BatteryMind · research behind the demo

How we estimate battery health from the data your BMS already records

BatteryMind estimates a battery's state of health (SOH) and remaining life from ordinary operating data. No test equipment is needed and the battery stays in service. Every method in the demo was first tested on open field data from real batteries, against real capacity measurements, with the pass criteria fixed before the test.

Abbreviations: SOH = state of health (capacity left vs. new) · SOC = state of charge · EOL = end of life · BMS = battery management system · EMS = energy management system · UPS = uninterruptible power supply · LFP = lithium iron phosphate · NMC = nickel manganese cobalt · LMO = lithium manganese oxide · pp = percentage points · Ah = amp-hours

60real field capacity tests used for end-to-end validation

49field battery systems analysed (21 home storage, up to 8 years each, plus 28 BMS-level systems)

20open datasets reviewed and loaded, lab and field

120+research papers reviewed

Try the Battery Health Check Open data used References

Pipeline

From a BMS export to a health report

1 · BMS / EMS exportPack voltage, current, SOC and temperature; min/max cell voltage with cell IDs if available. 5-minute resolution is enough.

2 · Data checksCurrent sign, idle-current offset, gaps, sensor resolution, series cell count.

3 · Health indicatorsCapacity between full and empty, internal resistance trend, weak-cell pattern, step changes.

4 · SOH and end of lifeMeasured health combined with fleet aging rates and a physics aging model, with 90 % ranges.

5 · ReportEvery number names its method and its uncertainty, and says when the data can't answer.

Principles

What makes it credible

Tested before used

A method enters the product only after it passes a test on real data, with the pass criterion written down in advance. Methods that failed are documented and kept out.

Validated against capacity tests

Real capacity tests on field batteries are the reference. The model's own predictions are never used as the reference.

Honest uncertainty

Ranges are calibrated on field results, and the report states how often such a range held the true value in validation.

Says "unknown" when it is

If the data can't support an SOH, for example because the battery never discharges fully, the report says so. It doesn't guess.

Methods

The methods and how each was tested

Capacity between full and empty

Main SOH method

Rests near full and near empty are identified by voltage. The charge moved between them gives the usable capacity, corrected for current-sensor offset. This reproduces the field method of Figgener et al. (2024, Nature Energy) and adds a check for typical 5-minute EMS data.

Capacity error at field tests with 5-minute data: 2.8 % (LMO), 3.5 % (NMC), 5.5 % (LFP). Aging rate matches the published estimates to 0.69 pp/yr across 21 systems. 5-minute data loses almost nothing compared with 1-second data.

Aging rate and end of life

Used

Combines (a) measured fade rates of comparable field systems, (b) the published LFP calendar and cycle aging models of Naumann et al. (2018, 2020), driven by the battery's own usage, and (c) the battery's own measured trend, once enough history exists.

The physics model alone, with no measurement, predicted SOH at 19 LFP field tests with a mean error of 3.3 points and no bias. Ranges were recalibrated after a first test showed them too narrow.

Weak-cell detection

Used

Uses only the min/max cell voltage and cell ID that standard BMS exports contain. If the same cell is lowest under discharge and highest under charge, it is most likely a weak or high-resistance cell.

Found the published weak cell in 10 of 13 field systems, against 2.5 expected by chance (p ≈ 4·10⁻⁶).

Internal resistance trend

Indicator

Voltage response to current steps, normalised for SOC and temperature. It tracks power capability and faults over time.

Tested as an SOH proxy and not used for that: for LFP and NMC, resistance and capacity don't move together closely enough. It is reported as a separate trend.

Step-change detection

Indicator

Flags sudden capacity changes, for example from a cell failure or a BMS update. A gradual-aging model would hide these.

Found 4 of 7 real steps in field data, with about 0.2 false alarms per system-year.

BMS-reported SOC

Low confidence

Capacity from charge ÷ SOC change is only valid if the BMS re-learns capacity. Many BMSs count against a fixed rated capacity and keep reporting it, so their SOC shows no fade.

On 28 real BMS systems, 10 of 12 such values came out above 105 % of nameplate. The report flags these as implausible.

Validation

End-to-end results

The complete service was run as a customer would receive it, on 21 home-storage systems. Data were cut three days before each of the 60 field capacity tests, so the test itself was never seen. To avoid tuning on the answers, the correction to full-power-test terms was learned with each system left out.

Chemistry	Tests	Mean error (SOH points)	Test result inside the reported 90 % range
LMO / NMC blend	22	2.4	95 %
NMC	19	6.2	79 %
LFP	19	7.0	68 %
All	60 (60 got an SOH)	5.1	82 %

Example from the demo's sample battery: estimated SOH 89.3 % in full-power-test terms. The real capacity test three days later measured 89.4 %. LFP is the hardest chemistry, because its flat voltage curve hides the state of charge. The report states that lower reliability instead of hiding it.

How much history is needed to see a battery's own aging

Smallest fade rate that can be told apart from noise (percentage points per year, median over field systems). Below this, the report relies on fleet aging rates and says so.

