Predictive Maintenance for Kids: How a Machine Warns You

A bearing that is starting to fail makes a sound at a specific, calculable frequency. Not a vague rattle. A rate you can work out in advance from the number of balls inside it, their diameter, the diameter of the ring they run on, and how fast the shaft turns. Every time a ball rolls over the damaged spot, it produces an impact, and those impacts arrive at that rate. Explaining predictive maintenance for kids starts there, because once a child understands that a machine’s complaint has a frequency, the rest of the field follows.

Key Takeaways

• Predictive maintenance means acting on the condition of a machine rather than on a calendar. The signal is a change from a known baseline.
• A damaged rolling-element bearing generates impacts at a frequency determined by its geometry and shaft speed, which is why engineers look for energy at specific frequencies rather than just “more noise.”
• Four main modalities: vibration, thermal imaging, oil analysis for wear particles, and ultrasonic or acoustic emission. Each sees a different failure mode.
• Machine learning does not replace the physics. It learns a normal envelope, flags deviation, and estimates remaining useful life, which requires having observed failures before.
• NASA’s Prognostics Center of Excellence publishes free run-to-failure datasets, including University of Cincinnati bearing experiments and C-MAPSS turbofan degradation simulations, so a teenager can do real analysis on real data.

Three ways to maintain anything

Reactive. Run it until it breaks, then fix it. Cheapest in planning, most expensive in consequences: unplanned downtime, collateral damage, emergency labour rates. Appropriate for a $4 light bulb. Inappropriate for the only pump in a building.

Preventive. Replace or service on a fixed schedule, regardless of condition. This is your car’s oil change interval. It prevents most failures and wastes life on components that were still fine, and it can introduce failures, because every time you open a machine you risk assembling it slightly wrong.

Predictive. Measure something that correlates with health, watch it over time, and intervene when the trend says so. More sensors, more analysis, and no wasted component life.

Most real operations run all three at once, matched to consequence. The interesting question is never “which is best” but “which does this component deserve.”

Why a failing bearing has a frequency

This is the mechanism worth getting right, because almost all popular explanations skip it.

A rolling-element bearing has an outer ring, an inner ring, and a set of balls or rollers between them held by a cage. When a tiny patch of the outer ring spalls — a flake of metal lifts away — a ball rolling over that patch produces a short, sharp impact. The next ball produces another.

How often? That depends on how often a ball passes any fixed point on the outer ring, which is set by the number of balls, their diameter, the diameter of the circle they travel on, the contact angle, and the shaft’s rotational speed. All of those are known from the bearing’s part number. So the impact rate can be calculated before anything breaks.

That is the whole trick. An engineer monitoring a machine is not listening for “a bad noise.” They are computing a handful of expected defect frequencies for each bearing, then watching for energy to appear at those specific frequencies in the vibration spectrum. A rise at the outer-race frequency says outer race. A rise at the inner-race frequency says inner race. Different numbers mean different parts.

The same logic applies to gear teeth, which produce a tooth-meshing frequency, and to electric motor faults, which show up as sidebands around the line frequency.

What each sensing method actually sees

Method	What it measures	Failure modes it catches early	Where it is weak
Vibration analysis	Acceleration, velocity or displacement over time, converted to a frequency spectrum	Bearing spalls, imbalance, misalignment, gear tooth damage, looseness	Needs good sensor mounting and a baseline; slow-speed machines are hard
Thermography	Surface temperature from infrared emission	Electrical connection faults, overloaded motors, bearing friction, insulation loss	Sees only surfaces; emissivity and reflections cause errors
Oil analysis	Wear particles, viscosity, water and additive depletion in lubricant	Internal metal-to-metal wear long before noise appears	Requires sampling discipline and a laboratory
Ultrasonic or acoustic emission	High-frequency sound from friction, leaks and electrical discharge	Early bearing lubrication problems, compressed air and steam leaks, partial discharge	Highly directional; background noise interferes
Motor current signature	Electrical current drawn, analysed for frequency content	Rotor bar damage, mechanical load changes, some bearing faults	Indirect; needs a known electrical baseline

The point of putting these in one table is that no single method sees everything. A plant that only does vibration will miss an electrical connection heating up. A plant that only does thermography will miss a bearing that is dry but not yet hot.

Where the machine learning actually goes

It would be easy to say “AI predicts failures,” and that sentence hides the only interesting part.

