How can I detect sensor drift in an industrial process?

What drift is and how it appears

Drift is a systematic change in a sensor's reading over time, while the input stays the same. It differs from noise because its direction is consistent and its rate is slow. A single reading rarely reveals it, since the effect can sit inside normal process variation. For that reason, drift is usually found by comparing against a reference, or by looking at a trend over a long time window.

Drift is a property of the sensor. Process drift means the process itself has changed. The two can appear on the same chart, yet their causes and the actions they need are different. Common causes of sensor drift include calibration shift, fouling, mechanical wear and temperature-dependent behaviour.

Example: two temperature sensors, 0.2 °C per month

Suppose two temperature sensors sit on the same pipe (hypothetical). At the start, the difference between them is 0 °C. Assume sensor A drifts by 0.2 °C per month relative to sensor B. After six months the difference is 6 × 0.2 = 1.2 °C. If a tolerance band of 0.5 °C is chosen, the difference exceeds it in the third month, when 3 × 0.2 = 0.6 °C. The difference alone does not say which sensor moved, so a third reference or a check at a known temperature is needed. To catch a small, sustained drift earlier, methods such as EWMA or CUSUM can signal sooner than a single threshold crossing.

Wrong conclusions and limits

The most important mistake is treating drift as process change, or the reverse. If the process really changed, correcting the sensor hides that change. The second mistake is retraining an ML model on drifting data: the model learns the new level as normal, so the anomaly never appears. The third is relying on statistics without a reference check. A single sensor can show a slow drift as normal when judged against its own past. Redundant sensors can also drift together if a common cause, such as a shared heat source, affects both.

Follow-up questions