Months of data	LMO	NMC	LFP
6	6.0	7.8	26.7
12	3.2	5.6	8.7
18	2.1	3.1	6.2
24	1.1	2.3	4.7
36	0.6	1.7	3.6

Typical fade is about 2 pp per year. So a battery's own trend becomes visible after about 18 months for LMO, about 30–36 months for NMC, and more than 3 years for LFP.

Open data

Datasets analysed

All datasets are publicly available from their authors. We thank them for publishing them. None of them is redistributed here, except the demo's sample: a two-year excerpt of one home-storage system, used under CC-BY-4.0.

Dataset	What it contains	Used for	Licence
RWTH home-storage field data Figgener et al. 2024	21 PV home-storage systems (LFP, NMC, LMO), up to 8 years at 1 s, 60 field capacity tests	Main validation: capacity method, aging rates, end-to-end test	CC-BY-4.0
Li-ion battery field data, 28 systems Schaeffer et al. 2024	28 LFP systems with BMS SOC and per-cell voltages, up to 5 years	Weak-cell detection, BMS-data case study	CC-BY-NC-4.0
M5BAT large-scale BESS RWTH Aachen ISEA	5–6 MW multi-technology grid storage, incl. lead-acid, 1 Hz field operation	Showing why BMS-SOC-based capacity is circular	CC-BY-4.0
Oxford energy-trading degradation Reniers, Mulder, Howey 2020	Cells aged on grid-arbitrage duty with capacity checks	Lab testbed: EMS-resolution method tests	ODC-ODbL
Long-term sweat testing, second-life Beatty et al. 2025	6 second-life modules on real grid and home use cases, 2019–2025	Duty-cycle reference	CC-BY-4.0
Second-life aging, grid-storage cycling Moy et al. 2024, Stanford	Retired EV cells on synthetic home and commercial storage duty, 24 months	Duty-cycle reference	CC-BY-NC
EVBattery field dataset He et al. 2023	Field charging data from hundreds of EVs with health labels	Labelling-method reference	CC-BY-4.0
MIT–Stanford–Toyota LFP cycle life Severson et al. 2019	124 LFP cells cycled to end of life	Lab background: lifetime prediction	CC-BY
HUST LFP dataset Ma et al. 2022	77 LFP cells, multi-stage discharge, cycled to end of life	Lab background	CC-BY-4.0
Multi-stage Li-ion aging study Sci. Data 2024	279 NMC cells, 71 conditions, combined calendar and cycle aging	Lab background: calendar vs cycle aging	CC-BY-4.0
TBSI–Sunwoda battery dataset Tao et al.	Commercial cells, multi-step fast charging	Lab background	CC-BY-4.0
Oxford battery degradation dataset 1 Birkl 2017	8 pouch cells, drive-cycle aging	Lab background	ODC-ODbL
CALCE battery data University of Maryland	Cycle aging, partial cycling, storage, anomaly sets	Lab background	Cite CALCE
NASA PCoE battery data	18650 aging and randomized-use packs	Lab background	Public domain

Lab datasets taught us what lab data can and can't predict. Deployed batteries have unknown history and no full test cycles, so the product is validated on field data only.

References

Key publications the methods build on

Figgener, J. et al. (2024). Multi-year field measurements of home storage systems and their use in capacity estimation. Nature Energy. doi:10.1038/s41560-024-01620-9 · capacity method we reproduce and extend
Figgener, J. et al. (2024). Degradation mode estimation using reconstructed open circuit voltage curves from multi-year home storage field data. arXiv. arXiv:2411.08025 · companion analysis
Schaeffer, J. et al. (2024). Gaussian process-based online health monitoring and fault analysis of Li-ion battery systems from field data. Cell Reports Physical Science. doi:10.1016/j.xcrp.2024.102258 · weak-cell reference
Naumann, M. et al. (2018). Analysis and modeling of calendar aging of a commercial LiFePO₄/graphite cell. Journal of Energy Storage 17. doi:10.1016/j.est.2018.01.019 · LFP calendar aging model
Naumann, M., Spingler, F., Jossen, A. (2020). Analysis and modeling of cycle aging of a commercial LiFePO₄/graphite cell. Journal of Power Sources 451. doi:10.1016/j.jpowsour.2019.227666 · LFP cycle aging model
Möller, M. et al. (2022). SimSES: A holistic simulation framework for modeling and analyzing stationary energy storage systems. Journal of Energy Storage 49. doi:10.1016/j.est.2021.103743 · simulation cross-check
Jacqué, K. et al. (2022). The influence of frequency containment reserve on the operational data and the state of health of the hybrid stationary large-scale storage system. Energies 15(4). doi:10.3390/en15041342 · grid-storage field SOH
Sulzer, V. et al. (2021). The challenge and opportunity of battery lifetime prediction from field data. Joule. doi:10.1016/j.joule.2021.06.005 · accuracy targets for field SOH
Severson, K. A. et al. (2019). Data-driven prediction of battery cycle life before capacity degradation. Nature Energy. doi:10.1038/s41560-019-0356-8 · lab lifetime-prediction background

Try the Battery Health Check