Back to Rahul's profile
12views

NavosEdge

Rahul Kumar Parida
  • Answered the full brief· 32 of 32
  • Expert reviewed· 5 reviews
NavosEdge

Submitted to The Edge AI Air Detective

Arduino board used?

Yes

Edge AI used?

Yes

AI/ML model used

1. GasNet MLP: Classifies environmental patterns and identifies likely pollution-source profiles using multi-gas and particulate data. 2. Anomaly Detection: Uses EWMA and dynamic Z-score analysis to identify sudden pollution spikes or abnormal environmental conditions. 3. PM2.5 Forecasting: Uses autoregressive/time-series forecasting to predict PM2.5 conditions at approximately +15, +30 and +60 minutes. 4. AQI & Advisory Engine: Uses EPA-style breakpoint calculations and deterministic rules to convert model results into actionable recommendations

Sensors used

1. MPM10-CS: Measures PM1.0, PM2.5 and PM10 particulate matter. 2. DHT22: Measures temperature and relative humidity. 3. MQ-2: Detects combustible gases, smoke, LPG and methane-related patterns. 4. MQ-9: Used for carbon monoxide and flammable-gas detection. 5. MQ-135: Provides broader air-quality/gas measurements including patterns associated with NH₃, NOx, alcohol, benzene, CO₂ and smoke.

Camera used

None

Modulino used

None

Problem being solved

Traditional air-quality monitors generally provide measurements such as PM2.5 or AQI values, but do not explain why pollution increased, what the likely source is, whether conditions are going to worsen, or what action the user should take. NavosEdge addresses this by combining multiple gas and particulate sensors with Edge AI. It can identify pollution patterns, detect abnormal changes, forecast short-term PM2.5 trends and provide actionable health recommendations. The system is also designed to remain useful during network or internet outages.

AI model training platform

The GasNet model is trained using environmental datasets, including generated and empirical sensor data. After training, the PyTorch model is exported into lightweight .npz weight matrices, allowing inference through a NumPy-based engine instead of requiring the full PyTorch runtime on the Edge device.

Arduino UNO Q features used

Heterogeneous MPU + MCU architecture: MCU handles real-time hardware operations while Linux handles higher-level processing and AI. Arduino RouterBridge + MessagePack-RPC: Used for communication between the Linux MPU and STM32U585 MCU. Hardware SPI: Drives the MPI3501 480×320 TFT display. Hardware Serial1 UART: Used for the MPM10-CS particulate sensor. 10-bit ADC: Used to acquire the MQ-2, MQ-9 and MQ-135 analog signals. Wi-Fi: Used for local communication with the parent Manager/dashboard and multi-node monitoring.

Deployment method

The STM32U585 MCU firmware is compiled for arduino:zephyr:unoq using arduino-cli and deployed using the project's flashing script. The Linux side runs the C++ hardware bridge and Python-based Intelligence Server directly on the UNO Q. A separate React + FastAPI Manager can run on a parent computer to monitor one or multiple Edge nodes over the local network.

Key innovation

The main innovation is combining real-time environmental sensing, lightweight Edge AI and actionable decision-making on a single UNO Q platform. Key aspects include: 1. Lightweight Edge AI: GasNet is exported into lightweight NumPy matrices, avoiding the need for the full PyTorch runtime during inference. 2. Observe → Decide → Predict → Explain: The physical display presents current air quality, recommended actions, future trends and model confidence. 3. Real-time architecture: Sensor acquisition and communication use non-blocking approaches so that sensing, AI processing and display operation can work continuously.

Solution summary

NavosEdge is an Edge AI-powered air-quality intelligence system built on the Arduino UNO Q. It combines five environmental sensors with local machine learning to go beyond simply displaying pollution values. The STM32U585 MCU handles real-time sensor acquisition and the physical TFT display, while the Linux MPU performs AQI calculation, pollution-source classification, anomaly detection, PM2.5 forecasting and advisory generation. Results are presented directly on the physical display and can also be synchronized to a local web dashboard through Wi-Fi.

