Setting a machine's baseline before calling anything an anomaly

Anomaly detection sells well in a demo and disappoints in a plant. The usual cause is that the system was switched on before anyone defined what normal looks like for that machine, so the first week is full of alarms about things that are fine.

Normal depends on the mode

A compressor at full load, at half load and idling draws very different current and vibrates differently. A single average and a single limit flags every load change. Record the operating mode alongside the signal: load setpoint, product being run, speed, ambient temperature, or whatever the machine controller exposes. Then build a baseline for each mode.

If the mode is not available from a controller, infer it. Motor current bands or spindle speed from a Modbus register often give a clean split.

How to collect a baseline

Run the machine in a known healthy state, ideally just after maintenance, and log through a full range of its normal work. A few days is the minimum for a machine with daily cycles. A machine with seasonal or weekly patterns needs longer. Two things to record:

• Raw or lightly processed values, such as RMS current, vibration RMS, temperature, with timestamps.
• Events, such as maintenance, product changes and known faults, so you can later separate real trends from process changes.

Check sensor quality first. A loose thermocouple or a noisy current clamp produces an anomaly signal that is a wiring fault.

From baseline to threshold

Start simple. For each mode, compute the mean and spread of each signal and flag values that sit beyond about three standard deviations for longer than a set time. A single spike is rarely a fault, a sustained shift often is. Add a rate-of-change check for signals such as bearing temperature, where the slope tells you more than the level.

Run in a quiet mode first: log what would have alarmed, and let the maintenance team judge each case. Tune until the flags they call useful clearly outnumber the ones they call noise, then switch on notifications. Only after that does a learned model add much.

Baselines age

Wear, a new supplier’s raw material or a firmware change on the drive all move the baseline. Re-baseline on a schedule and after major maintenance, and keep the old versions so you can see slow drift.

What this does not fix

A baseline tells you the machine differs from before. It does not tell you why, or whether the old behaviour was healthy. If the machine was already faulty when you recorded it, the system will treat the fault as normal.

---

If you want condition monitoring that your maintenance team trusts, book a discovery call. See industrial IoT and AI for devices and operations.