AI-based AHU Energy Analysis

Predictive Maintenance for Air Handling Units

Daewoong ENG
ML LLM RAG Predictive Maintenance IoT

Developing a predictive maintenance system for Air Handling Units (AHU) and core production utility equipment using machine learning and LLM-based anomaly detection. The system enables proactive maintenance scheduling and minimizes production downtime risks.

System Overview

Integrated CatBoost machine learning model with RAG-LLM pipeline for:

  • Digital Twin (CatBoost): Predicts whether current valve openings are normal based on historical operational patterns, considering features like outdoor temperature, humidity, return temperature, and other sensor data
  • Anomaly Detection: Real-time comparison between predicted normal behavior and actual sensor readings
  • Fault Diagnosis: Automated analysis of potential causes when anomalies detected
  • Predictive Alerts: Proactive notifications before equipment failure

Technologies

Machine Learning & AI

  • CatBoost (Digital Twin): Gradient boosting model trained on historical data to predict expected normal valve positions based on current environmental conditions (outdoor temperature, humidity, return temperature, supply temperature, etc.)
  • RAG/LLM Pipeline (Qdrant): Knowledge retrieval for fault diagnosis and maintenance recommendations when digital twin detects deviations from expected behavior
  • Vector Database: Historical maintenance logs and equipment manuals indexed for quick retrieval

Data Infrastructure

  • Airflow: Data pipeline orchestration for sensor data collection
  • SQL: Historical data storage and querying

User Interface

  • Next.js + TypeScript + Tailwind: Modern dashboard for real-time monitoring
  • Telegram Bot: Instant alerts to maintenance team

Key Contributions

Maintenance Operations

  • GMP-compliant preventive maintenance for:
    • Air Handling Units (AHU)
    • Air compressors
    • Constant temperature/humidity units
    • BMS/EMS systems

Monitoring & Analytics

  • Real-time dashboard tracking:
    • Heating/cooling valve positions
    • Fan power consumption
    • Temperature and humidity levels
  • Energy efficiency analysis and optimization recommendations

System Improvements

  • Minimized Downtime: Predictive alerts reduce unplanned shutdowns
  • Cost Optimization: Better spare parts management and maintenance scheduling
  • Compliance: Full GMP documentation and audit trail