ARIA: Hardware Neural Network Accelerator and Edge Air Quality Classifier on Tang Nano 20K

ARIA (Air Quality Real-time Intelligence Accelerator) is a real-time air quality classification system based on the Tang Nano 20K FPGA. Its core breakthrough is the implementation of an INT8-quantized two-layer neural network hardware inference pipeline, which completes the entire process from data collection to classification without relying on the cloud or external processors. The system adopts a dual-wearable device architecture (exposure monitoring + physiological response), fuses environmental and physiological data, and has the advantages of low latency, low power consumption, and strong privacy, making it suitable for wearable devices and personal health monitoring scenarios.

• Original author/maintainer: fareeha-ffk
• Source platform: GitHub, project link: https://github.com/fareeha-ffk/ARIA
• Release date: June 16, 2026 Traditional cloud-based inference has high latency, high power consumption, and privacy leakage risks; single environmental monitoring cannot combine individual physiological responses, and single physiological monitoring is prone to false alarms. ARIA aims to solve these problems and realize intelligent classification at the edge.

Dual-Device Architecture

1. Exposure monitoring device: Collects environmental data such as PM2.5, VOC, temperature, humidity, and CO2;
2. Physiological response device: Monitors physiological indicators such as heart rate, blood oxygen saturation (SpO2), respiratory rate, and heart rate variability (HRV);
3. Inference core: The Tang Nano 20K FPGA runs an INT8-quantized two-layer neural network, fusing 5-dimensional inputs (PM2.5, VOC, heat index, heart rate, SpO2) to output three types of results: safe/warning/dangerous. Hardware Implementation: Uses Verilog RTL-level design. INT8 quantization reduces resource usage, custom clock logic handles synchronization issues, and fault-tolerant mechanisms are designed for edge environments.

Resource Constraint Response: Through network pruning (two-layer minimal architecture), INT8 quantization (reduces storage and computation by 75%), fixed-point arithmetic, and RTL optimization to adapt to FPGA resources; Sensor Reliability: Clock synchronization mechanism ensures data consistency, outlier filtering handles faulty readings, and uses other data to compensate for inference when sensors fail; Power Balance: Local inference reduces communication power consumption, balancing computation and communication power consumption; Mobile Feedback Loop: The supporting App collects user symptom feedback for model verification, personalized calibration, and long-term health tracking.

Current Stage: Implements the INT8 two-layer neural network inference core on the Tang Nano 20K, supports 5-dimensional sensor input, and completes Verilog RTL design and verification; Future Plans: Integrate the complete dual-wearable device, introduce BiLSTM for time-series pattern analysis, and implement time-dimensional anomaly detection (e.g., identifying progressive health risks).

• Hardware-Algorithm Collaboration: Customize network architecture and quantization strategy for FPGA characteristics;
• Multimodal Fusion: Combining environmental and physiological data improves classification accuracy;
• Privacy First: Local inference protects sensitive health data;
• Open Hardware Ecosystem: The low-cost domestic FPGA platform (Tang Nano 20K) lowers the threshold for edge AI prototype development.

The ARIA architecture can be extended to:

1. Occupational health monitoring (exposure monitoring for factories, miners, etc.);
2. Sports health optimization (balancing air quality and physiological load during running/cycling);
3. Children's health protection (environmental monitoring in schools/kindergartens);
4. Elderly care (environmental safety and health monitoring for elderly living alone);
5. Urban environmental research (crowdsourced air quality map drawing).