Want to build a machine learning project, but tired of predicting who survived the Titanic? Fair enough. The Titanic is a great story, but it will not impress anyone who hires engineers.

The good news: mechanical engineers have some of the most interesting free datasets around. Vibrating bearings, degrading jet engines, worn cutting tools, ageing batteries. Real machines, real sensors, real mess.

This guide lists 15 free datasets I checked against their official pages. For each one you get what it contains, what you can predict with it, and how hard it is. If you are still deciding what to learn first, start with our ML roadmap for mechanical engineers, then come back here to pick a project.

One honest note: I have not run models on all of these, and I am not quoting accuracy scores. File sizes and terms can change, so always read the source page before you download.

Quick glossary (so nothing below is confusing)

• RUL (remaining useful life): how many more hours or cycles a machine can run before it fails.
• Regression: predicting a number, such as power output in megawatts.
• Classification: predicting a label, such as "healthy" or "inner race fault".
• Anomaly detection: spotting data that does not look normal, often with few or no failure examples.

Machine health and failure datasets

This is the classic territory for mechanical engineers: predicting when a machine will misbehave. Several of the first four come from one place, the NASA Prognostics Center of Excellence repository.

1. NASA C-MAPSS turbofan engine degradation

A simulation of jet engines wearing out under different operating conditions. You get sensor readings over time, and the job is to predict remaining useful life. NASA lists a second version too (Turbofan Engine Degradation Simulation-2) with run-to-failure trajectories under flight conditions. Level: intermediate, because you must handle time series properly. Find it on the NASA PCoE data repository.

2. IMS bearing run-to-failure data (NASA repository)

Bearing failure experiments from the University of Cincinnati’s Center for Intelligent Maintenance Systems. Good for anomaly detection and early-warning ideas, because you can watch a healthy bearing slowly become a bad one. Level: intermediate. Listed as "Bearings" on the same NASA repository page.

3. FEMTO (PRONOSTIA) bearing dataset

Accelerated bearing life tests from the FEMTO-ST Institute in France, also hosted in the NASA repository (listed as "FEMTO Bearing"). Because the bearings are aged on purpose, you get full life histories quickly. Typical task: RUL prediction. Level: intermediate.

4. NASA milling (tool wear)

Machine wear measurements at different cutting speeds, feeds and depths of cut, from the UC Berkeley BEST Lab. A very natural fit if you like manufacturing: predict tool wear from the cutting conditions and sensor signals. NASA also lists a separate CNC milling dataset for RUL estimation using dynamometer, accelerometer and acoustic emission sensors. Level: intermediate. Source: the same NASA repository.

5. Case Western Reserve University (CWRU) bearing data

Vibration data from a 2 hp Reliance Electric motor with deliberately seeded faults of 0.007 to 0.040 inches, placed on the inner raceway, the balls and the outer raceway. Tests ran at loads from 0 to 3 hp, with motor speeds between 1720 and 1797 RPM. Typical task: fault classification. It is hugely popular, so tutorials are easy to find. Level: beginner to intermediate. Source: the CWRU Bearing Data Center. Check the site for current terms.

6. Paderborn University bearing dataset

Vibration and motor current signals from a test rig run under different operating conditions, with bearing damage categorised systematically. Typical task: classifying bearing damage. The licence is CC BY-NC 4.0, which means non-commercial use with credit, and you are asked to cite the Lessmeier et al. benchmark paper. Level: intermediate. Source: the KAt Bearing DataCenter.

7. AI4I 2020 Predictive Maintenance

10,000 rows and 14 columns, with temperatures, rotational speed, torque, tool wear and a machine-failure label covering five failure modes. Important: it is a synthetic dataset, made to resemble real industrial data. That makes it tidy and beginner-friendly, but do not treat it as proof your model works on a real factory floor. Licence: CC BY 4.0. Level: beginner. Source: the UCI repository.

8. MIMII industrial machine sounds

Normal and abnormal sounds from valves, pumps, fans and slide rails, recorded with an eight-channel microphone array at 16 kHz, with factory background noise mixed in. Typical task: sound-based anomaly detection. Warning: the full set is about 100.2 GB, so check your disk space and data plan. Licence: CC BY-SA 4.0. Level: intermediate to advanced. Source: Zenodo.

Energy, power and thermal datasets

9. Combined Cycle Power Plant

9,568 hourly readings over six years (2006 to 2011): ambient temperature, pressure, humidity and exhaust vacuum, used to predict the plant’s net electrical output in megawatts. No missing values, four inputs, one target. A textbook regression problem. Licence: CC BY 4.0. Level: beginner. Source: UCI.

10. Gas Turbine CO and NOx Emission

36,733 rows of hourly sensor aggregates from a gas turbine, covering 2011 to 2015, with ambient variables, turbine parameters and CO and NOx emissions. Good for regression on emissions or energy yield. Licence: CC BY 4.0. Level: intermediate. Source: UCI.

11. Steel Industry Energy Consumption

35,040 records from a steel company in South Korea: energy use in kWh, reactive power, CO2 emissions, power factors, time of day and day type. Predict energy consumption, and learn why time features matter. Licence: CC BY 4.0. Level: beginner. Source: UCI.

