BREATHE: AIR QUALITY PREDICTION USING EMBEDDED MACHINE LEARNING AND DEEP LEARNING MODELS WITH QUANTIZATION TECHNIQUES

Abstract: Air quality degradation poses a significant global health challenge, necessitating accurate and intelligent systems capable of real time pollutant forecasting and actionable wellness guidance for everyday users. The complexity of spatial and temporal pollutant dynamics across urban environments demands advanced deep learning architectures capable of multi city, multi pollutant prediction while remaining deployable on resource constrained mobile devices. This project presents Breathe, a cross platform air quality intelligence system employing a hybrid GCN-TransGRU architecture combining Graph Convolutional Networks for spatial inter city relationship modeling with Transformer encoders and Gated Recurrent Units for capturing temporal dependencies across PM2.5, PM10, and NO2.

To enable efficient mobile deployment, a Knowledge Distillation strategy compresses a high capacity teacher model of approximately 2.52M parameters into a lightweight student model of approximately 466k parameters into a lightweight student model of approximately 43k parameters, exported as an ONNX float32 model (~1.51 MB) and deployed via ONNX Runtime for on-device inference without significant accuracy loss. The system is implemented across three integrated components including a deep learning forecasting model, a modular NestJS backend API managing authentication, real time air quality data, and trip planning, and a Flutter based mobile application featuring a dynamic AQI dashboard, personalized health profiles, and resilience on network loss (last loaded data retained in memory). By combining spatial temporal deep learning with scalable cloud infrastructure, Breathe contributes to improved public awareness, reduced health risks, and the advancement of technology driven air quality management.

Keywords: Air Quality Index (AQI), Deep Learning Forecasting, Multi-Pollutant Prediction, Spatial-Temporal Modeling, Graph Convolutional Networks (GCN), Transformer-GRU Hybrid Architecture, Knowledge Distillation, Model Compression, On-Device Inference, ONNX Runtime, Cross-Platform Mobile Applications, Flutter, NestJS Backend API, Modular Architecture, Software Engineering.