempSum.csv, EnvTempSum.csv	20 board + 10 ambient probes; ambient temperature already travels inside every curve row
Event.log, Exp.log	equipment event logs

Deleted vendor columns: circuit_voltage_drop_mv and cable_resistance (fixture contact metrics of the tester, not properties of the cell) and cdc_capacity_mah (a firmware artefact holding negative values).

Reproducing / uploading

pip install pandas pyarrow
python prepare.py --root /path/to/raw --out .        # emits both datasets
hf auth login
hf upload <user>/<dataset-name> . . --repo-type dataset

The source files are GBK / cp936 encoded, so reading them as UTF-8 fails.

Licence, privacy and intended use

License & Commercial Use： This sample dataset is released under the CC BY-NC 4.0 license for strictly non-commercial and academic research purposes.

Commercial Licensing: This repository contains only a small subset of a massive 1TB+ industrial-grade dataset, which includes full-lifecycle lithium battery manufacturing data (CCD vision defects, high-frequency formation time-series data, etc.). If you wish to use this data for commercial algorithm training, enterprise product development, or want to acquire the full dataset, please contact me for a commercial license:

Email: michael.smitch.20260912@gmail.com