Periodic BLE Calibration Campaigns

Motivation

This is the core thesis contribution. EXP-002 establishes that CSI models degrade over time. This experiment demonstrates that periodic BLE-driven recalibration restores accuracy cheaply.

Koo et al. (2026) ↗ explicitly calls for "lightweight self-calibration routines that run in the background without uploading data." No paper in the vault combines BLE-triggered calibration campaigns with CSI crowd counting recalibration. The closest work is BLE Can See ↗, which uses reinforcement learning to adapt BLE-based (not CSI-based) occupancy detection — but it does not use BLE as a calibration signal for a separate CSI system.

Setup

Prerequisites

• EXP-002 completed: degradation baseline established
• Same physical device deployment as EXP-001/002

Calibration Campaign Definition

A calibration campaign is a short, scheduled data collection session where:

1. The mobile app receives a TCP/UDP command to enter "calibration mode"
2. In calibration mode, the app increases BLE advertisement rate (e.g., 100ms interval instead of 1s)
3. Participants with the app walk through the room naturally for 10-30 minutes
4. The system collects CSI data with reliable BLE-derived ground truth labels
5. The existing model is fine-tuned on this fresh labeled data
6. The app exits calibration mode

Calibration Strategies

Strategy D is the control group — the frozen models from EXP-002.

Procedure

Phase 1: Strategy Comparison (4 weeks)

Run all strategies in parallel using separate model copies:

1. Week 1: Let all models degrade (continue EXP-002 baseline measurement)
2. Week 2: Start Strategy A (daily calibration) and Strategy B (weekly calibration)
3. Week 3: Start Strategy C (trigger-based calibration)
4. Week 4: Continue all strategies, collect final comparison data

For each calibration event:

1. Record pre-calibration accuracy (evaluate model on last hour of data before campaign)
2. Run the campaign: collect CSI + BLE ground truth
3. Fine-tune the model:
   • Fine-tune approach: update last N layers of the model using campaign data
   • Learning rate: 10x lower than original training (to avoid catastrophic forgetting)
   • Epochs: 5-10 (time-limited fine-tuning)
4. Record post-calibration accuracy (evaluate on next hour after campaign)
5. Log: campaign timestamp, duration, number of participants, data volume, training time

Phase 2: Campaign Duration Optimization (1 week)

Using the best strategy from Phase 1:

• Run campaigns of varying duration: 5 min, 10 min, 15 min, 30 min, 60 min
• Measure accuracy recovery vs. campaign duration
• Find the minimum viable campaign duration — the shortest campaign that restores accuracy to within 5% of Day 0

Phase 3: Mobile App Integration (1 week)

Test the full automated pipeline:

1. Server detects model drift (CSI prediction diverges from BLE count)
2. Server sends TCP/UDP push notification to mobile apps in the area
3. Apps enter calibration mode automatically
4. System collects calibration data without human intervention
5. Model is fine-tuned automatically
6. Apps exit calibration mode

Expected Outputs

• dataset/exp003/campaigns/ — one folder per calibration event with CSI + BLE data
• dataset/exp003/accuracy_timeline.csv — continuous accuracy with calibration event markers
• dataset/exp003/strategy_comparison.csv — per-strategy accuracy recovery metrics
• dataset/exp003/duration_sweep.csv — accuracy vs. campaign duration

Analysis Plan

Primary Metrics

1. Accuracy recovery: post-calibration accuracy / Day 0 accuracy (target: >95%)
2. Recovery speed: how many minutes of calibration data needed to reach 95% of Day 0?
3. Cost-effectiveness: accuracy improvement per minute of calibration
4. Trigger sensitivity: for Strategy C, what divergence threshold minimizes total calibration time while maintaining >90% accuracy?

Key Figures for Thesis

• Figure 1: Accuracy timeline with calibration events marked — shows the "sawtooth" pattern (degrade → calibrate → recover → degrade)
• Figure 2: Strategy comparison bar chart — accuracy maintenance cost vs. achieved accuracy
• Figure 3: Campaign duration vs. accuracy recovery curve — the "diminishing returns" curve
• Figure 4: Trigger-based strategy: detected drift events vs. actual environmental changes

Success Criteria

• At least one calibration strategy maintains >90% of Day 0 accuracy over 4 weeks
• 10-minute calibration campaign restores >85% of lost accuracy
• Strategy C (trigger-based) achieves comparable accuracy to Strategy A (daily) with <50% of total calibration time
• End-to-end automated calibration pipeline works without human intervention

Related Work in Vault

• Green Wireless Sensing ↗ — "lightweight self-calibration routines" as future direction
• WiFi Sensing Generalizability ↗ — continual learning after deployment
• Meneghello et al. 2023 ↗ — few-shot domain adaptation for CSI sensing
• Cross-Domain WiFi Sensing Survey ↗ — domain-invariant feature extraction
• BLE Can See ↗ — RL-based adaptation for BLE occupancy (our CSI equivalent)

Dependencies

• EXP-002 (degradation baseline)
• Mobile app must support TCP/UDP command for calibration mode toggle
• Fine-tuning pipeline for the trained models (scripted, not manual)

Notes

The trigger-based strategy (C) is the most interesting for the thesis because it is autonomous — the system detects when it needs calibration and requests it. This aligns with the "self-healing" vision from Koo et al. 2026 ↗.

The "sawtooth" accuracy pattern (Figure 1) is expected to be the signature result of the thesis — it visually demonstrates that BLE calibration campaigns are an effective, low-cost countermeasure to CSI model drift.