Residual score for sensor anomalies

Steps

Screens captured against Python 3 in your browser (JupyterLite) Python 3.12.7 / Pyodide 0.27.6 in JupyterLite 0.6.4; Chromium real browser capture, 2026-10-09 on 2026-10-09; the tool may have moved on — the text steps are the contract.

1. 1 Step 1

   Do

   Run the Step 1 notebook cell and inspect the drifting synthetic series with its planted outlier.

   You will see

   SYNTHETIC drifting temperature, seed=1414; Normal readings use fixed offsets; one planted +2.00 C outlier at t=7.; t value_C; 0 20.00; 1 20.30

   Why it matters

   The known planted event permits a direct check of detector behavior.

2. 2 Step 2

   Do

   Run the Step 2 notebook cell to inspect the trailing median baseline.

   You will see

   Trailing 3-reading median baseline; current point excluded; t value_C baseline_C residual_C flag; 8 21.70 21.20 +0.50 -; The baseline rises with the underlying drift.

   Why it matters

   A trailing baseline uses past readings and can follow slow drift.

3. 3 Step 3

   Do

   Run the Step 3 notebook cell to subtract the baseline and read residuals.

   You will see

   Residual = observed - trailing median baseline; t value_C baseline_C residual_C flag; 10 21.90 21.80 +0.10 -; At t=7 the residual is much larger than nearby values.

   Why it matters

   Residuals expose departures from the local expected value.

4. 4 Step 4

   Do

   Run the Step 4 notebook cell to apply the fixed 0.80 C threshold and note the flag.

   You will see

   Fixed alert rule: |residual| > 0.80 C; t value_C baseline_C residual_C flag; flagged timestamps=[7]; The planted event is t=7; post-event baseline may also be affected.

   Why it matters

   A fixed threshold converts scores into inspectable decisions.

5. 5 Step 5

   Do

   Run the Step 5 notebook cell to raise ordinary noise without changing the threshold.

   You will see

   Higher ordinary noise: random choices from -1.10, 0, +1.10 C; seed=1414; same planted t=7 value and same 0.80 C threshold; 9 21.80 22.30 -0.50 -; flagged timestamps=[4, 6, 7, 8]

   Why it matters

   Noise can produce ordinary false alerts under the same rule.

6. 6 Step 6

   Do

   Run the Step 6 notebook cell and compare flag counts and false positives.

   You will see

   RESULT CARD: synthetic rolling-median residuals; seed=1414; baseline threshold=0.80 C; More noise changes the alert count under the unchanged threshold.; This synthetic run does not establish field false-alarm rates.

   Why it matters

   A detector threshold needs evidence from the operating noise level.