Power system fault detection dataset: what an ML engineer can actually train on

Two records answer the query, and they solve different halves of the fault problem. Kaggle - Electrical Fault Detection & Classification is the labeled bench: ~12,000 simulation points covering healthy operation plus line-to-line and line-to-ground disturbances on a modeled transmission line. It scores 6 of 10 on Datadory's rubric - modest, because the underlying circuit is synthetic rather than field-recorded. Its neighbor Kaggle - Electric Motor Temperature Dataset solves the gradual-degradation half: 185 hours of 2 Hz sensor recordings from a permanent magnet synchronous motor test bench, scored 8 of 10, suited to thermal-anomaly and predictive-maintenance modeling rather than relaying classification.

Together they are the only two machine-learning-ready sensor records among the 18 primary heavy electrical equipment datasets in Datadory's catalog; everything else in the pool is standards metadata, market statistics or marketplace listings. Start with the line-fault corpus to get a working classifier in an afternoon, then decide whether the motor series or the engineering literature below matches your actual deployment - and request a sample cut to the fault classes you care about before either decision hardens.

The authors built a MATLAB Simulink model of a transmission power system with four generators rated 11 kV, arranged in pairs at each end of a transmission line with transformers between them, and studied faults occurring at the midpoint of the line. They simulated normal conditions and phase-to-phase and phase-to-ground combinations, recording measured line voltages and line currents at the output side of the system. Nearly 12,000 points were collected and labeled with fault type.

Ten columns define the schema, in two layers. Ia, Ib and Ic carry the phase A, B and C line currents at the measurement point; Va, Vb and Vc carry the corresponding line voltages. On top sit four boolean flags - G, C, B, A - marking ground or phase involvement, combinable into the LL, LLG, LLL and LLLG fault classes plus no-fault rows (the combination 1,0,0,1 reads as a line-A-to-ground fault). The measurement layer is the evidence, the flag 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 the field dictionary during sample preparation rather than assumed silently.

The intent stated in the accompanying documentation is pattern recognition for protective relaying - fast discrimination between faulty and healthy states, and identification of which phase is disturbed - with a methodological lineage crediting peer-reviewed work on artificial neural networks for fault classification. Coverage notes matter for anyone writing methods sections: geographic scope is not applicable because the system is synthetic, granularity is one row per simulated operating condition, and the whole corpus stands as a single frozen study snapshot, which is exactly what makes benchmark scores comparable across papers, teams and quarters.

The corpus arrives with opinions baked in - four anonymous 0/1 columns whose meaning has to be made explicit before any metric means anything:

1. Derive the target explicitly. Combine G, C, B and A into a single class label - no-fault, LG, LL, LLG, LLL, LLLG - or keep the binary encoding and let the model learn the grammar. Decide once, before any split, so runs stay comparable.
2. Split by scenario, not by row. Rows were generated by simulating each fault scenario, so shuffling blindly leaks near-duplicate conditions across train and test. Split by scenario if you want honest generalization claims.
3. Respect the geometry. Every faulted row was recorded at the line midpoint, a constraint no classifier can learn around - treat distance-to-fault as out of scope rather than hoping a feature absorbs it.
4. Expect rarity at the extreme. Three-phase-to-ground cases are the minority class; stratify or oversample so the condition your protection logic cares about most is not the one with the widest confidence interval.
5. Carry the lineage forward. The work descends from ANN-based protective-relaying pattern recognition and peer-reviewed neural-network fault classification; citing that lineage keeps methods sections defensible.

None of this requires owning a parser. Rows arrive typed against the ten-field dictionary, with the flag grammar documented and a collapsed single-label variant available as an engineered extension on the same identifiers.

Fault detection splits into discrete-event classification and gradual-degradation detection, and the second problem needs time series, not snapshots. Kaggle - Electric Motor Temperature Dataset provides exactly that: 185 hours of recordings from a permanent magnet synchronous motor - a German OEM's prototype - on a test bench at the LEA department of Paderborn University, sampled at 2 Hz across measurement sessions identified by profile_id, each session running one to six hours.

Documented columns cover d/q-axis currents (i_d, i_q) and voltages (u_d, u_q) from a standard field-oriented control strategy, achieved motor_speed and torque under hand-designed driving cycles that imitate real-world duty cycles, and four temperature targets: pm (rotor permanent magnet), stator_yoke, stator_tooth and stator_winding. An April 2021 revision expanded coverage from 138 to 185 hours by adding 17 profiles - roughly two million rows in one consistent panel.

Pick by task: relaying-style fault classes take the line-fault corpus; thermal drift, sensor-drift anomalies and predictive maintenance take the motor series. The two join cleanly - both arrive as typed rows keyed on session and profile identifiers, so a model trained on one borrows its preprocessing from the other.

Treat these as the citation and compliance envelope: the corpus teaches a classifier what a line-A-to-ground event looks like, and these three tell reviewers why that matters.

Simulated faults describe an idealized circuit; deployment happens on a fleet you can actually inventory. Two government records give that grounding. EIA Form EIA-860 — Annual Electric Generator Inventory is an annual census of every US power plant at 1 MW+ nameplate capacity - operable, proposed and retired generators, with plant, utility, owner and prime-mover detail - published as annual editions reaching back to 2001, and it scores 10 of 10 on Datadory's rubric. data.gov - US Federal Transformer Datasets aggregates the transformer angle across a harvest of 552,271 US datasets, surfacing EIA-860 form data, FERC filings, DOE distribution-transformer efficiency rulemaking materials and utility asset inventories as one record per published dataset.

Neither source contains waveforms - they tell you how many generators and transformers exist, who owns them and how they are configured. That is precisely what a deployment scoping section needs: the line-fault corpus teaches the classifier what a fault signature looks like, and EIA-860 says how many units such a classifier would ever monitor.

Only the two sensor corpora serve model training directly; the rest frame the literature, enforce the standards or size the deployment. The table ranks the options a fault-modeling project actually touches.

This cluster is the modeling leg of a broader pool. The heavy electrical equipment data guide maps all 30 pooled datasets with scale, coverage and personas compared, and pairs naturally with electric motor temperature dataset csv for the thermal-sensor half of equipment ML. Full field lists, sample rows and coverage chips live on the Kaggle - Electrical Fault Detection & Classification and Kaggle - Electric Motor Temperature Dataset records, and every source cited here is indexed in the heavy-electrical-equipment data hub - then request a sample cut to the fault classes and extensions you care about.