Eco-Track · PM2.5 Monitoring & Forecasting

About the Project

“Eco-Track: An AIoT System for Local PM2.5 Monitoring, Forecasting, and Real-Time Health Alerts” was conducted in the 2025–2026 academic year at Saigon International University (SIU). I lead the project with Bùi Lê Hồng Tâm and Võ Lê Thiên Ân, supervised by Nguyễn Đặng Thế Vinh, MSc.

The monitoring system uses ESP32 nodes, APM2000 particulate sensors, and SHT30 temperature/humidity sensors, aggregating measurements by minute. MQTT streams through Mosquitto/Telegraf write raw data to InfluxDB. A Python service standardizes data, performs quality control (QC), identifies and preserves pollution events, and synchronizes cleaned data to MongoDB.

The forecasting component compares Hybrid CNN-LSTM, LSTM Baseline, XGBoost, and Simple CNN-LSTM using historical OpenAQ data from District 3, Ho Chi Minh City. According to the report, Hybrid CNN-LSTM achieved MAE of 1.877 µg/m³ and RMSE of 2.507 µg/m³ on this experiment's test set.

Field experiments used co-located measurements in An Khanh Ward during March–April 2026, compared with a reference device. Keeping the model fixed and replacing raw inputs with QC/Event-processed data reduced one-hour-ahead forecast RMSE from 1.8365 to 1.7685 µg/m³ (3.70%) and MAE from 1.6307 to 1.5671 µg/m³ (3.90%). These results apply to the report's data and evaluation procedure.

The application provides a web dashboard, proactive Telegram alerts, multi-node management, device logs, and OTA firmware updates through an admin panel. The cleaned-data pipeline has a processing delay of several minutes to finalize event labels.

Awards: First Prize in the Computer Science section at the 8th SIU Student Research Competition, academic year 2025–2026; and a Consolation Prize at the Ministry of Education and Training's 2026 Science and Technology Awards for Students in Higher Education Institutions, project 205-102.

Future work includes longer reference-device comparisons and calibration, additional weather/traffic variables, and improved operational resilience and security. The report notes the short co-location period and incomplete integration of exogenous variables.

The project's source code is private.