AeroSense Edge
SSudhanshu Maharathi- Expert reviewed· 5 reviews
Submitted to The Edge AI Air Detective
- Arduino board used?
Yes
- Edge AI used?
Yes
- AI/ML model used
1. Chemical Stoichiometric Source Attribution Classifier: - Model: Gradient Boosting Classifier exported to ONNX Runtime with INT8 quantization (source_classifier.onnx, 1.37 MB, opset 12). - Purpose: Classifies 6 emission sources (Traffic Exhaust, Garbage/Waste Burning, Construction Dust, Kitchen Cooking Smoke, Crop Residue Burning, and Clean Baseline) using multi-pollutant stoichiometric ratios (PM2.5/PM10, CO/CO2, NO2/VOC). - Performance: 99.03% ± 0.33% accuracy via 5-fold stratified CV on synthetic dataset; 1.14 ms inference latency on laptop CPU. Includes rule-based fallback (95.4% accuracy). 2. Diurnal Boundary Layer Time-Series Forecaster: - Model: Autoregressive Multi-Layer Perceptron (timeseries_forecaster.onnx, 4.43 KB). - Purpose: Predicts +1h, +2h, +4h, and +6h PM2.5/PM10 diurnal trends and boundary-layer thermal inversion traps (MAE 11.93 µg/m³, beats persistence baseline by 27.1%). 3. On-Device Context-Aware Advisory Engine: - Heuristic rule engine and quantized SLM prompt synthesizer translating source probabilities into persona-specific actionable directives.
- Sensors used
1. DHT 22 2, MQ 2 3. MQ 135 4. PM 2.5 5. 4 chanel relay module
- Camera used
None
- Modulino used
Modulino Pixels / Display (I2C) for on-node visual air status, emergency hazard color bar (Green/Amber/Crimson/Purple), and diagnostic feedback without requiring an external monitor.
- Problem being solved
Current air quality monitoring suffers from the "interpretation gap": citizens receive an abstract, delayed city-wide number (e.g. "AQI 142 — Moderate") from government monitoring stations 10+ km away. This number fails to capture hyperlocal reality (such as roadside diesel exhaust or open plastic burning 200m away) and tells citizens nothing about what action to take (e.g., whether to close school windows, alter a morning jog, or dispatch waste security). Furthermore, traditional monitors stream raw data continuously to commercial clouds ($60–$240/node/year in recurring cellular/SaaS fees), failing completely during power cuts or network outages and creating privacy risks. AeroSense Edge solves this with a 100% offline, decentralized Edge AI station that attributes the exact chemical pollution source, forecasts local trends, and issues actionable directives at $0.00/year cloud OpEx.
- AI model training platform
- Scikit-learn (v1.9.0) with stratified 5-fold cross-validation and rolling-origin temporal CV for Gradient Boosting classification and MLP time-series regression. - ONNX & skl2onnx (opset 12) with INT8 dynamic quantization pipeline for edge execution on ARM Cortex-A53 / Qualcomm QRB2210 via ONNX Runtime. - Synthetic atmospheric chemistry generator (6,000 samples) grounded in atmospheric boundary-layer physics and stoichiometric distributions from EPA and CPCB field source apportionment studies.
- Arduino UNO Q features used
1. Heterogeneous Dual-Compute Architecture: - Qualcomm Dragonwing QRB2210 Linux MPU (Quad-Core Cortex-A53 @ 1.3 GHz): Executes ONNX Runtime INT8 ML inference (1.14 ms latency), 30-day circular SQLite database (WAL mode), Gorilla-style Delta-RLE telemetry compression (99.9% reduction), and embedded FastAPI Web SCADA daemon. - STM32U585 MCU (ARM Cortex-M33 @ 160 MHz): Microsecond deterministic sensor polling via DMA and hardware timers (PMS5003 UART, SCD41/AHT30 I2C, MiCS-6814 ADC) with zero Linux OS scheduling jitter. 2. Inter-Core IPC: High-speed internal virtual UART and lock-free shared ring buffer (RingBufferIPC.h) synchronizing sensor packets between STM32 and Qualcomm Linux. 3. On-Board Connectivity: Local Wi-Fi/mDNS for zero-cloud browser dashboard (http://localhost:8000) and peer-to-peer spatial mesh interpolation. 4. Power & Bus Management: Low-power MCU duty cycles and Modulino I2C bus integration.
- Deployment method
1. Linux Application Daemon: Python 3.11/FastAPI service (edge_core/app.py) running as a systemd service on the Qualcomm QRB2210 Debian Linux environment. 2. Bare-Metal Real-Time Firmware: Flashed to STM32U585 MCU via Arduino App Lab / Arduino CLI (uno_q_mcu.ino), running sensor drivers, calibration offsets, and DMA IPC ring buffer. 3. Zero-Cloud Web SCADA: Embedded glassmorphic web HUD served directly from the node via port 8000 using Vanilla HTML5/CSS3/ES6 (no external npm or CDN dependencies). 4. Flutter Mobile App: Companion Android/Desktop client syncing over local Wi-Fi/mDNS/BLE with offline simulation fallback.
