nasa-rw-2014

Tier 2 BSEBench-canonical Parquet harmonization of the NASA PCoE Randomized Battery Usage Data Set (2014, Bole, Kulkarni & Daigle). Hosted under the BSEBench organization for filter-benchmark reproducibility.

Disambiguation

This is the harmonization of the 2014 Random Walk dataset, not the older 2007–2008 NASA PCoE Battery Aging series (B0005, B0006, B0007, B0018). The 2007 B-series is excluded from BSEBench v1.0 — see the nasa-rw-2014-raw card for the rationale.

Status

🚧 Harmonization pending — placeholder dataset card. Awaiting :

1. Tier 1 raw mirror upload at bsebench-org/nasa-rw-2014-raw
2. Publication of bsebench-specs v0.2.0 with the canonical TimeSeriesSchema Pydantic model
3. Implementation + testing of the NASA RW harmonization adapter at bsebench-datasets/src/bsebench_datasets/adapters/nasa_rw_2014.py

What this will be

The NASA RW 2014 dataset (28 cells, 7 parts) converted to the BSEBench canonical Parquet schema :

Column	Type	Description
cell_id	string	identifier within the dataset (e.g. RW1, RW2, …, RW28)
time_s	float64	seconds since experiment start (t = 0 at first sample, monotonic across all cycles for a given cell)
voltage_V	float64	terminal voltage, always positive
current_A	float64	BPX-1.1 sign convention : charge = positive, discharge = negative
temperature_C	float64	cell-surface thermistor reading ; nullable when channel is masked (RW3 caveat)
cycle_number	int32	cycle counter, 1-indexed ; a single counter unifies Random-Walk segments and pulsed-load reference cycles
step_id	string	within-cycle step identifier (rw_discharge, rw_charge, cc_charge, pulsed_ref, eis, rest)
capacity_Ah	float64?	running coulomb-counted capacity (computed from current_A × Δt, sign per BPX-1.1)

Partitioned by cell_id for parallel reading. Expected size : ~3-5 GB uncompressed, ≤ 600 MB after Zstandard compression (estimate based on a 1 Hz sampling cadence and ≤ 200 days of cumulative recording per cell ; verify after first end-to-end run).

Mapping from original

The NASA RW .mat files use a nested MATLAB struct layout per cell :

data = struct(
    procedure = "Random Walk Cycling",
    description = "...",
    step = struct(
        type = ["D"|"C"|"R"|"P"|...],   % discharge / charge / rest / pulsed
        relativeTime = [...],             % seconds since cycle start
        time = [...],                     % MATLAB datenum
        voltage = [...],
        current = [...],
        temperature = [...],
        date = [...],                     % start of cycle
        comment = "..."
    )
);

The harmonization rewrites this into long-form Parquet with one row per sample :

.mat field	BSEBench column	Conversion
step.relativeTime (cumulative across cycles)	time_s	unchanged
step.voltage	voltage_V	unchanged
step.current	current_A	sign flipped (see below)
step.temperature	temperature_C	masked NaN if cell ∈ {RW3} (corrupted thermistor)
step.type	step_id	mapped to canonical step vocabulary
(cycle index)	cycle_number	re-numbered monotonically including reference cycles

Sign-convention notes

The NASA PCoE Random Walk dataset uses the convention positive current = discharge (load) — the legacy convention common to older battery test benches and NASA aerospace prognostics literature.

BSEBench follows BPX 1.1, which defines positive current = charge (into the cell). The harmonizer therefore applies a single sign flip when reading the step.current field :

current_A_bsebench = -1.0 * step.current_nasa

This is documented per-cell in the manifest YAML so that auditors can verify the conversion against Tier 1.

