← Back to the chapter · Next detail →

Technical detail · Time-Series Anomalies: Baselines and Residuals

Window completeness and response time

For a five-minute window sampled every 30 seconds, expect 10 readings. If only eight arrive before the one-minute watermark expires, choose explicitly: score a provisional residual, wait and violate the latency budget, or mark the window incomplete and route a sensor-health review. Store that policy in the incident record because it changes both quality and response time.

Without a watermark or late-data policy, a partial window can look abnormal and become normal after delayed readings arrive. Score only complete windows or mark low-confidence early alerts as provisional.

Carrying the last value forward through six missing samples may hide a pump shutdown; dropping the window may hide a short overload. Record the gap count, imputation method, clock source, model version, and baseline ID beside the residual. This lets reviewers distinguish a contextual anomaly from a timing or missing-evidence artifact.

A 30-minute STL window cannot support a two-minute safety response unless an edge rule raises a provisional alert first. A one-sample edge residual can be fast but weak. Attach the gateway or cloud’s later context review to the same incident ID so it does not become a disconnected second alert.

← Back to the chapter · Next detail →