PYTHON · NUMPY

Drift has a signature

Separate instrument drift from real change in the field using the spatial and temporal structure each one has.

01Challenge

One node in a mesh of twelve has been reading steadily higher for three months. Decide whether that node is drifting or whether that corner of the field is genuinely drying, using only the data the mesh already produces.

02Model

A drifting sensor and a changing field look identical if you only read one node. They stop looking identical the moment you read the mesh, because they have different shapes. A real change in the field is a physical event with a physical extent. It arrives at neighbouring nodes too, attenuated by distance and delayed by transport. Instrument drift has no extent at all. It lives in one device and stops at its enclosure, so it appears in exactly one series and in none of its neighbours. That difference is a test, and it needs no reference instrument and no field visit. Ask whether the anomaly is spatially coherent. A field event is coherent, a drift is not. The test costs nothing because the mesh is already reporting.

03Guided practice

For each node compute how far its recent change sits from the change seen by its nearest neighbours. A field event gives every node in the neighbourhood a similar change and a small residual. A drifting node gives a large residual because its neighbours did not move.

04Feedback

PASS WHEN You separated an instrument fault from a field event without a reference instrument and without leaving the office. The mesh already carried the answer.

If your run fails, check:

residual should compare node i's change to the average change of NEIGHBOURS[i], not to the whole mesh.
Both worlds put the same 0.030 anomaly on node 4. If your residuals are equal, the neighbour average is not being used.
A field event lifts the neighbours too, so its residual is small. That is the signal you are looking for.

05Retrieve & space

A field event is spatially coherent. Instrument drift stops at the enclosure.
Comparing a node to its neighbours separates the two without a reference instrument.
The test costs nothing because the mesh is already reporting.

06Mastery & project

Explain why this test gets weaker as node spacing grows, and say what quantity decides when it stops working.

Take a time series from any multi-node system and run the coherence test on its largest anomaly. Report the residual and say what you would do next if it came out ambiguous.

GO FURTHER · THE LIVE PLAYGROUND

Where the spacing question comes from ↗ · The neighbour test needs neighbours inside the correlation length. This instrument shows what happens below the sampling spacing.

← Agreement is not accuracy What recalibration is worth →

PYTHON · NUMPY

Drift has a signature

Twelve nodes, one anomaly. Decide from spatial coherence alone whether it is drift.

You read

the arrays and values already in scope

You change

the code you write in each cell

Fixed

the dataset and the checks that grade you