Live status · Read the latest
monad-knowledge Wi-Fi sensing lab · FIIT STU
superseded EXP-002

CSI Model Drift Measurement Over Time

In plain words

If you teach a Wi-Fi people counter once and then leave it alone, how quickly does it start giving the wrong number?

bookshelf moveddoor proppedheating on counting error thirty days, same room, nobody retrains
A model frozen on day zero is checked every day. The line is the error the design expects to climb; the dashed ticks are the days the room changed, whose size is the second thing the experiment records. Schematic of the design, not a measurement.
Why it matters
A wrong count looks exactly like a right one. Furniture moves, doors close, the heating comes on, and the model keeps answering with confidence.
How it is done
Freeze a trained model in a real room for 30 days, check its count every day against a Bluetooth reference, and write down what changed in the room. Two curves come out: error against days, and the jump in error against the size of each change.
Where it stands
Superseded historical design. Current work and evidence limits are recorded under EXP-F2; this card has no primary data.

Asks, formallyCSI-based crowd counting accuracy degrades significantly over 7-30 days without recalibration, driven by environmental changes (furniture, temperature, door state), with degradation rate correlated to the magnitude of environmental change.

FoundSuperseded design retained for provenance, reviewed September 15, 2026.

StandingReplaced by later work

RoomFIIT STU — meeting room, measured over 30 days

CSI Model Drift Measurement Over Time

Current assessment (2026-09-15)

Superseded design retained for provenance, reviewed September 15, 2026. No primary data are attributed to this card. The current drift series has data, but a calibrated count-error timeline needs independent labels and chronological scoring. An empty-room signal record alone is not a measured accuracy-decay curve.

The current drift series has data, but a calibrated count-error timeline needs independent labels and chronological scoring. An empty-room signal record alone is not a measured accuracy-decay curve.

Next evidence needed: Use EXP-F2 for the current protocol and evidence. The earlier plan below is not an instruction to reopen this experiment.

Basis: Thesis Narrative, the September 10 synthesis, and the September 15 register review. See the review record for evidence sources and outstanding checks.

Earlier protocol and dated results

The methods and observations below retain their original scope. Read their readiness statements and proposed next steps as historical. The current assessment above records what has changed; frozen thresholds have not been retrospectively revised.

The problem, in plain words

Think of calibrating a bathroom scale while it sits on a thick carpet. Later someone swaps the carpet for tile. Now the scale reads wrong — your weight never changed, its surroundings did. Wi-Fi crowd counting has exactly this disease. The model learns "this pattern of signal strength across the Wi-Fi subcarriers ≈ 5 people," but that pattern depends on where the walls, chairs, tables, and metal surfaces sit, because the radio waves bounce off all of them. Rearrange the room and the mapping quietly breaks, with no error message and no warning.

Machine-learning people have a name for this — concept drift, the input-to-output relationship changing over time — and a whole research field built around detecting and adapting to it (gama2014_8f2e ↗). The uncomfortable fact for radio sensing is that the drift is not gradual noise; it can be a step change the instant a table moves. The reason this thesis cares is practical: if you cannot say how fast a CSI counter decays and what drives the decay, you cannot argue for the fix (periodic BLE recalibration) that the rest of the experiment series proposes. This card produces the problem statement; the calibration cards produce the cure.

What we are trying to prove

  • Hypothesis (falsifiable): a frozen CSI crowd-counting model loses meaningful accuracy — a >10% absolute drop from Day-0 — within 7–30 days of deployment, and the size of each error jump correlates with the magnitude of the environmental change that preceded it (furniture moved, object added near a link). Both halves must hold.
  • What a null result would mean: if the model stayed within a few percent of Day-0 for a month with the room in normal use, the thesis' central premise weakens — "recalibrate periodically" would be solving a problem that barely exists at this room's timescale, and the calibration cadence question (EXP-F2) would need re-motivating around a slower or rarer failure mode.
  • The subtle failure to guard against: finding drift but being unable to attribute it. If accuracy drops but shows no relationship to the logged environmental changes, then either the log is too coarse or the drift is driven by something unmeasured (hardware thermal effects, Wi-Fi channel reassignment, firmware). A drift curve without an attributable cause is a weaker thesis figure than one with a clear "chair moved → error jumped" panel.