- Key innovation
1. Chemical Stoichiometric Source Fingerprinting: Instead of treating all particulate matter the same, our edge model calculates real-time chemical stoichiometric ratios (PM2.5/PM10 particle ratio, CO/CO2 combustion inefficiency, and NO2/VOC vehicular index) to attribute emissions to specific causes (diesel traffic, garbage burning, construction dust, cooking) in 1.14 ms on-device. 2. Actionable Guidance vs. Abstract Numbers: Converts complex multi-gas telemetry into persona-specific action directives for school administrators, joggers, residents, and municipal patrols. 3. Dual-Silicon Jitter Elimination: Decouples microsecond sensor acquisition on the STM32 MCU from heavy AI and web workloads on the Qualcomm MPU, eliminating sensor drift caused by OS scheduling. 4. Zero-Cost Gorilla Delta-RLE Compression & 100% Privacy: Compresses 24 hours of 1-second telemetry by 99.9% (17.56 MB raw down to 19.48 KB), guaranteeing $0.00/year cloud OpEx while keeping all raw data, voice, and camera strictly on-device.
- Solution summary
AeroSense Edge is an autonomous, decentralized Edge AI air intelligence and source attribution station purpose-built for the Arduino UNO Q. By exploiting the board's heterogeneous dual-compute architecture (STM32U585 MCU for microsecond deterministic sensor DMA reads and Qualcomm QRB2210 Linux MPU for on-device machine learning), the station closes the traditional "interpretation gap" in environmental monitoring. It deploys an INT8 quantized ONNX Gradient Boosting classifier (99.03% CV accuracy, 1.14 ms inference) that distinguishes 6 pollution sources using chemical stoichiometric ratios (PM2.5/PM10, CO/CO2, NO2/VOC), paired with an autoregressive forecaster predicting +1h to +6h diurnal trends (beating persistence by 27.1%). Instead of abstract numbers, AeroSense Edge issues context-aware actionable guidance for schools, joggers, and municipal teams via an embedded zero-cloud Web SCADA interface and a Flutter mobile app. Equipped with 30-day circular SQLite storage, 99.9% Gorilla-style Delta-RLE compression (17.56 MB → 19.48 KB), and LiFePO4 solar off-grid power, the station delivers 100% privacy and $0.00/year cloud OpEx.
- Report (PDF)
- Report (PDF)PDF
- Arduino board used?
Yes
- Edge AI used?
Yes
- AI/ML model used
1. Source Attribution Classifier (source_classifier.onnx, 1.34 MB): scikit-learn GradientBoostingClassifier (120 trees, max_depth=5) exported to ONNX Runtime running on Qualcomm QRB2210 Linux using 9 stoichiometric/sensor features (PM2.5, PM10, CO, CO2, NO2, VOC, PM2.5/PM10 ratio, CO/CO2 ratio, NO2/VOC ratio) + acoustic/rate-of-change fusion to classify 6 trained pollution sources and detect sudden Diwali firecracker spikes. 2. Multi-Horizon Forecaster (timeseries_forecaster.onnx, 4.3 KB): StandardScaler + MLPRegressor (32->16 ReLU) predicting +1h, +2h, +4h, and +6h PM2.5 trends on-device. 3. Edge Impulse TinyML Bundle (aerosense_edge_impulse.eim, ARM64 Linux ELF) for windowed multi-sensor + acoustic inference. 4. Statistical Spike vs. Drift Detector & Smart Notification Filter: Rolling baseline Z-score + dPM2.5/dt slope separation (sudden spike vs. gradual buildup) and 7-day hourly adaptive baseline.
- Sensors used
- Sharp GP2Y1010AU0F Optical Dust Sensor (PM2.5 / PM10 / PM1.0 via A0 analog + D2 IR LED pulse) - MQ-2 Gas/Smoke Sensor (CO & NO2 equivalent on A1, 15-sample median + T/RH compensation + Kalman filter) - MQ-135 Air Quality Gas Sensor (CO2, VOC, NH3 on A2) - DHT22 Digital Temperature & Humidity Sensor (D4) - Acoustic Analog Envelope / MEMS Sound Sensor (A4) (Plus 1.3" SH1106 I2C OLED Display and 4-Channel Relay Actuators on D7-D10 for HEPA purifier, ventilation lockout, and alarm siren)
- Camera used
None.
- Modulino used
None.
