Hybrid Fusion and Cross-Geometry Transfer — BLE+CSI fusion across phone penetration and across rooms

Motivation

EXP-F3 closes the field-side hybrid-fusion + cross-geometry story. Merges the superseded EXP-004 (multi-room transfer) and EXP-005 (BLE+CSI fusion) into one 2-D ablation per diary/2026-05-20 - Experiment Series Reorganization:

• Axis 1: phone penetration (S1 → S5). What happens to the BLE+CSI fusion as the BLE-side signal degrades?
• Axis 2: cross-room transfer (A → B). What happens to the fusion model when geometry changes?

Both axes are ablations on the same fusion model, share data infrastructure, share evaluation. Running them as one experiment is operationally identical.

Setup

Two-room deployment

Identical kit per room — 4× Pi 5 + AX210, 2× ESP32-C6 BLE listeners, 2× ESP32-C5 supplementary CSI. The EXP-F1 / F2 room is room A. Room B is the next room PostGIS polygonizes (target: a seminar room of different dimensions than A — see diary/2026-05-20 - Experiment Series Reorganization §"GIS state").

If only one room is available when EXP-F3 starts, the cross-geometry axis is deferred but the phone-penetration axis runs in room A immediately. Don't block the whole experiment on the second-room polygonization.

Phone-penetration scenarios

Same scenario tree as the superseded EXP-005, app-driven:

App enforces who carries an active beacon by toggling BLE adv on/off per-phone (programmable via the gamification loop from EXP-P1 Phase C).

Fusion model

Feature-level fusion model — input is the concatenation of per-window per-link CSI features and per-cell BLE features. Same model used in EXP-F2 (Bagging-LGBM and the PINN-with-continuity-residual), updated to accept BLE features.

Train in room A on scenarios S1 + S5 (the extremes), evaluate on S2 / S3 / S4 (the held-out penetrations).

Procedure

Phase A — Fusion baselines in room A

1. Deploy room A kit. Train Day-0 fusion model (LGBM + PINN) on S1 + S5 with full BLE-anchored ground truth.
2. Evaluate on S2 + S3 + S4 (held out across penetrations).
3. Compare against BLE-only baseline (count of unique adverts) and CSI-only baseline (the EXP-F2 frozen models).
4. Headline figure: phone-penetration sweep — MAE vs penetration rate, one line per modality (BLE-only, CSI-only, fusion).

Phase B — Zero-shot transfer A → B

1. Deploy room B kit (when room B polygonization arrives).
2. Apply room-A fusion model to room B without retraining. Run S3 (50 % penetration) for 1 h.
3. Record accuracy → zero-shot baseline.

Phase C — BLE-calibrated transfer A → B

1. Run a 30-min BLE calibration campaign in room B per the EXP-F2 protocol — but starting from the room-A model.
2. Fine-tune the model on the 30-min room-B campaign.
3. Run S3 for 1 h, record accuracy.
4. Repeat with 10-min and 20-min campaigns for the duration sweep.

Phase D — Native room-B baseline

1. Train a fresh fusion model in room B on its own S1 + S5 data (∼1 h labelled).
2. Evaluate on S3 for 1 h.
3. Comparison: BLE-calibrated-transfer (Phase C) vs native room-B (Phase D) — what fraction of native accuracy does the transfer achieve?

Expected outputs

• _attachments/exp-f3/room_a/<scenario>/ — per-scenario room-A data.
• _attachments/exp-f3/room_b/<scenario>/ — per-scenario room-B data.
• _attachments/exp-f3/models/<phase>/<model>_<run_id>.pt — checkpoints per phase.
• _attachments/exp-f3/results/penetration_sweep.parquet — Phase A metrics.
• _attachments/exp-f3/results/transfer.parquet — Phase B/C/D metrics.
• _attachments/exp-f3/figures/ — penetration sweep, transfer comparison, cross-modal failure-mode plots.

