Predictive Maintenance: Start with the Data You Already Have

Predictive maintenance projects fail by starting with platforms instead of failure modes. Reverse it: pick the five failure modes that cost you the most downtime, then check — the data that predicts them is usually already on the network.

Data hiding in plain sight

• Variable frequency drives report motor current, torque, thermal state and energy per shift. A rising current trend at constant load is a mechanical problem announcing itself — a bearing, a blocked filter, a worn impeller. Every ACS580 or PowerFlex exposes these over Modbus or Ethernet.
• Smart sensors with IO-Link — the SICK W16, for example — expose contamination level, operating hours and temperature, so a dirty lens is replaced on a planned stop instead of stopping the line.
• PLCs already count cycles, travel times and reject rates. A pneumatic cylinder that takes 15 % longer than its baseline is wearing out; the PLC knows before the operator does.
• HART instruments carry diagnostics most plants never read: sensor drift, loop current deviations, electronics temperature.

The 90-day pilot that works

1. Pick one line with a known pain point.
2. Historize 10–30 signals — an edge gateway such as the Moxa UC-2112 or the existing SCADA historian is enough.
3. Establish two weeks of baseline per operating mode.
4. Set simple statistical thresholds: alert when a signal exceeds baseline + 3σ for N minutes. No machine learning yet — threshold-on-trend catches a large share of developing failures.
5. Review weekly with maintenance; tune the thresholds.

Add sensors only where data is missing

Vibration on critical bearings (velocity RMS per ISO 10816 first, spectra later), temperature on gearboxes, pressure differential across filters. Each added sensor should map to a named failure mode. "We might see something interesting" is how projects drown in data.

Close the loop or it dies

An alert must create a work order with a human decision recorded: real or false. That feedback tunes thresholds and, later, justifies real models. Dashboards nobody acts on are the tombstones of IIoT pilots.

What success looks like after 90 days

A short list of signals that predicted at least two real problems, a maintenance team that trusts the alerts, and a business case for the next line. From there, scale the architecture — not the ambition — one line at a time.

We deliver starter packages — edge gateway, sensor kit, historian configuration and threshold design — sized for one line and expandable when it proves itself. Our Vietnamese page on predictive maintenance for pumps and rotating equipment shows a worked example. Contact our engineers.

Quick answers

Do we need vibration sensors to start?

Not usually. Motor current and torque from the VFD, cycle times from the PLC and dirt-level data from IO-Link sensors catch a large share of developing failures. Add vibration sensors on critical bearings in the second phase, mapped to named failure modes.