AFR & Throttle Anomaly Detection for BMW MHD Logs

This release provides a fully functional anomaly detection tool for analyzing MHD log files from BMW platforms (e.g., N54/N55/S55). The model uses an Isolation Forest trained on real-world driving data and is tailored to detect high-impact engine behavior anomalies—AFR spikes, throttle lag, hesitation zones, and logging artifacts.

Key design goal: prioritize real-time usability and clear diagnostic output to assist enthusiasts, tuners, and techs reviewing logs from MHD or similar tools.

What It Detects

• AFR spikes (lean conditions, sensor faults, invalid data)
• Throttle mismatch (pedal vs. throttle lag—e.g., torque intervention or hardware delay)
• RPM hesitation in the 1600–1800 band (common for airflow or tuning lag)
• Idle AFR instability (misfire, fueling imbalance, or IACV behavior)
• MHD logging artifacts (repeating 235.19 AFR values from sensor dropout or ECU noise)

Included in This Release

• analyze_log.py: Main script to parse and flag anomalies in .csv logs
• models/train_model.py: Pipeline for training the Isolation Forest
• plugins/anomaly_detector.py: Modular inference engine for flagging and formatting
• sample_log.csv: Realistic example to test pipeline
• Clear printout format with timestamps and feature flags
• All dependencies listed in requirements.txt

Training Summary

Model trained on engineered binary features extracted from real logs:

• AFR_spike, TPS_bug, Pedal_Throttle_Mismatch, RPM_idle, RPM_hesitation
• IsolationForest (contamination=0.05) with sklearn 1.3.0
• Saved model weights included (models/anomaly_model.pkl)

Use Cases

• Quick triage of log health before tuning review
• Detection of soft misfires or airflow instabilities
• Early diagnosis of sensor failure or ECU behavior issues
• Log quality checks (sensor dropout, logging corruption)

Next Goals

• Add log visualization tools (e.g., AFR over time, throttle overlays)
• UI interface for easier usage by non-technical users (e.g. Streamlit)
• Group anomaly sequences into meaningful diagnostic chains
• Boost logic for WOT log validation and ignition timing analysis

This tool is fully usable as-is. Output is human-readable, explainable, and tailored to real-world BMW diagnostics—not abstract ML outputs.