Anomaly Detection with Vibration Sensors — A Maker's Introduction

Predictive maintenance — detecting equipment faults before they cause failure — is one of the most commercially valuable AIoT use cases. Industrial companies spend significant resources deploying vibration sensors and ML models on motors, pumps, fans, and conveyors to catch abnormal vibration patterns that precede failures.

The underlying approach is accessible to makers. You don't need an industrial sensor network to understand how vibration anomaly detection works and to build a working prototype. A ₹200 ADXL345 accelerometer, an ESP32, and a few hours with Edge Impulse is enough to build a system that can distinguish "motor running normally" from "motor with a loose bearing" or "motor with a worn belt."

This post covers the approach — not as a production engineering guide but as a starting point for understanding what's involved.

What normal vibration looks like

Every rotating machine has a characteristic vibration signature: frequencies and amplitudes that are consistent when the machine is running correctly. This signature is determined by the rotation speed, bearing geometry, blade/impeller count, and mounting characteristics.

The signature isn't a single frequency. It's a spectrum: the fundamental rotation frequency, harmonics at integer multiples, and sometimes sub-harmonics. Specific bearing defect frequencies are predictable from bearing geometry. Belt frequency is predictable from belt length and pulley diameter.

"Anomaly" means the measured spectrum deviates from this learned normal signature. A worn bearing produces energy at specific defect frequencies. An unbalanced rotor produces increased amplitude at the rotation frequency. A loose component produces broadband noise or impact impulses.

The ML approach for makers

For a maker project, the most accessible approach is a classification model (normal vs. X fault conditions) rather than full spectral analysis.

Data collection: mount an ADXL345 or MPU-6050 on the machine. Sample at high frequency (100–400Hz, limited by your ADC and microcontroller). Collect data in several conditions: normal operation, simulated fault condition (slightly unbalance a rotating component, loosen a mount bolt), and powered-off noise floor.

Feature extraction in Edge Impulse: use the spectral analysis block, which computes FFT features automatically. This converts time-domain vibration data into frequency-domain features that the model can learn from.

Training: a simple neural network trained on these features can classify normal vs. anomalous operation with reasonable accuracy if the fault conditions are distinct enough.

Deployment: once trained, the model runs on the ESP32 with inference every few seconds. When anomalous vibration is detected, trigger an alert.

Practical limitations

The model only knows about conditions it was trained on. It will detect deviations from normal — but "anomalous" might mean "the fan is dusty" rather than "the bearing is failing." You need to understand what anomaly means in your specific context and design your data collection to cover the cases you care about.

Single-axis vs. three-axis: a 3-axis accelerometer gives you more information. Many faults produce off-axis vibration components that aren't visible on the primary axis. If your budget and compute allow, use three-axis features.

Mounting consistency: the sensor must be mounted consistently relative to the machine. If the sensor position changes between training and deployment, the vibration signature changes and the model's accuracy degrades.