- Problem being solved
Protecting an elderly person living alone during sudden indoor air pollution emergencies (such as Diwali firecracker PM2.5/PM10 & combustion gas spikes at night) when they may be asleep or unable to operate appliances. Unlike static-threshold alarms, AeroSense separates sudden spikes from gradual drift/noise, prevents unsafe outdoor-air intake (which pulls firecracker smoke indoors) by locking out ventilation and activating indoor HEPA recirculation instead, sounds a local alert with an "I'm OK / Silence" check-in timer, and escalates to family/caregivers if unacknowledged. (Safety-support concept only — not a medical diagnostic device.)
- AI model training platform
Python scikit-learn + skl2onnx (exporting GradientBoostingClassifier -> source_classifier.onnx and StandardScaler+MLPRegressor -> timeseries_forecaster.onnx for ONNX Runtime), plus Edge Impulse Studio (240 multi-sensor time-series CSV windows compiled into an ARM64 Linux .eim TinyML bundle), cross-validated against synthetic physics benchmarks and public UCI Air Quality / CPCB / WHO reference data.
- Arduino UNO Q features used
- Dual-Compute Architecture: STM32U585 Cortex-M33 MCU for deterministic ADC/pulse sampling (GP2Y1010AU0F dust, MQ-2, MQ-135, DHT22, acoustic envelope), Kalman filtering, I2C 1.3" SH1106 OLED rendering, and 4-channel relay control (D7-D10); Qualcomm QRB2210 Linux MPU for on-device Edge AI inference, SQLite WAL storage, rules engine, and web/API hosting. - Arduino_RouterBridge & arduino-router IPC (127.0.0.1:7500 / serial bridge) for bidirectional telemetry streaming and Edge Core -> OLED synchronization. - Dual-Node Fleet Operation (Indoor Bedroom node + Outdoor Perimeter node) + BLE BTHome v2 & Bluetooth RFCOMM / local Wi-Fi / USB ADB failover + Hardware Watchdog (IWatchdog) & systemd power-cut recovery.
- Deployment method
Fully autonomous on-device edge deployment across two Arduino UNO Q nodes: - MCU firmware compiled & flashed via arduino-cli / OpenOCD SWD to STM32U585. - Python Edge Core + ONNX Runtime AI engine deployed on Qualcomm QRB2210 Debian Linux as an auto-restarting systemd service (aerosense) with SQLite WAL persistence and corruption recovery. - Local web dashboards (SCADA, Executive, Elderly/Citizen portals served on port 8090/8091) and Flutter Android companion app with automatic transport failover (USB ADB, Local Wi-Fi, Tunnel / Firebase P2P relay, and Bluetooth).
- Key innovation
1. Outdoor-vs-Indoor Ventilation Safety Lockout: Cross-compares outdoor perimeter and indoor bedroom UNO Q nodes plus spike dynamics to automatically block fresh-air intake/exhaust (preventing negative pressure from drawing firecracker smoke inside) while engaging HEPA recirculation. 2. Multi-Modal Edge AI Beyond Fixed Thresholds: Fuses ONNX GradientBoosting source attribution, MLP multi-horizon forecasting, acoustic burst + dPM2.5/dt slope & Z-score spike-vs-drift discrimination, and 7-day adaptive hourly baselines. 3. Asleep-Resident Check-In Escalation & Safe Demo: Unacknowledged audible warnings trigger a timed "I'm OK" window before escalating to caregiver webhooks/email/push alerts, paired with a zero-smoke Diwali Firecracker safe demo preset and synchronized hardware OLED display.
- Solution summary
AeroSense Edge is a zero-cloud dual-node environmental safety system built on two Arduino UNO Q boards (STM32U585 MCU + Qualcomm QRB2210 Linux MPU) designed to protect an elderly person living alone during sudden indoor air quality deterioration (e.g., Diwali firecracker emissions at night). The STM32U585 continuously samples optical particulates (GP2Y1010AU0F), combustion gases (MQ-2, MQ-135), temperature/humidity (DHT22), and acoustic impulse levels using 15-sample median and Kalman filtering, streaming 1 Hz frames via Arduino_RouterBridge IPC to the QRB2210 Edge Core. On-device AI runs an ONNX GradientBoosting source classifier, an ONNX MLP multi-horizon PM2.5 forecaster (+1h to +6h), and statistical spike-vs-drift detection (dPM2.5/dt + rolling Z-score) alongside 7-day adaptive hourly baselines. Under the Elderly Care profile, outdoor-vs-indoor differential checks and spike detection automatically lock out fresh-air intake/exhaust (preventing outdoor firecracker smoke from being pulled indoors) while energizing indoor HEPA recirculation (Relay 2) and sounding a local safety warning (Relay 3 + OLED + mobile/web alert). An "I'm OK / Silence" timer monitors for resident acknowledgment; if unresponsive, AeroSense escalates to caregiver email/webhook/push trigger alerts. (Safety-support concept only — not a medical diagnostic device.)
- Report (PDF)
- Report (PDF)PDF