Kaggle Electrical Fault Detection & Classification Data

It is the reference bench behind a generation of protective-relaying coursework and papers. The corpus comes from a MATLAB/Simulink study of one transmission line under stress: four generators rated 11 kV, arranged in pairs at each end of the line with transformers in between, and a fault point deliberately placed at the line midpoint. The circuit ran through healthy operation and through the full menu of line-to-line and line-to-ground disturbances while the measurement point recorded per-phase line currents and line voltages - roughly 12,000 labeled data points covering no-fault, LG, LL, LLG, LLL and LLLG conditions.

The intent is stated in the accompanying documentation: pattern recognition for protective relaying - fast, reliable discrimination between faulty and healthy states, and identification of which of the three phases is disturbed. The methodological lineage credits peer-reviewed work on artificial neural networks for fault classification, and its standing as the go-to teaching set for line-fault classification has made it one of the power-systems cornerstones of the Kaggle catalog.

Datadory packages the corpus as typed rows under Heavy Electrical Equipment, with the flag grammar documented and collapsible rather than left as four cryptic 0/1 columns. See the Kaggle source profile for how community benchmarks move through the catalog.

One row per simulated line condition, flat enough to read without a parser. The layout below shows the ten columns as they arrive, followed by the flag grammar that turns four bits into a diagnosis:

# ONE ROW PER SIMULATED LINE CONDITION -- columns as documented, values illustrative,
# confirmed against delivered rows during sample preparation

Ia       Ib       Ic       Va       Vb       Vc       G  C  B  A
-0.5     -0.4     -1.1     0.0      0.3      0.9      0  0  0  0    # healthy line
 2.10    -0.60    -0.70    0.00     0.40     0.50     1  0  0  1    # line A to ground (LG)
-0.30     1.80    -0.90    0.20     0.05     0.55     0  0  1  1    # line B to line C (LL)

# THE FLAG GRAMMAR -- FOUR BINARY COLUMNS RESOLVE SIX CONDITIONS
G C B A   ->  condition
0 0 0 0   ->  no fault (healthy line)
1 0 0 1   ->  line A to ground          (LG)
0 0 1 1   ->  line B to line C          (LL)
1 0 1 1   ->  lines B and C to ground   (LLG)
0 1 1 1   ->  three-phase               (LLL)
1 1 1 1   ->  three-phase to ground     (LLLG)

# SHAPE
rows        : ~12,000 simulated operating conditions
columns     : 10  (6 measurement channels + 4 fault flags)
scenario    : four 11 kV generators paired at each end of one line, transformers between
fault point : line midpoint

Numeric values above are illustrative of the corpus ranges; the column arrangement and every definition confirm against delivered rows during sample preparation. Your sample arrives cut to the fault classes you care about - all six, or just the ground-involving ones.

Ten columns, in two layers. A measurement layer - Ia, Ib, Ic, Va, Vb, Vc - holds the six instantaneous per-phase readings a relay actually sees: currents that swing on the conductor taking the hit, voltages that collapse on the faulted phase. A label layer - G, C, B, A - holds four binary flags marking ground participation and which conductors carried the fault; read together they resolve each row into one of six conditions.

That split is the whole pedagogical point. The measurement layer is the evidence, the label layer is the verdict, and a classifier's job is the argument between them. Column naming follows the corpus's standard layout and is confirmed against delivered rows during sample preparation rather than assumed silently.

• Geography: not applicable by design. This is a synthetic model of a generic transmission system, so there is no country, utility or substation dimension to bias a benchmark - and no confidentiality constraint anywhere in it.
• Temporal: a frozen study corpus. The ~12,000 conditions were captured as a single snapshot rather than accumulating over time, which is exactly what makes scores comparable across papers, teams and quarters.
• Granularity: one row per simulated operating or fault condition, each carrying six instantaneous per-phase readings and four fault flags. Roughly 12,000 rows by ten columns, compact enough to iterate on interactively.

API, files, or your warehouse. Daily, weekly, or hourly.

A corpus this size rewards the simplest posture: take all ~12,000 rows once as typed columns and iterate locally, with scheduled refreshes only if the collapsed-label or symmetrical-component extensions need to track your own feature pipeline. Files suit classroom labs and benchmark runners, API payloads suit products scoring batches of readings inside an app, and warehouse delivery suits teams joining fault signatures against their own asset registers in SQL. The ten-column dictionary travels unchanged across all three channels - changing cadence is a settings conversation, never a re-integration project.

Six jobs the corpus settles outright:

1. Protective-relaying model prototyping - train and stress classifiers on labels that are exact rather than inferred, so accuracy is measurable instead of argued; see the data scientists workflow.
2. Teaching power-system protection - students meet the waveform signature of each fault class without any utility data in the room.
3. Benchmark comparability - a frozen reference set keeps scores comparable across papers, teams and quarters.
4. Relay-logic regression testing - labeled conditions run through protection logic before hardware-in-the-loop, so obvious misoperations die in software.
5. Feature-engineering practice - symmetrical components, phase relationships and ratios built from six channels, with immediate feedback on class separation.
6. Course projects and interview take-homes - a complete, self-contained modeling exercise that fits a weekend and ships without legal review.

Data scientists and ML engineers get the rarest thing in applied classification work: exact multi-class labels on a tidy table that trains on arrival; see data scientists working in heavy electrical equipment. Journalists, academics and students get the corpus cited whenever protection-relaying arguments need a reproducible reference, with the flag grammar documented rather than reverse-engineered; see journalists and academics. Developers and builders get realistic fault signatures for simulators and demos - a line that actually misbehaves beats a sine wave every time; see developers and builders. Competitive-intelligence and product teams get a neutral bench for scoring protection-analytics vendors, and market researchers and consultants get concrete texture for grid-protection narratives expressed as data rather than adjectives.

Because the corpus arrives with opinions baked in: four anonymous 0/1 columns whose meaning lives in a forum post, headers worth confirming rather than trusting, and a class taxonomy everyone reconstructs by hand. We ship it typed against the ten-field dictionary above, with the flag grammar documented and collapsible into a single class label on request, per-unit normalization and symmetrical components available as engineered extensions on the same identifiers, and split presets that put the rare three-phase-to-ground cases into every fold.

Get a sample of this dataset scoped to the fault classes and extensions you care about, then put it on whatever cadence your project runs on.

Notes worth reading next:

• Electric Motor Temperature Dataset - fault physics on the rotating-machine side of the house; the sensor-panel counterpart to this line-fault corpus.
• EPRI Power Grid Equipment Research - utility-funded applied R&D where protection and diagnostics research lands once it leaves the classroom bench.
• Open Power System Data time series - hourly load, wind and solar series; the live-drift half of a monitoring story this static corpus anchors.
• EIA Form 860 generator inventory - the US installed base these fault scenarios abstract away.

Browse the vertical on the Heavy Electrical Equipment data hub or the ranked best heavy electrical equipment datasets. Before requesting a sample, decide the class scope: all six conditions, or the ground-involving subset - the flag columns make either cut deterministic rather than keyword-fuzzy.