Sentinel: anomalies by shape

Four different assets, each scored the same way: every day-long window is compared with the nearest window from the asset's own normal operation, and the difference is scaled by how much normal days differ from each other. That puts a transformer, a pump, a flow meter and a wind turbine on one score axis with one alert threshold. The aim is fewer false alerts: no model or limit to tune per asset, and every alert comes with the normal behaviour it was compared against.

SentinelIllustrative, synthetic data

The fleet, in each asset's own units

Four asset signals over three weeks, each in its own units, with its alarm limit, shape alert episodes and the time a fault was injected.

One score axis: distance from the nearest normal shape

The shape score of all four assets on one axis, with a single shared alert threshold that can be dragged.
signalnormal library (first 7 days)shape alertshared thresholdlimitlimit alarmfault injected

The window and its nearest normal

The selected day-long window beside the nearest window from normal operation, with the difference shaded.
selected windownearest normal windowdifference

False alerts and faults caught

AssetFaultShape alertLimit alarm

Click an alert, a row or anywhere on a chart to see that window beside the normal behaviour it was compared with. Each limit was set for its own asset; the shape threshold is the same number for all four. Times are days from the start.