Anomaly Detection on Diagnostics
Flagging diagnostic signals that depart from normal behavior, often without labeled examples of every fault.
The task
A fusion device carries many diagnostics, and any can drift, saturate, or fail. Anomaly detection watches these streams and flags samples that do not look like normal operation, catching sensor faults, off-normal plasma behavior, or precursors to events, often before a labeled classifier could.
Why unsupervised
There are far too few examples of each specific fault to train a supervised classifier for all of them. Anomaly detection instead learns what normal looks like and measures deviation, so it can flag faults it was never shown, at the cost of not naming them.
- Reconstruction-based: autoencoders that fail to reconstruct anomalies
- Density-based: flag low-probability samples under a learned model
- Distance-based: flag points far from normal clusters
- Forecast-based: flag large prediction residuals
Thresholds and drift
An anomaly score must be thresholded to raise an alert. Set it too low and normal variation triggers alarms; too high and real faults are missed. Because normal operation itself drifts over a campaign, thresholds and the normal model may need periodic updating.
From detection to action
A flag is a request for attention, not a diagnosis. Useful systems route anomalies to the right context: is it a sensor fault, a genuine plasma event, or a data-pipeline glitch? Pairing detection with the raw signal and its recent history lets an expert or a downstream classifier resolve it.
Evaluation
Because true anomalies are rare and sometimes unlabeled, evaluation uses injected faults, held-out known events, and expert review. Report the false-alarm rate at a useful detection rate, and be explicit about which fault types were tested.