Analysis plan

Primary metrics

1. Fusion gain vs single modality at each penetration rate — MAE delta.
2. Penetration robustness — fusion MAE curve as penetration drops from 100 % to 0 %.
3. Critical penetration — at what penetration rate does CSI alone match fusion? (where fusion stops paying for itself).
4. Zero-shot transfer accuracy — fraction of native-room-B accuracy achieved without retraining.
5. BLE-calibrated transfer accuracy — fraction achieved with 10 / 20 / 30 min campaign.
6. Data efficiency — room-A 1 h + room-B 30 min ≈ how much room-B labelled data does this save?

Key figures

• Fig 1. Phone-penetration sweep — MAE vs penetration, one line per modality. The fusion line should lie below both single-modality lines at all penetrations.
• Fig 2. Cross-geometry transfer ladder — zero-shot → 10 min → 20 min → 30 min → native, MAE bar chart.
• Fig 3. Per-regime failure-mode breakdown — at low penetration, where does fusion fail (which cells, which regimes)?

Success criteria

• Penetration axis: fusion achieves lower MAE than the best single modality at penetration rates 25–75 %.
• Graceful degradation: at 0 % penetration (S5), fusion ≈ CSI-only (no degradation).
• At 100 % penetration: fusion outperforms CSI-only by leveraging BLE.
• Cross-geometry: 30-min BLE-calibrated transfer achieves ≥ 80 % of native-room-B accuracy.
• Mechanism mirror: PINN-fusion transfers better than LGBM-fusion (the mechanism story holds in the field — see EXP-S1 Phase B).

Risks and mitigations

1. Two-room availability. Mitigation: penetration axis runs in room A alone; cross-geometry deferred until room B polygonization arrives.
2. Different participant pools for room A vs B confound the transfer measurement. Mitigation: same participants traverse both rooms in one session if possible; if not, log participant overlap as a covariate.
3. 30 min may not be enough room-B data. Mitigation: sweep 10 / 20 / 30 min in Phase C; if 30 min is the floor of the curve, extend; if 10 min already plateaus, document.
4. The "critical penetration" finding (where CSI overtakes BLE) is environment-specific. Mitigation: report it from both rooms; if it varies across rooms, that is the finding — the BLE-vs-CSI breakeven is geometry-dependent.
5. App-driven phone-penetration toggle may be defeated by participants. Mitigation: server-side validation of which phones are advertising at each minute; reconciliation against the prescribed scenario.

Dependencies

• EXP-F2 passed — the CSI-only baseline models exist.
• Room B polygonization in PostGIS (for Phase B onwards).
• Two-room kit availability (8× Pi 5 + AX210, 4× ESP32-C6, 4× ESP32-C5).
• ≥ 8 participants for full N = 8 scenarios.

Related work in vault

• shahbazian2023_1172 — "Combined BLE and WiFi" section; survey of separate-system approaches.
• chaudhari2024_6efc ↗ — call for BLE+WiFi integration.
• santos2024_1e39 — the cross-room problem statement.
• chen2023_5cbd ↗ — five-axis DG taxonomy.
• zou2018_1590 ↗ — TKL as a fusion baseline reference (parked here, may be added if Ch7 needs it).
• torun2026_72aa — stationary-crowd regime, relevant to the "where fusion fails" analysis.

Cross-experiment links

• EXP-S1 — provides the PINN $\lambda^*$ used in the fusion model; the cross-geometry transfer test mirrors EXP-S1 Phase C's geometry-B follow-on.
• EXP-F1 — the ground-truth chain.
• EXP-F2 — the CSI-only baseline models + the calibration protocol Phase C reuses.

Notes

EXP-F3 ships two distinct figures — Fig 1 (penetration sweep) and Fig 2 (transfer ladder). The former is the field-validation of L4½.4 (fusion lift); the latter is the field-validation of L4½.6 (mechanism transfers, model parameters don't). Together they close the new physics-informed-fusion chapter introduced in 2026-05-19 - GAT thesis deployment strategy.