Report (PDF)
Report (PDF)PDF
Arduino board used?

Yes

Edge AI used?

Yes

AI/ML model used

TinyGasNet (lightweight two-head neural network with gas classification and safety detection), implemented using NumPy for edge inference with Monte Carlo Dropout (20 passes) for uncertainty estimation; Decision Tree Classifier with feature-similarity scoring (IQR/MAD) for pollution source classification; Autoregressive AR(p) model for particulate matter forecasting; EWMA/MAD-based anomaly detection and baseline estimation; EPA/CPCB breakpoint-based AQI calculation; deterministic rule-based advisory engine.

Sensors used

MQ-2 – combustible gases and smoke; MQ-9 – carbon monoxide and flammable gases; MQ-135 – air-quality-related gas response (NH3, NOx, VOCs); DHT22 – temperature and relative humidity; MPM10-CS – PM1.0, PM2.5 and PM10 particulate matter.

Camera used

None

Modulino used

None

Problem being solved

Conventional environmental monitoring systems mainly provide sensor readings and limited real-time analysis, often depending on centralized infrastructure or continuous internet connectivity. NavosEdge addresses the need for an affordable, offline-first edge environmental monitoring system that collects particulate matter, gas-response, temperature and humidity data; calculates AQI; identifies potential pollution patterns and anomalies; forecasts short-term particulate trends; and generates actionable environmental and safety advisories directly on the edge device.

AI model training platform

Python-based training environment on a development machine using PyTorch for TinyGasNet neural network training, Scikit-learn for Decision Tree-based pollution source classification, and NumPy for preprocessing and model inference. Trained neural network weights are exported into lightweight NumPy-compatible (.npz) artifacts for deployment on Arduino UNO Q. Training and edge inference are separated to minimize onboard computational requirements.

Arduino UNO Q features used

Dual-core architecture using Qualcomm Linux MPU and STM32U5 microcontroller (Zephyr). STM32U5 GPIO, ADC and UART interfaces for sensor acquisition; Linux MPU for Python-based AI inference, FastAPI intelligence server, C++ hardware bridge, local data processing and communication. Arduino RouterBridge and MessagePack-RPC for inter-core communication, SPI interface for the 3.5-inch TFT display, and local processing for offline operation.

Deployment method

Edge deployment on Arduino UNO Q. Sensor firmware is compiled and flashed to the STM32U5 MCU. The C++ hardware bridge and Python FastAPI intelligence server run on the Linux MPU. Trained models are exported as lightweight NumPy artifacts (.npz) and loaded for local inference without requiring PyTorch on the edge device. A separate laptop-based Parent Edge Manager communicates with multiple nodes over a local Wi-Fi network. Supports simulated and physical sensor operating modes.

Key innovation

An offline-first, multi-node environmental intelligence system that combines real-time multi-sensor monitoring, edge-based AI inference, pollution source classification, anomaly detection, short-term air-quality forecasting, AQI calculation and automated environmental advisories. Its key innovation is deploying lightweight intelligence directly on the Arduino UNO Q using its dual-core architecture, reducing dependence on cloud computing while enabling local decision-making and centralized monitoring of multiple edge nodes.

Solution summary

NavosEdge is an offline-first Edge AI environmental monitoring system built on the Arduino UNO Q. It integrates MQ2, MQ9, MQ135, DHT22 and MPM10-CS sensors to monitor gas-response patterns, temperature, humidity and particulate matter (PM1.0, PM2.5 and PM10). Its onboard intelligence system performs AQI calculation, gas-pattern classification, anomaly detection, pollution source analysis, short-term particulate forecasting and deterministic advisory generation. A physical TFT display presents real-time environmental information and recommendations. The system uses lightweight NumPy-based inference on the Linux MPU and communicates with a laptop-based Parent Edge Manager for multi-node monitoring over local Wi-Fi. Its offline-first architecture enables local monitoring and decision-making without continuous internet or cloud connectivity.

Report (PDF)
Report (PDF)PDF
Schematic
Schematic: NavosEdge