Anomaly detection
- Anomaly detection flags behavior that deviates from a learned model of normal — across many variables at once, not one threshold at a time.
- It catches multivariate drift: combinations of small in-limit changes that together predict scrap or failure.
- Alert quality is the make-or-break: flags must arrive with ranked contributing variables and respect operating regimes.
- It is the front end of a pipeline — detection, then attribution, then action — not a product category by itself.
Anomaly detection is the workhorse of industrial ML: a model learns the joint behavior of a line's variables during normal production and scores live data against that model. A high score means “this moment doesn't look like any good history,” — often long before any single variable crosses an alarm limit.
What it catches that thresholds miss
Single-variable alarms miss three failure shapes that dominate real losses. Combinations: five variables each drift 1% — individually fine, jointly a scrap signature. Context: 180 °C is normal for product A and disastrous for product B; a learned model knows the difference, a static limit doesn't. Slow drift: a three-week creep toward the edge of the envelope that no daily glance at trends will register. Process-signal drift is also increasingly used as an earlier failure precursor than vibration for some equipment classes.
Avoiding the alert cannon
The classic failure mode is a system that cries wolf until operators mute it. The defenses are attribution (every flag names its top contributing variables so an engineer can triage in seconds), regime awareness (changeovers, startups and product mixes are modeled, not flagged), and severity scoring tied to consequences the plant cares about — predicted scrap or quality impact, not statistical distance alone. Evaluate exactly this in a pilot: alerts per shift, share of alerts an engineer judged actionable, and time from flag to explanation.
FAQ
- Is anomaly detection the same as predictive maintenance?
- Predictive maintenance is one application of it. Anomaly detection on machine signals predicts failures; the same math on process signals predicts quality problems and drift. Platforms differ in which signals they specialize in.
- How much history does an anomaly model need?
- Enough normal production to cover the line's real operating regimes — typically months. More regimes (products, seasons, shifts) need more history; a line with stable products can train usefully on less.
Every number on this site — in one sourced spreadsheet.
The $43.6B market math, use-case shares, adoption barriers and published deployment results, with the source column intact. Ready to paste straight into your business case.