Telemetry Anomaly Lab

Synthetic sensor stream, real detectors. Three channels of invented telemetry flow through three online anomaly detectors implemented in this page's JavaScript. Use the buttons to break the sensors and watch each detector respond. Score is kept honestly: detection latency and false alarms both count.

Inject fault: Speed

What is running

Adaptive z-score keeps a rolling mean and variance per channel and flags points beyond 4σ — fast on spikes, blind to slow drift. EWMA tracks an exponentially weighted mean and flags sustained departure — catches drift the z-score misses. CUSUM accumulates small deviations and alarms when the cumulative sum crosses a decision boundary — the classical change-point detector, tuned here with k=0.5σ, h=5σ. A stuck-at fault is caught by a variance-collapse check, which neither of the others sees. No single detector wins on every fault type; the engineering is in running them together and triaging their disagreement. That triage pattern is the transferable piece.

Why this matters, in plain terms: machines and systems rarely announce that they are failing; they drift, stick, or blip in ways a human watching a dashboard will miss at 3 a.m. Software like this is the night-shift watchman that never blinks: it learns what normal looks like and raises a hand the moment something stops being normal, whether that is a factory sensor, a server fleet, or a financial feed. Catching the problem an hour earlier is often the difference between a maintenance ticket and a very bad day.