Unusual, Bad and Actionable: Anomaly Detection from Control Charts to Isolation Forests

Deep End Era. Engineering with Sankalp · Oct 28, 2026 · 1h

Monday morning at the Kestrel compressor station: 412 anomaly alerts from one weekend, and buried among them a bearing on compressor four that fails on Tuesday. The system had worked. It had flagged something unusual, which is not the same as something bad and actionable. Part seven of The Deep End's Machine Learning series follows anomaly detection from nineteenth-century astronomers deciding whether to discard an observation, through Walter Shewhart's 1924 memo and Mahalanobis's tilted ellipses, to isolation forests that count how many random questions it takes to get a point alone…