What a model usually does is one of two jobs. The first is anomaly detection: learn what the spectrum of this machine looks like when it is healthy across a range of loads and temperatures, then flag when today’s reading falls outside that envelope. This needs lots of healthy data and no failure examples, which is convenient because healthy data is what every factory has.

The second job is remaining useful life estimation. Given the degradation trajectory so far, how many operating hours remain before this component crosses a failure threshold? This needs examples of components that actually went all the way to failure, which is expensive and rare, and is exactly why public datasets matter.

NASA’s Prognostics Center of Excellence maintains a free data repository of exactly that: time-series data tracking systems from a prior nominal state to a failed state. Its collection includes bearing run-to-failure experiments donated by the University of Cincinnati’s Center for Intelligent Maintenance Systems, turbofan engine degradation simulations produced with the Commercial Modular Aero-Propulsion System Simulation under various operating conditions and fault modes, lithium-ion battery impedance experiments across charge and discharge temperatures, and milling insert wear at varying speeds, feeds and cutting depths. NASA notes plainly that users employ the data at their own risk.

Two honest limits. First, a model that has never seen a failure mode cannot predict it; it can only notice that something is unusual. Second, every alarm threshold trades false alarms against missed failures, and that trade has no free lunch. Schönherr and colleagues demonstrated the same arithmetic in a different domain in 2020, finding hundreds of accidental triggers across 11 smart speakers from 8 manufacturers because detection thresholds had been tuned toward sensitivity. A maintenance alarm tuned the same way sends technicians to healthy machines until they stop trusting it.

Also worth saying: the dramatic cost-savings percentages that circulate in predictive-maintenance marketing usually trace back to vendor material rather than peer-reviewed measurement. The engineering case is solid. The specific percentage in any given slide deck probably is not, and this article is not going to repeat one.

How to Teach Your Kid About Predictive Maintenance

Ages 5–8: the listening walk

Walk the house and listen to things that run: the fridge, a fan, the washing machine, a bike wheel spun by hand. For each one, ask them to describe the sound in words and then to predict what it would sound like if it were broken. Come back a month later and listen again. You are teaching the idea of a baseline, which is the entire foundation.

Ages 9–12: make the fault on purpose

Spin a bike wheel and listen. Now tape a coin to one spoke and spin it again. The wheel is now unbalanced, and the vibration arrives once per revolution, which they can feel through the frame and count. Add a second coin opposite the first and the imbalance mostly cancels. They have just produced, diagnosed and corrected a once-per-revolution fault, which is the most common machine fault in the world.

Ages 13+: plot a real bearing failing

Download one of the University of Cincinnati bearing run-to-failure sets from the NASA repository and plot a simple statistic over time: root mean square of the vibration signal, or its kurtosis, which is a measure of how spiky a signal is. Both rise as the bearing degrades. Then ask the hard question: at which point on that curve would you have ordered a replacement, and what did waiting longer buy you? Our walkthrough of building a sensor that watches one thing is the right warm-up, because a self-built sensor makes the data concrete.

The question to ask: “If this machine sounds exactly the same as yesterday, is it healthy?”

The correct answer is “not necessarily, because some failures do not change sound at all.” Oil contamination, insulation degradation and electrical connection heating are all silent. A good answer names a different sense.

What to do at home

Keep a baseline log for three household machines

Car, washing machine, heating system. Write down one sentence a month: unusual noises, where, under what condition. When something changes, you will know it changed, which is more than most households can say. A change from baseline is the signal; the absolute value almost never is.

Use your phone as a crude vibration sensor

Most phones have an accelerometer and free apps that plot its output. Tape the phone to a washing machine during a spin cycle and look at the trace. You will see the once-per-revolution component clearly. This is genuinely the same measurement an industrial sensor makes, at lower quality.

Teach the difference between a symptom and a cause

A hot bearing is a symptom. The cause might be lack of lubricant, misalignment, overload or a failing seal. Replacing the bearing without finding the cause gives you the same failure again in four months. Children are naturally good at “why” chains and this is a satisfying place to use one.

Point an interested teenager at the occupational data

U.S. installation, maintenance and repair occupations are projected to grow faster than average from 2025 to 2035, with roughly 569,000 annual openings. Industrial machinery mechanics, machinery maintenance workers and millwrights had a 2025 median of $64,100; calibration technologists and technicians $67,820. Meanwhile production occupations are projected to show little or no change, with assemblers and fabricators at a $45,450 median. Our fuller treatment is in the factory jobs that actually exist in 2026.

