Sentinel

Name: Sentinel: half-minute tour
Uploaded: 2026-10-11
Duration: 39 s
Description: Find unusual behaviour in a sensor table and see which sensors drove it. Isolation Forest, PCA (Hotelling T² and SPE) and EWMA, in your browser with C/WebAssembly; nothing is uploaded.

Load your sensor table. Sentinel learns normal operation, flags unusual periods and shows which sensors carry the deviation. Your data stays in this browser.

Try it now ↓Watch the half-minute tour0:39

Sentinel in half a minute

A short narrated tour: how normal operation is learned, how to read an alert and where to start.

Read the transcript

A machine has dozens of sensors. A failing seal rarely breaks one limit. It quietly changes how the sensors move together. Sentinel learns normal operation from the first part of your data. Then it scores every row with three methods, and raises an alert only when the deviation persists. For each alert you see which sensors carried the deviation, so you know where to look first. Load a CSV, or try a built-in scenario. Everything runs in your browser.

Narration: synthetic voice (Piper, voice “Joe”, CC0 public domain). Captions are burned into the video.

In three steps.

Learns normal

Sentinel studies the first part you call “normal”: how the sensors move together.

Scores every row

Three methods measure how far each row sits from normal. A combined deviation index above 1 means a limit was crossed.

Shows why

For each unusual period you see which sensors contributed most. The decision stays with you.

Data

Pick a scenario. All are mathematically generated examples, not real plant data.

My own sensor table

Drop a CSV file here

Numeric columns are sensors; a date column becomes the time axis.

Settings

60%

Sensitivity ·

PCA · +

Timeline: deviation index

When the line crosses 1 a limit is exceeded. Hover to read; drag to zoom.

Unusual periods

Pick one to see the reasons below.

Which sensors drove it?

Bars show how the deviation is shared between sensors. Small charts compare each sensor with its own normal band.

Normal-region map: T² and SPE

Each point is a row. Horizontal: size within the normal relationships; vertical: how far those relationships broke; both relative to their own limit (1). Outside the box means a limit is crossed. Faint points are the training part.

Export

How do I read the result?

1Timeline. The line is the . Below 1 is normal, above is a crossed limit. Shaded areas are alert periods.
2Colour strip. Each row is one cell; the darker, the further from normal. The black line separates training from test.
3Sensor shares. Taken from the and decomposition, they show which sensors carry the deviation.
4Map. Inside the box is normal. Points that move right are unusually large values; points that move up mean the relationships between sensors broke.

Sentinel says “unusual”, never “faulty”. A period may be maintenance, a product change or a real problem; the engineer decides. If the training part contains a fault, the model treats it as normal.

Frequently asked

What kind of anomaly does Sentinel find?

Rows and periods where several sensors together depart from the learned normal operation. It also catches cases where each sensor is within its own limits but the combination is unusual.

Is “unusual” the same as “faulty”?

No. Sentinel only says a row departs from the normal it learned. A commissioning step, a product change or maintenance also look unusual. The engineer decides the cause; the per-sensor shares for each period support that review.

Is my data uploaded?

No. The table is processed in your browser by a WebAssembly module written in C. Nothing is sent anywhere and it is gone when you close the page. Files you export stay with you.

What if the training part contains faulty data?

The model then treats that behaviour as normal and may miss similar events later. Check that the training part really represents normal operation; Sentinel also shows the alarm rate inside the training part.

How many rows and sensors are needed?

In practice at least a few hundred normal rows and two to twenty-four numeric sensors. Sentinel reads up to 20,000 rows; model training uses up to 4,096 rows, selected at even spacing.