Consider a vibration sensor mounted on a factory conveyor motor, streaming two readings continuously: vibration amplitude and temperature deviation from baseline. On a healthy day, nearly every reading clusters tightly around normal operating values. Occasionally, a genuine mechanical fault, a bearing wearing out, a belt beginning to slip, produces readings that sit well outside that normal cluster, sometimes hours or days before the fault becomes severe enough to cause an unplanned shutdown.
Scoring isolation, not fixed limits
Rather than setting a fixed acceptable range for each sensor independently, an isolation-style detector scores each reading by how far it sits from the dense cluster of normal points in the combined feature space, typically measured as average distance to its nearest neighbours. Readings deep inside the normal cluster score low; readings out on their own score high, regardless of which specific combination of vibration and temperature produced that isolation.
The threshold decision is the whole ballgame
Once a detector produces a score, the alert threshold determines everything that follows operationally. Set it too low, and maintenance staff get paged for ordinary operating noise, and eventually start ignoring the alerts entirely, a well-documented failure mode called alert fatigue. Set it too high, and a genuine developing fault can run for days before crossing the threshold, closing the window for a cheap scheduled repair instead of an expensive emergency one.
What this looks like in practice
Real deployments typically start with a conservative threshold, tuned against historical fault data where available, then adjust based on observed false-alarm rates and any faults that were caught too late. The goal is not a perfect detector, but a threshold that keeps the cost of missed faults and the cost of alert fatigue both within what the maintenance team can actually absorb.
Try it yourself
The AI Anomaly Detection Lab lets you adjust the alert threshold against a simulated dataset with known true faults, and see directly how many faults get caught, how many false alarms fire, and how many faults slip through.
🧪 Try it yourself: the AI Anomaly Detection Lab simulation lets you experiment with everything described above directly in your browser.