Asking the fleet what it is doing…
monad-knowledge Wi-Fi sensing lab · FIIT STU
WS501 active 5 / 13 weeks

Wi-Fi CSI Sensing from First Principles

A twelve-week track that builds Wi-Fi Channel State Information sensing from the physical channel upward. The throughline: no number is asserted that the learner cannot reproduce in a notebook by Friday. Every week pairs a written lecture, a Jupyter workbook with sample executions, worked examples, and a Slidev deck.

The 13-week arc

5 of 13 lectures published — each a self-contained read with a slide deck and runnable workbook. Start at the first published week and work forward.

Wk Theme & focus Read
01 The wireless channel start here Multipath, fading, why walls matter — how a signal *becomes* a channel, before any CSI read → slides · workbook
02 From OFDM to CSI Subcarriers, CFR vs CIR, what the matrix **H** the NIC reports actually is read → slides · workbook
03 Extraction toolchains Getting real CSI off commodity hardware: FeitCSI vs Nexmon vs PicoScenes, formats & gotchas read → workbook
04 The standard: IEEE 802.11bf The 802.11bf sensing session (setup→sounding→report), what a standardized CSI report costs vs the raw AX210 tensor (grouping + quantisation), why scheduling beats ambient traffic, and the calibration/ground-truth/drift gap the standard leaves open read → workbook
05 Preprocessing Amplitude/phase sanitisation, AGC removal, denoising — why raw CSI lies. First-party anchor now exists (2026-07-27 monad01→monad02 AX210 link): raw per-packet phase intercept std 3.5 rad → two-chain CSI-ratio circ-std 0.26 rad, per-tone amplitude CV 0.32 → 0.20 — teach the sanitisation ladder against these measured numbers, not only synthetic CSI; two active RX chains are a hard capture requirement (zeng2021 ratio model, wu2022 Fresnel→ratio) read → slides · workbook
06 Feature engineering Statistical / spectral / learned features, and which survive a hardware change planned
07 Occupancy detection The simplest useful task, end-to-end, on a real(istic) split planned
08 Activity & gesture recognition Windowing and temporal models; where occupancy methods break planned
09 Localisation from CSI Fingerprints vs geometry, and the limits of each indoors planned
10 The drift problem 94% on Monday, guessing by Friday — measuring the collapse (cf. EXP-002). First-party drift phenomenology in hand: the 19 h overnight ambient capture (monad01, 588k records) shows quasi-static night plateaus punctuated by discrete state jumps, not a smooth ramp. Safe decay citations: jung2025 (F1 100→66.1 %/month) + Brunello (TRRS drop within 300 s); the zou2017 "99→66 % in a week" framing is a documented mis-citation — true one-week endpoint ≈53 % accuracy, recovered ~92 % only via TKL domain adaptation planned
11 Domain adaptation & online recalibration Keeping a model alive as the channel moves under it planned
12 Reproducibility & dataset hygiene 28-datasets-one-with-video; logging firmware/hardware as data planned
13 Capstone Design and run a reproducible CSI occupancy study end-to-end planned

Capstone

Week 12 is a self-directed study: pick a room, a piece of commodity hardware, and a question; produce a notebook that someone else could re-run and a Slidev deck that defends the result honestly — including what drifted and what you logged.