Battery datasets

Batteries are a mechanical-and-electrical crossover, and a hot topic for EV careers. If you want the big picture first, read our post on how an electric car estimates its battery state of charge.

12. Panasonic 18650PF Li-ion battery data

Tests of a 2.9 Ah Panasonic cell at the University of Wisconsin-Madison: HPPC pulses, drive cycles and impedance spectroscopy at five temperatures. Built for state-of-charge estimation and battery modelling. Check the source page for current licence terms. Level: intermediate. Source: Mendeley Data.

13. CALCE battery data

Open test data from the University of Maryland on cylindrical, prismatic and pouch cells (LCO, LFP and NMC chemistries): cycling, storage, dynamic driving profiles and impedance. Files come as Excel, text and zip. Users should cite the CALCE articles listed for each dataset. Typical task: capacity fade and health estimation. Level: intermediate. Source: CALCE Battery Data.

Aerodynamics and vehicle datasets

14. UCI Airfoil Self-Noise

1,503 rows from NASA wind-tunnel tests of airfoil sections. Five inputs (frequency, angle of attack, chord length, free-stream velocity, suction-side displacement thickness) predict the scaled sound pressure level in decibels. Small, clean, and a great first fluid-mechanics regression. Licence: CC BY 4.0. Level: beginner. Source: UCI.

15. Auto MPG

398 cars with cylinders, displacement, horsepower, weight, acceleration, model year and origin, used to predict fuel consumption in miles per gallon. The horsepower column has missing values, which makes it a gentle lesson in data cleaning. It is old, so use it for learning, not for claims about modern cars. Licence: CC BY 4.0. Level: beginner. Source: UCI.

Comparison table

Dataset	What it is	ML task	Level
NASA C-MAPSS	Simulated turbofan degradation	RUL prediction	Intermediate
IMS bearings	Bearing failure experiments	Anomaly detection, RUL	Intermediate
FEMTO (PRONOSTIA)	Accelerated bearing life tests	RUL prediction	Intermediate
NASA milling	Tool wear at varied cutting settings	Wear regression	Intermediate
CWRU bearings	Vibration with seeded faults	Fault classification	Beginner to intermediate
Paderborn bearings	Vibration and motor current	Damage classification	Intermediate
AI4I 2020	Synthetic machine failure data	Failure classification	Beginner
MIMII	Industrial machine sounds	Anomaly detection	Intermediate to advanced
Combined Cycle Power Plant	Hourly plant readings	Power regression	Beginner
Gas Turbine CO and NOx	Hourly turbine and emission data	Emission regression	Intermediate
Steel Industry Energy	Steel plant electricity use	Energy regression	Beginner
Panasonic 18650PF	Li-ion cell lab tests	State-of-charge estimation	Intermediate
CALCE batteries	Li-ion ageing and cycling	Capacity and health estimation	Intermediate
Airfoil Self-Noise	Wind-tunnel noise tests	Noise regression	Beginner
Auto MPG	Car specs and fuel use	MPG regression	Beginner

Which dataset should I start with?

• Brand new to ML: Auto MPG or Airfoil Self-Noise. Small, clean, and perfect for practising linear regression.
• Want a portfolio project: Combined Cycle Power Plant, then something with time, like C-MAPSS.
• Love rotating machinery: CWRU first, then IMS or FEMTO for run-to-failure work.
• Into EVs: Panasonic 18650PF or CALCE.
• Manufacturing fan: NASA milling for tool wear.

New to code as well? Sort out your setup with our Python basics guide first. It takes ten minutes and saves hours of frustration.

Tips before you download

• Split by time, not randomly. If your data is a time series, train on the earlier part and test on the later part. Random shuffling lets the model peek at the future.
• Watch for data leakage. Leakage means information sneaks into training that would not exist in real use. For run-to-failure data, keep all rows from one machine or bearing in the same split, otherwise your test score will look far better than reality.
• Read the licence and cite the source. Most UCI sets above are CC BY 4.0, Paderborn is non-commercial, and MIMII is share-alike. Follow each page’s citation request.
• Check the units. Inches or metres, Wh or kWh, psi or kPa. A unit mix-up is the most mechanical-engineer mistake there is.
• Know if the data is synthetic. AI4I and C-MAPSS are simulated. Great for learning, but say so in your write-up.

Frequently asked questions

Can I use these datasets in a college project or portfolio?

Generally yes, if you follow each licence and cite the source. Paderborn is non-commercial, so check before using it in anything paid. When in doubt, read the source page.

Do I need a powerful computer?

Not for the small ones. Auto MPG, Airfoil and Combined Cycle Power Plant run on a basic laptop or free Google Colab. MIMII is the exception, at roughly 100 GB in full.

Are synthetic datasets useless?

No. They are excellent for learning the workflow. Just be honest that results on synthetic data do not guarantee results on a real machine.

Final thought

You do not need a factory to do machine learning on machines. You need one good dataset, a clear question and the patience to check your units. Pick one from the list, finish a small project end to end, and write down what went wrong. That write-up is worth more than any certificate.

Which of these datasets will you try first, and what would you like to predict with it? Tell us in the comments.