Predictive Maintenance with Vibration Sensors and Edge AI

Unplanned downtime usually starts as something small: a bearing that is beginning to wear, a fan that has picked up some imbalance, a coupling that has drifted out of alignment. Each of these changes the way a machine vibrates long before it fails.

Vibration-based predictive maintenance measures those changes and warns you early enough to plan a repair instead of reacting to a breakdown. With modern sensors and edge AI, it is no longer limited to large plants with dedicated reliability engineers.

Preventive vs predictive maintenance

Predictive maintenance doesn't replace good preventive practice. It tells you which machines need attention now and which can safely run longer.

What vibration tells you

Rotating equipment produces vibration at frequencies linked to its running speed. Different faults show up in characteristic ways:

• Imbalance typically raises vibration at once per revolution (1× running speed).
• Misalignment often shows up at 2× running speed, with high axial vibration.
• Mechanical looseness produces many harmonics of running speed.
• Bearing defects create vibration at specific bearing fault frequencies, which depend on the bearing's geometry, often at high frequencies and early in the failure.

International standards such as ISO 20816 give guidance on judging overall vibration severity for different classes of machine, which is a useful starting point for alarm levels.

Choosing sensors

MEMS accelerometers are small and inexpensive, and they work well for trending overall vibration on motors, fans and pumps.

Piezoelectric (IEPE) accelerometers are the industrial standard. They have wider frequency range and lower noise, which matters for early bearing fault detection and high-speed machines.

ImportantHow you mount the sensor matters as much as which sensor you buy. A stud mount on a clean, flat surface near the bearing gives the most reliable readings; a magnetic mount is convenient for surveys; adhesive mounts sit in between. Mount every sensor in the same place and direction so readings can be compared over time.

Sample at more than twice the highest frequency you care about (the Nyquist limit), and in practice well above that, to capture bearing-related frequencies.

From raw signal to useful features

Sending every raw sample to the cloud is expensive and rarely needed. Instead, the edge device computes features from short windows of data:

• RMS (root mean square): overall vibration energy, the most common trend value.
• Peak and crest factor: spikes relative to the average, useful for impacts.
• Kurtosis: how "spiky" the signal is, often sensitive to early bearing damage.
• Frequency spectrum (FFT): energy at 1×, 2× and bearing fault frequencies.

These features are a few numbers per minute instead of thousands of samples per second.

Where edge AI comes in

Simple thresholds on RMS catch large problems, but every machine is different and loads change through the day. Machine learning helps in two ways.

Anomaly detection: learning "normal"

Failure data is scarce; most machines rarely fail. So the practical starting point is anomaly detection. The model learns the normal pattern of features for each machine, across its normal operating states, and scores how unusual new readings are. Algorithms such as isolation forests or small autoencoders are common choices.

The model then raises a warning when the machine drifts from its own baseline, even if it is still under a generic alarm level.

Classification: naming the fault

Once you have labelled examples of real faults, a classifier can suggest what is wrong: imbalance, misalignment or bearing wear. This usually comes later in a program, once anomaly detection has been running for a while.

Running models on the edge

Small models run comfortably on an industrial gateway, and tiny ones can run on microcontrollers using frameworks such as TensorFlow Lite for Microcontrollers. Running at the edge means:

• far less data to send,
• alerts keep working if the network is down,
• faster response for critical machines.

Running a pilot

1. Pick three to five critical machines whose failure causes the most downtime.
2. Install sensors and collect a baseline over a few weeks of normal operation, across all products and loads.
3. Start with trends and simple alarms on RMS and temperature so the maintenance team trusts the data.
4. Add anomaly detection per machine and tune it with feedback from technicians.
5. Close the loop. Every alert should lead to an inspection, and every inspection result should be recorded so the system improves.

NoteThe goal of a pilot is not a perfect model. It is one planned repair that would otherwise have been a breakdown. That single result builds the case for scaling up.

Getting help

CrimsonEdge Systems builds the full chain: sensor selection and mounting, edge gateways, feature extraction, anomaly detection models and dashboards. Read how we retrofit legacy machines with IoT sensors, explore our AI solutions, or contact us to plan a pilot.