Summary

This issue proposes adding a new plugin called Anomaly Guardrail to the OnAIR framework.
The plugin introduces configurable anomaly detection, severity classification, and severity-to-action mappings, providing a modular guardrail system for telemetry data streams.

Motivation

• Enhances OnAIR with real-time anomaly guardrails for telemetry.
• Supports severity-based responses (LOG_ONLY, REDUCE_LOAD, SAFE_MODE).
• Complements existing plugins (e.g., Kalman, generic) by focusing specifically on fault detection and classification.
• Provides a reusable framework for anomaly detection that can be tested with both CSV and Redis telemetry sources.

Features

• Outlier detection using Z-score or Interquartile Range (IQR).
• Frozen signal (“stuck” values) detection.
• Severity classification: minor, moderate, critical.
• Configurable severity-to-action mapping, e.g., minor:LOG_ONLY, moderate:REDUCE_LOAD, critical:SAFE_MODE.
• CSV audit logging of anomaly events.
• Log post-processor (count_anomalies.py) to collapse redundant anomaly runs.
• Redis-based publisher (redis-experiment-publisher-anomalies.py) to simulate mixed normal + anomalous telemetry for testing.

Usage

1. Run with the sample config:

   python driver.py onair/config/config_anomaly_guardrail.ini

2. Publish synthetic telemetry for testing:

   python3 redis-experiment-publisher-anomalies.py --normals 100 --anomalies 50 --repeat 3 --shuffle

3. Post-process anomaly logs:

   python3 count_anomalies.py --in logs/anomaly_events.csv --out logs/anomaly_events_clean.csv --mode edge --emit-end

Files to be Added

• plugins/anomaly_guardrail_plugin.py – core anomaly detection and classification logic
• onair/config/config_anomaly_guardrail.ini – configuration of thresholds, severity mapping, monitored columns
• redis-experiment-publisher-anomalies.py – telemetry simulator for Redis
• count_anomalies.py – log summarization and cleanup utility
• doc/AnomalyGuardrail.md – usage guide

Reference

Detailed usage and context included in this commit:
👉 AnomalyGuardrail@ commit