How the experiment works (plain method)

  1. Day 0 — train and freeze. Collect ~2 hours of CSI with a co-located BLE ground-truth count (occupancy stepped 0–10 people). Train three models of deliberately different appetite — a lightweight Random Forest, an LSTM on amplitude sequences, and a CNN on subcarrier×time "images" — record Day-0 accuracy on a held-out set, then freeze all three. No retraining for 30 days is the entire point.
  2. Days 1–30 — watch it rot. Run continuous CSI capture during working hours with BLE ground truth in parallel. Every day, score the frozen models on that day's data: mean absolute error of the count, occupancy-band accuracy (0 / 1–2 / 3–5 / 6–10), and per-band F1.
  3. Log the room. Daily photographs of furniture, door/window state, count of extra electronics, temperature and humidity, and any maintenance event. This log is what lets a drop be attributed rather than merely observed.
  4. Optional controlled nudges — if the room is too stable to produce natural drift: move a chair (Day 7), swap table/chair positions (Day 14), place a large object near a TX–RX link (Day 21), restore the original layout (Day 28). The Day-28 restoration is the informative one: does accuracy recover, proving the environment (not the hardware) was the culprit?
  5. Analyse. Fit error-vs-time to get a degradation rate and a "half-life" (days to lose half the Day-0 headroom); correlate per-day error jumps against environmental-change magnitude; rank the three architectures by how slowly they decay.

Prerequisite: EXP-001 BLE-Assisted CSI Ground Truth Collection must be validated first — the whole experiment leans on the BLE count being trustworthy, because it is the only ground truth during the 30 days.

The plan (unexecuted) — what "done" would look like

Because nothing has run, the criteria below are targets, not results:

  • 30 consecutive days of paired CSI + BLE data (weekend gaps allowed).
  • A measurable degradation documented — >10% absolute accuracy drop from Day-0 on at least one frozen model.
  • At least one model showing a statistically significant relationship between logged environmental change and accuracy drop (with the unit of replication and multiplicity handled — see the Statistician below).
  • Enough signal to motivate the recalibration cadence work in EXP-F2.

The central deliverable is one figure: accuracy vs. days since training, one line per model, with the environmental-change events annotated on the time axis. That single plot is the thesis' "without recalibration, CSI crowd counting becomes unreliable within X days" claim — and right now X is unknown for this project's hardware and room.

Key references

  • brunello2025_d781 ↗ — the closest external result: CSI-sensing accuracy collapses under temporal (train/test time-split) shift. The paper this card most needs to reproduce for counting.
  • gama2014_8f2e ↗ — the drift-detection framing (concept drift, detectors, adaptation) the analysis inherits.
  • billah2021_69a2 ↗ — BLE occupancy F1 92.2%→70.3% in 7 days; the drift-is-real anchor (different modality — borrowed, not first-party).
  • santos2024_1e39 ↗ — environment dependence measured for the exact task (CSI crowd counting).
  • wang2026_2758 ↗, chen2023_5cbd ↗, jiang2018_77f6 ↗ — the domain-adaptation alternative the drift curve must be positioned against.
  • meneghello2023_0a93 ↗ — day-to-day CSI variability with the room untouched; the radio-stack confounder the Field Scientist flags.
  • demrozi2021_bf55 ↗, longo2019_b72f ↗ — BLE occupancy references, the ground-truth modality's real-world basis.
  • koo2026_a08d ↗ — recalibration burden differs across model families; the facilities/energy framing.
  • zhang2026_ccac ↗ — the reproducibility protocol the SWE voice wants adopted before Day 0.

Provenance

—
not recorded

Data types

  • csi-amplitude
  • csi-phase
  • ble-ground-truth
  • environmental-log
  • model-accuracy-timeseries