Predictive Maintenance with Edge AI

The Cost of Unplanned Downtime

A failed bearing in a production motor does not just cost the bearing. It costs the production line stoppage, the emergency repair crew, the scrapped in-process material, and the missed delivery deadline. Industry estimates put the average cost of unplanned downtime at $5,000-$50,000 per hour for mid-size manufacturing operations.

Traditional maintenance follows two patterns:

Reactive maintenance. Run the machine until it breaks, then fix it. Cheapest in labor — catastrophic in downtime costs. A typical bearing failure on an unmonitored motor goes from first detectable vibration anomaly to catastrophic failure in 2-8 weeks.

Scheduled maintenance. Replace parts on a fixed calendar (every 6 months, every 10,000 hours). Safer, but wasteful — you replace parts that still have months of useful life, and you still miss failures that happen between intervals.

Predictive maintenance. Monitor the actual condition of the equipment and intervene when degradation is detected. This is where edge AI fits.

Why Edge, Not Cloud

Predictive maintenance is one of the strongest use cases for edge AI over cloud AI. The reasons are specific to factory environments:

Data volume. Vibration data at 3.2 kHz sampling generates approximately 70 MB per hour per 3-axis sensor (int16 format). Streaming this to the cloud is impractical. Processing on-device reduces the output to kilobytes of anomaly events.

Connectivity. Factory floors are RF-hostile environments. Metal enclosures, electromagnetic interference from motors, and thick concrete walls make Wi-Fi unreliable. An edge device works offline after deployment.

Data sensitivity. Vibration patterns can reveal production volumes, process parameters, and equipment configurations. Many manufacturers will not send this data to external servers.

Scale economics. Cloud API costs scale linearly with inference frequency — at production rates across hundreds of sensors, costs can reach six figures annually. Edge devices have zero per-inference cost after the hardware investment.

Sensing Modalities

Vibration Analysis (Most Common)

Vibration is the primary signal for rotating machinery. Bearing wear, shaft misalignment, gear damage, and imbalance produce characteristic vibration patterns detectable before human-noticeable symptoms appear.

Hardware: MEMS accelerometer (3-axis, minimum 1 kHz sample rate) mounted on the machine housing near the bearing. Rigid coupling is essential — magnetic mounts work for testing, epoxy or stud mounts for permanent installation.

ML approach: Capture FFT spectra of vibration data. Train an anomaly detection model on spectra from healthy operation. The model flags spectral patterns that deviate from the learned baseline.

Example: ESP32 with Edge Impulse reads a 3-axis accelerometer, computes spectral features, and runs inference in under 20 ms.

Temperature Monitoring

Temperature rise in motors, bearings, and electrical panels indicates increased friction, overload, or insulation breakdown. Temperature changes are slower than vibration signals — useful as a secondary indicator or confirmation channel.

Hardware: Thermistor, thermocouple, or IR temperature sensor. Sample rate of 1-10 Hz is sufficient.

ML approach: Track temperature gradients over time. A sudden rise relative to the ambient baseline — not just a high absolute value — is the signal.

Current Monitoring

Motor Current Signature Analysis (MCSA) detects faults without physical contact with the motor. Broken rotor bars, stator faults, and bearing defects modulate the motor current in detectable patterns.

Hardware: Current transformer (CT) on the motor supply cable. Non-invasive installation — no need to touch the motor.

ML approach: Analyze the frequency spectrum of the current signal. Specific fault types produce sidebands at characteristic frequencies around the supply frequency (50/60 Hz).

ML Approaches

Anomaly Detection (Start Here)

Anomaly detection learns “normal” and flags deviations. This is the right starting point because:

• You only need data from healthy operation — no failure data required
• Works across different failure modes without pre-defining them
• Models are small (5-30 KB) and fast (1-10 ms inference)

The trade-off: anomaly detection tells you “something is different,” not “the bearing will fail in 3 weeks.” It is a screening tool, not a diagnostic tool.

Model types: Autoencoders, statistical models (Mahalanobis distance), or classifiers trained on spectral features of normal operation. The STM32F4 with TFLite runs an autoencoder anomaly detector in under 10 ms.

Failure Classification (Advanced)

If you have historical failure data — labeled examples of different fault types — you can train a classifier that identifies the specific failure mode: bearing wear, misalignment, imbalance, looseness.

This requires:

• Labeled failure data (from accelerated life tests or historical maintenance logs)
• Separate model per equipment type
• More training data (50-200 samples per failure class)

Most teams start with anomaly detection and add classification later, after collecting enough anomaly events with confirmed root causes.

Hardware Selection

Priority	Recommended MCU	Why
Wireless alerts	ESP32	Wi-Fi/BLE for MQTT alerts, $5-15
Battery-powered	STM32L4	< 100 nA sleep, years on battery
Maximum accuracy	STM32H7	480 MHz, 1 MB SRAM for complex models
Budget fleet	ESP32-C3	$1-3 per chip, Wi-Fi included
Industrial reliability	STM32F4	Industrial temp range, long lifecycle

For a first deployment, the ESP32 with Edge Impulse is the fastest path. Built-in Wi-Fi means anomaly alerts reach a dashboard without additional hardware.

Use the MCU Selector to compare options by connectivity, power budget, and cost for your specific deployment.

ROI Calculation

A realistic ROI model for a 10-motor pilot deployment:

Cost Item	Amount
10x MCU nodes + sensors + enclosures	$500-1,500
Edge Impulse (free tier for development)	$0
Installation and commissioning (8 hours)	$400-800
Dashboard setup (Grafana + MQTT broker)	$200-500
Total pilot cost	$1,100-2,800

Benefit Item	Annual Value
1 prevented unplanned downtime event	$5,000-50,000
Extended maintenance intervals (30% longer)	$2,000-10,000
Reduced spare parts inventory	$1,000-5,000
Total annual benefit	$8,000-65,000

Payback period: 1-4 months for a typical mid-size manufacturer.

These ranges are wide because they depend on equipment criticality. A $200K CNC machine has different downtime costs than a $5K pump motor. The ROI argument is strongest for critical-path equipment where a single failure stops the production line.

Limitations

Edge AI is a screening tool, not a diagnostic tool. It tells you “this motor’s vibration pattern changed.” It does not tell you “the inner race bearing will fail on Thursday.” A maintenance technician still diagnoses the root cause — edge AI tells them where to look.

Sensor placement is make-or-break. An accelerometer loosely attached with tape produces unreliable data regardless of model quality. Industrial deployments need rigidly mounted, properly rated sensors.

Not all failures are detectable. Sudden failures (electrical shorts, seal blowouts) produce no gradual degradation signal. Predictive maintenance works for progressive mechanical wear — bearings, gears, belts, and alignment issues.

Training data quality matters more than model complexity. A simple anomaly detector with clean, representative training data outperforms a complex model trained on noisy data. Invest time in sensor placement and data collection, not model architecture.

Getting Started

1. Pick one critical motor — the one that causes the most pain when it fails
2. Mount a sensor — MEMS accelerometer, rigidly attached near the bearing
3. Collect baseline data — 1-2 weeks of normal operation across different loads and shifts
4. Train an anomaly detector — Edge Impulse is the fastest path; see the ESP32 predictive maintenance guide
5. Deploy and monitor — set conservative thresholds initially, tighten as you learn the signal
6. Validate with maintenance — track hit rate: how often does a flagged anomaly match a real issue?
7. Scale — after validating on one motor, replicate to the next 10, then 100