ADR-048: Adaptive CSI Activity Classifier

Field	Value
Status	Accepted
Date	2026-03-05
Deciders	ruv
Depends on	ADR-024 (AETHER Embeddings), ADR-039 (Edge Processing), ADR-045 (AMOLED Display)

Context

WiFi-based activity classification using ESP32 Channel State Information (CSI) relies on hand-tuned thresholds to distinguish between activity states (absent, present_still, present_moving, active). These static thresholds are brittle — they don't account for:

• Environment-specific signal patterns: Room geometry, furniture, wall materials, and ESP32 placement all affect how CSI signals respond to human activity.
• Temporal noise characteristics: Real ESP32 CSI data at ~10 FPS has significant frame-to-frame jitter that causes classification to jump between states.
• Vital signs estimation noise: Heart rate and breathing rate estimates from Goertzel filter banks produce large swings (50+ BPM frame-to-frame) at low confidence levels.

The existing threshold-based approach produces noisy, unstable classifications that degrade the user experience in the Observatory visualization and the main dashboard.

Decision

1. Three-Stage Signal Smoothing Pipeline

All CSI-derived metrics pass through a three-stage pipeline before reaching the UI:

Stage 1: Adaptive Baseline Subtraction

• EMA with α=0.003 (~30s time constant) tracks the "quiet room" noise floor
• Only updates during low-motion periods to avoid inflating baseline during activity
• 50-frame warm-up period for initial baseline learning
• Subtracts 70% of baseline from raw motion score to remove environmental drift

Stage 2: EMA + Median Filtering

• Motion score: Blended from 4 signals (temporal diff 40%, variance 20%, motion band power 25%, change points 15%), then EMA-smoothed with α=0.15
• Vital signs: 21-frame sliding window → trimmed mean (drop top/bottom 25%) → EMA with α=0.02 (~5s time constant)
• Dead-band: HR won't update unless trimmed mean differs by >2 BPM; BR needs >0.5 BPM
• Outlier rejection: HR jumps >8 BPM/frame and BR jumps >2 BPM/frame are discarded

Stage 3: Hysteresis Debounce

• Activity state transitions require 4 consecutive frames (~0.4s) of agreement before committing
• Prevents rapid flickering between states
• Independent candidate tracking resets on new direction changes

2. Adaptive Classifier Module (adaptive_classifier.rs)

A Rust-native environment-tuned classifier that learns from labeled JSONL recordings:

Feature Extraction (15 features)

#	Feature	Source	Discriminative Power
0	variance	Server	Medium — temporal CSI spread
1	motion_band_power	Server	Medium — high-frequency subcarrier energy
2	breathing_band_power	Server	Low — respiratory band energy
3	spectral_power	Server	Low — mean squared amplitude
4	dominant_freq_hz	Server	Low — peak subcarrier index
5	change_points	Server	Medium — threshold crossing count
6	mean_rssi	Server	Low — received signal strength
7	amp_mean	Subcarrier	Medium — mean amplitude across 56 subcarriers
8	amp_std	Subcarrier	High — amplitude spread (motion increases spread)
9	amp_skew	Subcarrier	Medium — asymmetry of amplitude distribution
10	amp_kurt	Subcarrier	High — peakedness (presence creates peaks)
11	amp_iqr	Subcarrier	Medium — inter-quartile range
12	amp_entropy	Subcarrier	High — spectral entropy (motion increases disorder)
13	amp_max	Subcarrier	Medium — peak amplitude value
14	amp_range	Subcarrier	Medium — amplitude dynamic range

Training Algorithm

• Multiclass logistic regression with softmax output
• Mini-batch SGD (batch size 32, 200 epochs, linear learning rate decay)
• Z-score normalisation using global mean/stddev computed from all training data
• Per-class statistics (mean, stddev) stored for Mahalanobis distance fallback
• Deterministic shuffling (LCG PRNG, seed 42) for reproducible results

Training Data Pipeline

1. Record labeled CSI sessions via POST /api/v1/recording/start {"id":"train_<label>"}
2. Filename-based label assignment: *empty*→absent, *still*→present_still, *walking*→present_moving, *active*→active
3. Train via POST /api/v1/adaptive/train
4. Model saved to data/adaptive_model.json, auto-loaded on server restart

Inference Pipeline

1. Extract 15-feature vector from current CSI frame
2. Z-score normalise using stored global mean/stddev
3. Compute softmax probabilities across 4 classes
4. Blend adaptive model confidence (70%) with smoothed threshold confidence (30%)
5. Override classification only when adaptive model is loaded

3. API Endpoints

Method	Endpoint	Description
POST	/api/v1/adaptive/train	Train classifier from train_* recordings
GET	/api/v1/adaptive/status	Check model status, accuracy, class stats
POST	/api/v1/adaptive/unload	Revert to threshold-based classification
POST	/api/v1/recording/start	Start recording CSI frames (JSONL)
POST	/api/v1/recording/stop	Stop recording
GET	/api/v1/recording/list	List available recordings

4. Vital Signs Smoothing

Parameter	Value	Rationale
Median window	21 frames	~2s of history, robust to transients
Aggregation	Trimmed mean (middle 50%)	More stable than pure median, less noisy than raw mean
EMA alpha	0.02	~5s time constant — readings change very slowly
HR dead-band	±2 BPM	Prevents display creep from micro-fluctuations
BR dead-band	±0.5 BPM	Same for breathing rate
HR max jump	8 BPM/frame	Outlier rejection threshold
BR max jump	2 BPM/frame	Outlier rejection threshold

Consequences

Benefits

• Stable UI: Vital signs readings hold steady for 5-10+ seconds instead of jumping every frame
• Environment adaptation: Classifier learns the specific room's signal characteristics
• Graceful fallback: If no adaptive model is loaded, threshold-based classification with smoothing still works
• No external dependencies: Pure Rust implementation, no Python/ML frameworks needed
• Fast training: 3,000+ frames train in <1 second on commodity hardware
• Portable model: JSON serialisation, loadable on any platform

Limitations

• Single-link: With one ESP32, the feature space is limited. Multi-AP setups (ADR-029) would dramatically improve separability.
• No temporal features: Current frame-level classification doesn't use sequence models (LSTM/Transformer). Could be added later.
• Label quality: Training accuracy depends heavily on recording quality (distinct activities, actual room vacancy for "empty").
• Linear classifier: Logistic regression may underfit non-linear decision boundaries. Could upgrade to 2-layer MLP if needed.

Future Work

• Online learning: Continuously update model weights from user corrections
• Sequence models: Use sliding window of N frames as input for temporal pattern recognition
• Contrastive pretraining: Leverage ADR-024 AETHER embeddings for self-supervised feature learning
• Multi-AP fusion: Use ADR-029 multistatic sensing for richer feature space
• Edge deployment: Export learned thresholds to ESP32 firmware (ADR-039 Tier 2) for on-device classification

Files

File	Purpose
crates/wifi-densepose-sensing-server/src/adaptive_classifier.rs	Adaptive classifier module (feature extraction, training, inference)
crates/wifi-densepose-sensing-server/src/main.rs	Smoothing pipeline, API endpoints, integration
ui/observatory/js/hud-controller.js	UI-side lerp smoothing (4% per frame)
data/adaptive_model.json	Trained model (auto-created by training endpoint)
data/recordings/train_*.jsonl	Labeled training recordings