Anomaly detection is the data-analysis technique that identifies unusual behaviour in time series or industrial patterns (a changing vibration, a consumption spike, a lengthening cycle) without anyone having previously labelled what is “normal” and what is not. The output is a signal, not a diagnosis: it says something changed, and the cause still has to be established.
How anomaly detection works
Starting from process history (temperatures, vibrations, consumption, cycle times), a model of normal behaviour is built, from statistical thresholds and process control to models that learn relationships between variables, up to deep learning for complex multivariate patterns. When the current observation departs significantly from what the model expects, an alert fires. The practical key is tuning: too sensitive produces false alarms the team learns to ignore; too lax, warnings that arrive late.
Why it matters for industrial SMEs
For an SME it is often the entry point to industrial AI, because it does not require a labelled failure history, which almost no plant has. Typical uses: anticipating equipment degradation, catching in-process quality drift, spotting anomalous energy consumption that reveals leaks or off-hours equipment, and monitoring data quality itself. The prerequisite is continuous, reliable data capture.
Related terms
Anomaly detection underpins predictive maintenance and remaining useful life. It feeds on data captured via OT/IT integration, often runs at the edge, can trigger an industrial AI agent, and the same techniques monitor models against drift.