Electric Motor Temperature Dataset

It is 185 hours of argument that a thermometer is optional. The dataset packages recordings from a permanent magnet synchronous motor - a German OEM's prototype - mounted on the test bench of Paderborn University's LEA department, excited by hand-designed driving cycles that mostly wander the speed-torque plane as random walks to imitate real-world duty.

What makes the panel useful is that every row holds the whole causal chain at once: the d/q voltage and current components commanded by standard field-oriented control, the speed and torque actually achieved, and the four temperatures that followed - permanent-magnet rotor temperature (pm), stator yoke, stator tooth and stator winding. Sessions run one to six hours and are keyed by profile_id; the April 2021 revision removed anonymization from every value and lifted coverage from 138 to 185 recorded hours by adding 17 measurement profiles. Adoption at research time ran to roughly 291,000 views and 590 upvotes.

Inside Datadory's 1,744-dataset catalog this record anchors the measured-hardware end of the eighteen-dataset Heavy Electrical Equipment slice, scored 8/10 against a catalog mean of 7.81. Amaze yourself with what is in it, then take the rows that matter: Get a sample of this dataset.

One delivered row per 2 Hz sample, flat enough to load without a parser:

# one delivered row per 2 Hz sample - shape below, values from the verified field examples
profile_id     : 72                   # measurement-session key; each session runs 1-6 h
u_q            : -0.628               # quadrature-axis voltage component
u_d            : 0.342                # direct-axis voltage component
i_q            : 12.41                # quadrature-axis current component
i_d            : -1.87                # direct-axis current component
motor_speed    : 1545.2               # rotor speed achieved under the driving cycle
torque         : 3.94                 # torque achieved under the driving cycle

# the four targets - what thermal models learn to predict
pm             : 24.55                # permanent-magnet (rotor) temperature - primary target
stator_yoke    : 24.02                # stator yoke temperature
stator_tooth   : 23.95                # stator tooth temperature
stator_winding : 24.31                # stator winding temperature

# rows repeat every half-second within a session; a new profile_id opens the next 1-6 h run
profile_id     : <next session id>

The values above are the verified examples attached to each documented field during the August 2026 research pass - shown so you can see the anatomy before committing. Read it as a story: the control strategy pushes with u_q, u_d, i_q and i_d, the machine answers with motor_speed and torque, and the four temperature channels record what the pushing cost. Half a second later, another row; when profile_id changes, a new one-to-six-hour run begins. Your sample arrives cut to the channels and sessions you name, in exactly this shape.

Eleven columns, three jobs. An excitation layer (u_q, u_d, i_q, i_d) states what the field-oriented control strategy imposed. A response layer (motor_speed, torque) records what the machine did about it. A target layer (pm, stator_yoke, stator_tooth, stator_winding) holds the temperatures - with pm carrying the commercial weight, because permanent-magnet temperature is precisely the quantity production motors cannot afford to instrument. profile_id ties every row to its one-to-six-hour session.

Every definition was verified against the published documentation rather than reconstructed from memory, which is why the examples carry observed values. Three gaps stay honestly open: the row total (~2 million) is an inference from duration and sampling rate rather than a counted figure, per-column engineering units go unstated upstream, and whether the 17 profiles added in 2021 reused identical sensor placement is undocumented. None of the three blocks modeling; all three belong in a validation pass, which is what the sample is for.

• Geography - deliberately narrow. One test bench at Paderborn University's LEA department, one PMSM prototype from a German OEM. This is a dataset you can know completely: every degree in it came off the same windings.
• Temporal - ~185 recorded hours accumulated across multi-hour measurement sessions, raised from 138 hours in the April 2021 revision that added 17 profiles. The corpus is closed; nothing newer arrives, so longitudinal claims come from scheduled deliveries diffed against each other rather than from the record itself.
• Granularity - 2 Hz time-series rows, one reading every half second, grouped into one-to-six-hour sessions by profile_id - roughly two million rows by inference. Dense enough to resolve thermal transients, coarse enough to stay a working panel rather than a firehose.

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

Because the corpus has a fixed length - 185 recorded hours, closed since the April 2021 revision - cadence buys synchronization rather than freshness. Take the whole panel once as typed columns for a benchmark study, or keep a warehouse table warm so every notebook on the team reads the same session ids; hourly feeds suit CI rigs that want the corpus ready for each training run. Deliveries arrive normalized to the eleven-column dictionary above with profile_id intact, so splits and joins hold without remapping.

Name the channels you model - the four temperature targets, the d/q electrical pairs, speed and torque - and the sample ships shaped to that scope before any commitment.

Five jobs the catalog settles outright:

1. Sensorless thermal estimation - train estimators that read rotor and stator temperatures from electrical signatures alone, retiring thermometers from places they are expensive or impossible to mount; pm is the headline target.
2. Thermal-network identification - fit lumped-parameter thermal models channel by channel from paired excitation and response.
3. Torque estimation for control efficiency - reproduce achieved torque from d/q quantities and study how much sensing a control loop truly needs.
4. Predictive-maintenance method benchmarking - a fixed 185-hour panel makes comparisons reproducible; see ml model training use cases.
5. Teaching physical machine learning - visible physics and clean column semantics make this the rare benchmark a course finishes end to end.

Data scientists and ML engineers lead fit at 3/3 - the record sits first in the industry's data-science ranking on the strength of measured hardware rather than simulation; see data scientists use cases. Developers and builders (2/3) stand predictive-maintenance demos on a self-contained eleven-column panel; see developers builders use cases. Journalists, academics and students (2/3) teach regression and thermal modeling on a documented public test bench; see journalists academics use cases. Competitive intelligence and product teams (1/3) benchmark condition-monitoring features against published recordings; see competitive intel product teams use cases. Investors and quants (1/3) cite test-bench modeling in industrial electrification diligence; see investors quants use cases. Market researchers and consultants (1/3) ground motor-efficiency analyses in a named Paderborn bench; see market researchers use cases.

This record owns the measured-hardware corner of the eighteen-dataset Heavy Electrical Equipment slice, and its neighbours mark the boundary precisely. Kaggle - Electrical Fault Detection & Classification covers failure modes but simulates them - about 12,000 modeled points against 185 hours of measured healthy operation, which makes the pair complementary rather than redundant. EPRI - Power Grid Equipment Research holds utility-funded applied R&D records; CIGRE Technical Brochures the peer-reviewed consensus literature; IEC Standards Webstore normative standard metadata at roughly 30,482 publications with 263 hits for power transformers.

Datadory rates this record 8/10 against a catalog mean of 7.81 across 1,744 datasets - a band only 1,096 records (62.8%) reach. For the macro view of the sector, the head-to-head against India's heavy electrical vertical lives on the Kaggle PMSM Motor Data vs India MHI comparison, and the full ranking sits on best heavy electrical equipment datasets.

Three things worth settling upfront.

First, scope by channel and session. Eleven columns and 185 hours sound small until a thermal model meets autocorrelated rows. Say which channels you model and how you want sessions grouped, and the sample arrives pre-split by profile_id rather than shuffled into leakage.

Second, know the documented gaps: the ~2 million row total is inferred rather than counted, per-column engineering units go unstated upstream, and sensor placement for the 17 profiles added in 2021 is not documented as identical. None is disqualifying; all three belong in the validation pass your sample exists to fund.

Third, decide whether you need faults. This corpus covers healthy operation across demanding duty cycles. If labeled fault conditions are part of the question, the simulated fault-current record beside it completes the pair - say so when requesting the sample and the scoping reflects both.