Aeris: An Intelligent Air Quality Monitoring and Anomaly Detection System Based on Time-Segmented Machine Learning

Aeris is an intelligent air quality monitoring and anomaly detection system designed specifically for Surabaya, Indonesia, developed by a student team from the 2026 cohort of the Electronic Engineering Polytechnic Institute of Surabaya (PENS). The system adopts a time-segmented machine learning strategy, combines PyCaret AutoML to achieve 3-hour pollution prediction, uses Isolation Forest for real-time anomaly detection, and strictly follows the official ISPU standards of Indonesia's Ministry of Environment and Forestry (KLHK). The project source code is open-sourced on GitHub, aiming to provide accurate data support for air quality management.

Urban air quality is highly dynamic, with significant differences in pollution characteristics across different time periods in Surabaya: During the morning rush hour (06:00-11:59), PM2.5 surges due to exhaust emissions and temperature inversion; at noon (12:00-17:59), high temperature and strong sunlight accelerate ozone formation; in the evening/night (18:00-05:59), evening rush hour and temperature inversion lead to pollutant accumulation. Traditional systems use a single model to cover the entire day, ignoring time-varying patterns, which limits prediction accuracy.

Aeris is a graduation project by students majoring in Data Science at PENS, with the team dividing tasks to complete the entire process from data collection to deployment. The core innovation is time-segmented modeling: dividing a day into 3 time periods with unique pollution characteristics, training 3 period-specific models for each of the five pollutants (PM2.5, PM10, CO, NO2, O3) (total 15 models), and the system automatically selects the model based on the current time to improve prediction accuracy.

Technical Architecture: The layered design includes a PostgreSQL data layer, a machine learning layer with PyCaret + Isolation Forest, a FastAPI backend API, a responsive front-end dashboard, and an APScheduler scheduling system.

Data Pipeline: Extracts 21-dimensional features, including time features (hour, day of the week, etc.), lag features (1/3/24-hour lag values), rolling statistics (mean/standard deviation/maximum), rate of change features, meteorological features (temperature/humidity, etc.), and ISPU features.

Model Selection: Uses PyCaret AutoML to automatically compare 15 regression algorithms (Random Forest, XGBoost, etc.), selects the optimal model for each pollutant-time period combination, and MLflow is used for experiment tracking and version management.

Anomaly Detection: Runs the Isolation Forest module in parallel to determine whether the current observation is abnormal (sudden pollution or sensor failure), complementing the prediction model to provide comprehensive data support.

ISPU Standards: Strictly complies with KLHK Ministerial Decree No.14 of 2020, classifying air quality into 6 levels (Good - Hazardous), with the dashboard using corresponding color coding, and abnormal levels automatically triggering warnings.

Deployment Operations: Deployed via Docker containerization, supporting one-click startup; automatically retrains models weekly to ensure performance; provides local development guidelines and supports custom configurations.

The project's slogan is "Clean Air, Bright Future - Early Detection, Wise Action", embodying the concept of technology serving society. The system can be extended to other cities with similar pollution patterns, providing a reference for global air quality monitoring. Aeris demonstrates the practical application of machine learning in environmental science and is an ideal output combining academic research and engineering practice.