Caveat — confidence MEDIUM : the discharge-positive convention for NASA PCoE Random Walk has not been confirmed from a primary NASA technical report at the time of this card's draft. Confirmation will be performed during adapter implementation by inspecting cycle end-points (a CC-discharge cycle should drive voltage down) and the result locked in the manifest with a verification SHA. If the convention turns out to be charge-positive, the sign flip is removed — the harmonization remains correct.

Time origin

NASA stores absolute timestamps (step.time in MATLAB datenum format) plus per-cycle relative timestamps (step.relativeTime). BSEBench's time_s column is built by concatenating relative times across cycles in chronological order, yielding a monotonic time_s from 0 through the cell's full operating life. Absolute MATLAB datenum is preserved in the manifest YAML for users needing wall-clock alignment.

Cycle numbering

A single cycle_number counter is used across both Random-Walk segments and pulsed-load reference cycles. The step_id column distinguishes their roles. This avoids the asymmetric cycle counters that appear in the raw .mat (where reference cycles sometimes restart at index 0).

Excluded cells

By default, two cells are flagged but not removed :

• RW3 : temperature_C masked NaN throughout (corrupted channel).
• RW20 : marked with a per-cell flag quality_warn = true. Rows are retained for completeness ; users running estimator-quality metrics should filter on cell_id != 'RW20'.

The default loader exposes these as opt-in / opt-out flags ; raw rows are preserved verbatim in Parquet to allow downstream re-evaluation.

Why "canonical" tier

This is the version users should consume for filter benchmarking. It :

• Maps NASA's struct layout to BSEBench's standardized columnar schema
• Harmonizes sign convention to BPX-1.1
• Unifies cycle numbering across the seven parts and reference cycles
• Strips MATLAB-specific metadata not needed for state-estimation
• Loads in 1 line via bsebench-datasets

For provenance / audit, see the Tier 1 raw mirror.

Anchor standards

• Sign convention : BPX 1.1 (charge = positive, discharge = negative)
• Time-series schema : LF Energy BDF (Dec 2025) where applicable

License

• Underlying NASA data : public domain / CC0 1.0 (per NASA Open Data Policy and the NASA PCoE Data Repository terms — see the Tier 1 card for the verbatim disclaimer).
• BSEBench-derived harmonization layer (column mapping, sign flip, Parquet schema, cycle numbering) : CC-BY-4.0, per the BSEBench organization's standard policy for derived data.

Users who consume only the Tier 2 Parquet files should comply with both licenses (CC0 for raw content + CC-BY-4.0 for the BSEBench harmonization layer) — in practice this means citing both NASA and BSEBench, which is what we recommend regardless.

How to use (target API)

# Once shipped (M2.1)
from bsebench_datasets import load_bsebench

ds = load_bsebench("nasa-rw-2014", revision="v1.0")
print(ds.chemistry)            # "LCO"
print(ds.cell_count)           # 28
print(ds.first_cell_data().head())  # canonical Parquet columns

Citation

Cite the NASA PCoE record (no DOI) :

@misc{bole2014_nasa_rw,
  author    = {Bole, Brian and Kulkarni, Chetan S. and Daigle, Matthew},
  title     = {Randomized Battery Usage Data Set},
  year      = {2014},
  publisher = {NASA Ames Prognostics Center of Excellence (PCoE)},
  howpublished = {NASA PCoE Data Repository, Item~\#11},
  url       = {https://www.nasa.gov/intelligent-systems-division/discovery-and-systems-health/pcoe/pcoe-data-set-repository/}
}

@misc{bsebench2026,
  author = {Akir, Oussama and BSEBench Contributors},
  title  = {{BSEBench}: an open-source benchmark for battery
            state-estimation filters},
  year   = {2026},
  url    = {https://bsebench.org},
}

Provenance

Derived from bsebench-org/nasa-rw-2014-raw via bsebench-datasets/src/bsebench_datasets/adapters/nasa_rw_2014.py. The harmonization script's commit SHA is recorded in the manifest YAML at bsebench-datasets/manifests/nasa_rw_2014.yaml.