Connect it to the robots

Agility Robotics’ Digit 5, covered by IEEE Spectrum on September 15, 2026, is sold with a service model estimated around $8,500 a month. That service line is predictive maintenance as a business model: someone is monitoring actuators, bearings and batteries and intervening before a robot stops a production line.

What not to do

Do not let a child conclude that more sensors is always better. A sensor nobody looks at, or one with a threshold set so loose that every alarm is ignored, is worse than no sensor, because it creates confidence without information. Our career-focused piece on the IoT engineers behind machines that know when they will break covers how that work is actually organised.

What to Watch For Over the Next 3 Months

• Week 4: Start the baseline log. One sentence per machine. The value is entirely in having the earlier entry to compare against, so month one feels pointless and month six does not.
• Month 2 red flags: A new noise that correlates with load rather than appearing randomly. A warm smell near a motor. A puddle that was not there. All three are classic early indicators, and all three are usually noticed and then forgotten.
• Month 3 self-check: Did anyone act on something before it failed? Tightened a belt, topped up oil, called a technician about a noise. That is the behaviour the whole discipline exists to produce, and it is as valid for a washing machine as for a turbine.

Frequently Asked Questions

What is predictive maintenance in simple terms?

Fixing something based on its measured condition rather than on a schedule or after it breaks. You watch a signal that tracks health, such as vibration or temperature, and act when it drifts away from its normal baseline.

How can a sensor know a bearing is about to fail?

Because a damaged bearing produces impacts at a frequency set by its geometry and shaft speed. The number of balls, their diameter, the diameter of the ring they run on, the contact angle and the rotation rate determine how often a ball passes the defect, so engineers look for energy at that specific calculated frequency.

Does predictive maintenance need artificial intelligence?

No. Vibration analysis predates machine learning by decades and works through physics and frequency analysis. Models help with anomaly detection across large fleets and with estimating remaining useful life, but they need examples of real failures, which is why public datasets like NASA’s matter.

Can we try this at home?

Yes, at a basic level. Tape a phone to a washing machine during the spin cycle and plot the accelerometer trace, or spin a bike wheel with a coin taped to one spoke to create and then cancel an imbalance. Both reproduce the real measurement.

Is this a good career for my kid?

The occupational data is encouraging. U.S. installation, maintenance and repair work is projected to grow faster than average through 2035, with about 569,000 annual openings, and medians from $64,100 for industrial machinery mechanics to $67,820 for calibration technicians. Entry is usually through apprenticeship or a two-year programme rather than a four-year degree.

---

About the author

Ricky Flores is the founder of HiWave Makers and an electrical engineer with 15+ years of experience building consumer technology at Apple, Samsung, and Texas Instruments. He writes about how kids learn to build, think, and create in a tech-saturated world. Read more at hiwavemakers.com.

---

Sources

1. NASA. “Prognostics Center of Excellence Data Set Repository.” NASA Intelligent Systems Division. https://www.nasa.gov/intelligent-systems-division/discovery-and-systems-health/pcoe/pcoe-data-set-repository/
2. U.S. Bureau of Labor Statistics. “Installation, Maintenance, and Repair Occupations.” Occupational Outlook Handbook, last modified August 27, 2026. https://www.bls.gov/ooh/installation-maintenance-and-repair/home.htm
3. U.S. Bureau of Labor Statistics. “Production Occupations.” Occupational Outlook Handbook, last modified August 27, 2026. https://www.bls.gov/ooh/production/home.htm
4. U.S. Bureau of Labor Statistics. “Industries at a Glance: Manufacturing (NAICS 31-33).” Data reference September 2026. BLS. https://www.bls.gov/iag/tgs/iag31-33.htm
5. IEEE Spectrum. (2026, September 15). “Digit 5 May Be the First Humanoid Robot Worker That’s Truly Safe.” IEEE Spectrum. https://spectrum.ieee.org/humanoid-robot-safety
6. Schönherr, L., Golla, M., Eisenhofer, T., Wiele, J., Kolossa, D., & Holz, T. (2020). “Unacceptable, where is my privacy? Exploring Accidental Triggers of Smart Speakers.” arXiv:2008.00508. https://arxiv.org/abs/2008.00508