A motor bearing rarely fails without warning. Days before it seizes, it vibrates differently, runs hotter, and draws current in a slightly odd pattern. The problem is that no human is standing at every machine listening for it. Predictive maintenance puts a sensor there instead — one that watches continuously and raises a flag while there's still time to act.
This guide explains why reactive maintenance costs so much, how IoT sensing plus edge machine learning predicts failures, and how an Indian SMB can pilot it affordably.
What's wrong with reactive maintenance?
Most small and mid-size plants in India run on one of two models, and both are expensive:
- Reactive ("run to failure") — fix it when it breaks. The breakdown always comes at the worst time, takes the line down unplanned, damages neighbouring parts, and forces overtime and rush-order spares.
- Scheduled ("calendar-based") — service everything every N months whether it needs it or not. You replace healthy parts early and waste labour, yet still miss the failure that happens between service dates.
Both treat every machine the same. Predictive maintenance treats each machine by its actual, measured condition.
How does IoT predict a failure?
Three physical signals reveal a developing fault, and cheap sensors can read all three:
- Vibration — a high-rate accelerometer captures the tiny repeating shocks a damaged bearing, imbalance or misalignment produces. This is the richest early signal.
- Temperature — friction from wear shows up as rising heat well before total failure.
- Current — the motor's current draw shifts as mechanical load changes, and on many motors it also gives a reliable read of shaft speed.
A controller samples these signals, extracts features that correlate with specific fault types, and compares them against what "healthy" looks like for that exact machine.
The faults these signals reveal
The value isn't just "something's wrong" — it's what's wrong, which tells a technician what to fix. Vibration signatures in particular separate the common rotating-equipment faults:
- Bearing wear — the most common failure, showing up as characteristic high-frequency vibration that grows as the bearing degrades.
- Imbalance — vibration synchronised with the shaft's rotation, from a worn or fouled rotor or fan.
- Misalignment — coupled shafts that aren't true, producing their own distinct pattern.
- Looseness — mounting or structural slack that worsens over time.
Catching the type of fault early means the right part is ordered and the right fix is planned, instead of a technician opening the machine mid-breakdown to find out what failed.
What is edge ML (TinyML), and why run it on the device?
TinyML means running a machine-learning model directly on a small, low-power microcontroller at the machine — not streaming raw data to the cloud. It matters for three reasons:
- Bandwidth and cost — raw vibration data is huge; sending only a health verdict and a few features keeps data and connectivity cheap.
- Resilience — the device keeps working through a network or power dip and buffers until it reconnects.
- Privacy and latency — the decision happens on-site, in near real time.
In SVB Tech Labs' own R&D, a fault classifier was compressed to under 3 KB and runs on an ESP32-S3 microcontroller, producing the same results on the device as on a laptop — small enough to sit on the machine, capable enough to diagnose the fault type.
Does it actually work? Our R&D findings
We validated this approach on public run-to-failure datasets (including NASA's bearing rigs) before building any hardware. A few results worth knowing:
- On four real bearing failures, the system raised a sustained alarm 42 to 120 hours ahead of the actual failure — days of warning, not minutes.
- It judges each machine against its own baseline rather than a generic threshold. That cut false alarms dramatically: population-style thresholds flagged 8–18% of healthy readings, while the per-machine approach produced zero false positives on healthy machines and well under one nuisance alert per machine-month.
- Alerts escalate cleanly through Watch → Warning → Critical, so a technician can plan the repair rather than scramble.
These are controlled R&D figures on reference data, not a client guarantee — real machines and real plants vary, which is exactly why the system learns each installation individually and why a pilot matters.
What does a sensor node actually look like?
A monitoring node is deliberately simple so it's cheap to deploy and rugged enough to live on a machine: a microcontroller (our R&D uses the ESP32-S3), an accelerometer for vibration, a temperature sensor, optionally a current sensor clamped on the motor supply, and a power feed. It mounts on or near the machine, runs the classifier on-device, and reports over Wi-Fi or another available link.
Because the intelligence runs locally, the node doesn't need a fat, always-on connection or a nearby PC — it sends a compact health verdict, not a raw data firehose. That's what makes a multi-machine rollout affordable: each node is a small, self-contained unit, and the network only carries results.
What's the ROI for an SMB?
The payoff is avoided unplanned downtime. One prevented line stoppage — the lost production, the rushed spare, the overtime, the collateral damage to connected parts — typically dwarfs the cost of a sensor node. Predictive maintenance also extends the useful life of healthy parts you'd otherwise swap on a calendar, and turns emergency repairs into scheduled ones. For a plant where a stopped machine means stopped revenue, catching the failure days early is the whole game.
To make the ROI real, write down two numbers before you start: your unplanned downtime hours over the last year, and your spend on emergency repairs and rush-order spares. Those are the baseline the pilot is measured against. If a handful of sensor nodes prevents even one of last year's stoppages, the programme is already ahead — and every prevented failure after that is clear gain.
How do I pilot predictive maintenance?
Keep the first step small and measurable:
- Pick your most critical or most trouble-prone machine — the one whose failure hurts most.
- Screen it at install. A healthy baseline only works if the machine is healthy when you start, so check its condition first.
- Let it learn, then watch. The system learns that machine's normal signature over a short baseline period, then monitors against it.
- Track the warnings against reality for a few weeks, and compare downtime and repair cost to your previous baseline.
- Roll out to the next machines once the pilot proves its numbers.
Frequently asked questions
What is predictive maintenance?
Predictive maintenance uses sensors and analytics to judge a machine by its actual, measured condition and predict a failure before it happens — so you fix it on a planned schedule rather than after it breaks (reactive) or on a fixed calendar regardless of condition (preventive).
How much warning does it give?
It depends on the fault and the machine, but meaningful warning is measured in hours to days, not minutes. In our R&D on reference bearing datasets, sustained alarms fired roughly 42 to 120 hours ahead of failure — enough to plan the repair into a break.
Do I need to send data to the cloud?
No. With edge ML (TinyML), the model runs on the device at the machine and sends only a health verdict, so it keeps working through network or power dips and keeps data and connectivity costs low. The cloud is optional, for dashboards and trends.
Which machines is it best for?
Rotating equipment — motors, pumps, fans, compressors, gearboxes — where vibration, temperature and current reveal developing faults. Start with your most critical or most trouble-prone machine, the one whose failure costs you the most.
Conclusion
Predictive maintenance replaces guesswork and calendars with the machine's own data. With affordable IoT sensing and on-device TinyML, days of warning before a failure is realistic — and for an SMB, one avoided breakdown can pay for the whole programme. Start with one critical machine and let the results decide the rest.
Want to pilot predictive maintenance on a critical machine? Explore our Machine Health Monitor or see the full AIoT & smart automation offering from SVB Tech Labs.