AI Strategy: Wi-Fi Collapse Detection

Technical Brief for Development Team

This document outlines the machine learning pipeline for detecting Wi-Fi "Collapse" events (Hidden Node, Saturation, Interference) using ESP32 sensors. We utilize an LLM-Assisted Weak Supervision approach to overcome the lack of labeled training data.

🚀 The Core Concept

We cannot hard-code thresholds because RF environments vary wildly. Instead, we use Gemini (LLM) as a "Senior Network Engineer" to label raw logs, and then train a fast, lightweight model (Random Forest/XGBoost) to mimic that decision logic in real-time.

Phase 1: Firmware Data Engineering

The ESP32 firmware is the data generator. It must output time-series features suitable for ML, not just human-readable logs.

1. Feature Vector (The CSV)

Every 1 second, the firmware writes a CSV line to the internal storage partition. This is our feature set:

Timestamp: Epoch time.
RetryRate: (Float 0-100) Percentage of frames requiring retransmission.
AvgNAV: (UInt16) Average Network Allocation Vector duration (microseconds).
MaxNAV: (UInt16) Peak contention window seen in that second.
Collisions: (UInt8) Count of inferred collision events.
AvgPHY: (UInt16) Average data rate (Mbps). Low PHY + High NAV = Bad.
Mismatches: (UInt8) Count of packets where duration > expected airtime.

2. Storage Strategy

We use a custom partition table to allocate ~5-13MB for LittleFS/SPIFFS. NVS is strictly for config. This allows 24-72 hours of continuous logging.

Phase 2: The "Weak Supervision" Pipeline

We solve the "Cold Start Problem" (having data but no labels) by using Generative AI.

Step A: Context-Aware Capture

Technicians capture logs in known scenarios. The filename conveys the context:

microwave_interference.csv
hidden_node_scenario_A.csv
clean_baseline.csv

Step B: LLM Labeling (Gemini)

We feed the raw CSV chunks to Gemini via API with a prompt that injects domain knowledge:

"Act as a Wi-Fi expert. Analyze this CSV log. This data represents a Hidden Node scenario. Look for periods where AvgNAV is high (>10ms) but AvgPHY remains normal, yet Retries spike. Label each timestamp row as 'Normal', 'Congestion', or 'Collapse'."

Result: A "Silver Standard" labeled dataset ready for supervised learning.

Phase 3: Training & Inference

Gemini is too slow for real-time packet analysis. We train a classical model for the actual work.

1. Model Selection

Algorithm: Random Forest Classifier or XGBoost.

Why? They excel at tabular data, handle non-linear relationships (e.g., high NAV is fine unless Retries are also high), and have microsecond inference times.
Input: The 7-column Feature Vector from Phase 1.
Output: Probability Score (0.0 - 1.0) for "Collapse".

2. Runtime Inference Architecture

Deployment Target: Linux Gateway / Edge Server.

1 ESP32 streams CSV lines via UDP (Broadcast/Unicast) to Linux.
2 Python script listens on UDP port.
3 Script loads pre-trained model.pkl.
4 Incoming CSV -> Feature Vector -> model.predict().
5 If Collapse_Prob > 0.8 for 3 consecutive seconds -> TRIGGER ALERT.