Anomaly detection

Machine anomaly detection before it becomes a breakdown

A fixed limit only catches a reading once it is already too high. Nisthora learns what is normal for each machine, for what it is doing right now, and flags readings that do not fit, even when they are still inside the limits.

Normal range for this machine (learned) Unusual reading flagged Time →
Illustration. The band is learned from the machine's own recent history.
Machine health

Four checks, all learned from your own machines.

Each check starts once it has enough of your plant's own data. Until then, the screen says so, for example "Learning — 12 of 60 days the machine ran", instead of showing a guess.

CheckWhat it answersLearns from
Sensor healthIs this sensor telling the truth? Flags a sensor that has gone quiet, is stuck on one value, is out of range, or sends bad signals.3 days
Anomaly detectionIs a reading not normal for what the machine is doing? Normal is learned per sensor for running, stopped and each product.60 days the machine ran
Cycle-time driftIs the machine slowly getting slower? Compares the last 7 days with the 30 before, and says how many pieces it cost.14 days with production
Micro-stopsHow much do the short stops nobody writes down cost? Counted from the gaps between pieces.30 pieces
Machine Health page

What to work on, in plain words.

Every open warning in one list: which machine, what the readings show, since when, and whether someone has picked it up.

  • Maintenance due, with roughly how many days are left
  • Readings that are not normal, and machines that behave unlike their sister machines
  • Sensors that cannot be trusted, and micro-stops nobody recorded
Machine Health page listing open warnings such as a press needing maintenance in about 5 days
How it decides

Careful about false alarms, honest about its record.

Level and relation

A reading can be far from its usual level, or wrong for what the other sensors say: 140 A can be normal at 70 strokes a minute and a fault at 50.

No alert floods

An anomaly opens when most of the last few minutes are out of line, and closes after ten normal minutes. Two related sensors raise one anomaly, naming both.

A measured false-alarm rate

Limits are set to about one false alarm per machine per month of running, then tested on days they never saw. The measured rate is shown next to every result.

Your team teaches it

Mark each anomaly "real problem" or "normal running". Normal running is learned; a real fault never becomes normal.

Where you see it

Built into the screens you already use.

Machine Health page

Every machine of the plant, its checks and its open anomalies.

Health tab on each machine

The machine's sensors, anomalies, cycle time and micro-stops.

Alerts

Anomalies and sensor faults arrive on the Alerts page like any other alert.

Dashboards

Add an anomaly widget to any board, or start from the Anomaly Feed template.

See Nisthora on your own machines.

Book a demo. We will show you live data and talk through your machines, lines and shifts.