Unplanned downtime costs manufacturers a fortune per hour — and most of it is announced in advance, in vibration and current signatures that an embedded sensor could have flagged weeks earlier. Predictive maintenance is where edge AI earns its keep most directly.
This guide covers the sensing, the models, and the deployment realities of condition monitoring on the factory floor.
What faults look like in data
Rotating machinery announces its failures. Bearing wear shows up as energy in characteristic defect frequencies; misalignment adds harmonics of the rotation speed; imbalance multiplies the fundamental; looseness smears broadband noise. Current draw (MCSA — motor current signature analysis) reveals electrical faults and load changes without touching the machine.
The signature lives in the frequency domain — raw time-series vibration is noise; the FFT spectrum is where faults speak. Feature extraction (band energies, crest factor, kurtosis, harmonic ratios) converts a few kHz of vibration into a handful of numbers a small model can reason about.
Anomaly detection first, classification second
Real fleets don't have labeled failure data — machines that actually failed while instrumented are rare. The standard progression: deploy anomaly detection trained on normal operation (the healthy baseline), which flags anything unusual; then, as flagged events accumulate and get diagnosed, build supervised classifiers for the fault types you actually see.
Per-machine baselining matters more than model sophistication. Every motor has its own normal — mounting, load, age. Models that learn machine-level baselines, or normalize against machine context, flag faults instead of flagging 'this machine sounds different from the training fleet'.
- Baseline: weeks of healthy data per machine type
- Stage 1: anomaly detection on vibration/current features
- Stage 2: supervised classifiers as diagnosed events accumulate
- Per-machine normalization beats fleet-average models
Why on-device, specifically on the factory floor
Factories are hostile RF environments — shielded concrete, metal enclosures, interference — and IT policy often forbids production data leaving the building. Edge inference answers both: the node analyzes continuously, reports compact events upstream, and keeps working when the network doesn't.
Latency matters too. A bearing approaching catastrophic failure can't wait on a cloud round trip for the shutdown signal — the interlock decision belongs on the node. Cloud connectivity adds fleet dashboards and trend analytics on top, not instead of.
Installation and maintenance realities
Sensing quality is mechanical: a poorly coupled accelerometer (magnetic mount on painted steel, adhesive on a vibrating housing) corrupts the signature before the model sees it. Stud-mount or industrial adhesives, consistent placement across a fleet, and documented torque — the unglamorous details decide detection performance.
Power strategies split by site: line-powered nodes infer continuously; battery/harvesting nodes duty-cycle (wake, sample a few seconds, infer, sleep) and still catch slow-evolving faults. Bearing failures develop over weeks — a node checking every few minutes misses nothing.
Where it shows up
Motor & bearing health
Vibration sensors classify bearing wear, imbalance, and misalignment weeks before failure — flagging the exact fault frequency on the node.
Spindle & tool monitoring
Current and vibration signatures catch tool wear, chatter, and spindle anomalies on CNC machines during the cut.
Pump & compressor diagnostics
Cavitation, valve wear, and flow anomalies detected from fused pressure + vibration + current data at the machine.
Frequently Asked Questions
What sample rate do I need for vibration monitoring?
Enough to capture the fault frequencies of interest — bearing defect frequencies can reach a few kHz, so 5–20 kHz sampling is typical for bearing monitoring; imbalance and misalignment live under 10x rotation speed and need much less. Oversampling costs power and memory, so match the rate to the faults you're hunting.
How early can edge AI detect a bearing fault?
Characteristic bearing frequencies become visible in vibration data weeks to months before audible or thermal symptoms — the classic P-F curve. Early detection depends more on sensor coupling quality and measurement cadence than on model choice; a well-mounted accelerometer with modest ML beats a sophisticated model on a loose magnetic mount.
Do I need one model per machine type?
Not necessarily — anomaly detection baselines transfer reasonably within machine families, and normalization layers handle unit differences. The honest answer is machine-context-aware models: conditioned on machine type, speed, and load. Fixed-speed, fixed-load machines (HVAC motors, conveyor drives) are far easier than variable-speed equipment.
Can this run without internet on the floor?
That's the point of the edge architecture — nodes analyze locally and report events over whatever link exists (industrial Ethernet, gateway mesh, even store-and-forward). The monitoring system keeps working in air-gapped facilities; cloud connectivity is an optional analytics layer, not a dependency.
Building something that should run AI on-device?
Edgehound designs, compresses, and deploys TinyML models on microcontrollers — from feasibility audit to field-ready firmware.