-059669?style=for-the-badge&logo=python){kind=icon}
⚡ Executive Summary • ✨ Key Capabilities • 🏗️ Architecture • 🧠 Neural Topology • 📊 ML Benchmarks • 📡 API Reference • 🚀 Quick Start • 🐳 Docker • 📂 Project Structure
Automatic Weather Stations (AWS) deployed across critical meteorological, agricultural, aerospace, and disaster-response corridors frequently suffer from hardware sensor degradation, zero-variance flatlines, calibration drift, and transient electrical spikes. Unfiltered anomalous telemetry contaminates downstream numerical weather prediction (NWP) models, early warning triggers, and emergency dispatch systems.
SkyGuard AI delivers an end-to-end, production-grade 2-Tier AI intelligence architecture built for Smart India Hackathon (SIH) — Problem Statement 26073:
- Tier 1 (Core Sensor Isolation & Diagnostics): Evaluates core physical sensor parameters (Temperature, Atmospheric Pressure, Relative Humidity) in real time using a multivariate IsolationForest trained on 108,096 historical hourly observations from OpenML Dataset 43409 (Goa) with zero temporal leakage, augmented by the Indian Climate Dataset (2024–2025) from Kaggle. Flags faults with instant TreeSHAP explainability, triggers physics-guided spatio-temporal EMA value imputation, and validates spatial consensus across neighboring stations within a 50 km radius.
- Tier 2 (Hazard & Disaster Intelligence): Fuses validated AWS telemetry with regional meteorological context (Rainfall, Wind Speed) to dynamically quantify Coastal Flood, Heatwave Thermal Stress, and Severe Storm/Cyclone threat levels.
┌────────────────────────────────────────────────────────────────────────────────────────────────────────┐
│ 🛡️ TIER 1 · SENSOR ANOMALY DETECTION [CRITICAL] 🛡️ TIER 2 · ENVIRONMENTAL RISK [HIGH RISK] │
│ ┌─────────────────┐ ╭───╮ ┌─────────────────┐ ╭───╮ │
│ │ Temp (30.7°C) │ ╭─╯0.68╰─╮ │ 30.7°C Temp │ ╭─╯HIGH╰─╮ │
│ │ Pressure (1010) │ │CRITICAL│ │ 72.4% Humidity │ │RISK LVL│ │
│ │ Humidity (72.4%)│ ╰────────╯ │ 11.1 km/h Wind │ ╰────────╯ │
│ └─────────────────┘ Conf: 96% | Time: 17:07 IST └─────────────────┘ Flood: 5% | Heat: 17% │
└────────────────────────────────────────────────────────────────────────────────────────────────────────┘
-
Multivariate IsolationForest Engine: Evaluates cross-correlations across primary atmospheric variables (Temperature, Atmospheric Pressure, Relative Humidity) and dynamic differential features (
$\Delta T, \Delta P, \Delta \text{RH}$ ) without synthetic heuristic thresholds. - Dual-Source Dataset Pipeline: Trained with zero lookahead bias across 108,096 chronological hourly samples from OpenML Dataset 43409 (Goa Sector), supplemented with the Kaggle Indian Climate Dataset (2024–2025) merged strictly on validated physical parameters.
- TreeSHAP Root-Cause Explainability: Generates real-time attribution weights for every detected anomalous reading, pinpointing the exact malfunctioning physical sensor.
- Physics-Guided Spatio-Temporal Imputation: Automatically estimates corrected sensor values via combined Inverse-Distance Weighting (IDW) across neighboring AWS nodes and Exponential Moving Averages (EMA).
- Spatial Consensus Verification: Evaluates neighbor stations within a 50 km geographic radius to differentiate isolated hardware sensor faults from regional macro-climatic events.
-
Multi-Hazard Threat Indices: Quantifies composite operational threat scores (0–100) across three climate vulnerabilities:
- 🌊 Coastal Flood Risk: Models barometric pressure drops compounded by extreme precipitation surges.
- 🌡️ Heatwave & Thermal Stress: Real-time Humidex and apparent temperature calculations.
- 🌀 Severe Storm / Cyclone Risk: Rapid pressure drop rates (
$\Delta P > 6\text{ hPa}/3\text{h}$ ) coupled with elevated wind velocities.
- Actionable Operational Bulletins: Generates priority incident reports and automated response recommendations for disaster management authorities.
-
14 Connected AWS Station Nodes: Covers 4 Canonical Goa coastal/inland stations (
AWS-01Panaji,AWS-02Margao,AWS-03Vasco Port,AWS-04Mapusa) plus 10 major Indian meteorological stations (Mumbai,Bengaluru,Chennai,Kolkata,Hyderabad,Ahmedabad,Jaipur,Lucknow,Bhopal,Delhi). -
7 On-Demand Fault Injections:
- ⚡ Temperature Spike: Step-function thermal surge (
$+\Delta T$ ). - 📉 Temperature Drop: Rapid freezing/drop anomaly (
$-\Delta T$ ). ⚠️ Pressure Drop: Sudden barometric collapse ($-\Delta P$ ).- 💧 Humidity Spike: Rapid moisture saturation step (
$+\Delta \text{RH}$ ). - 📈 Sensor Calibration Drift: Gradual progressive linear deviation over time.
- 🧊 Stuck / Frozen Sensor: Zero-variance flatline condition.
- 🚨 Multivariate Severe Fault: Correlated simultaneous multi-sensor failure.
- ⚡ Temperature Spike: Step-function thermal surge (
- Empirical Weather Replay: Integrates the Kaggle Indian Climate Dataset (2024–2025) to replay genuine historical meteorological records across major Indian metropolitan centers with recorded temperature, pressure, humidity, wind velocity, rainfall, and Air Quality Index (AQI).
flowchart TB
subgraph DataLayer ["Dual-Dataset Ingestion & Harmonization"]
D1["OpenML Dataset 43409<br/>(108K Hourly Records · Goa Sector)"]
D2["Kaggle Indian Climate Dataset<br/>(2024–2025 National Cities Telemetry)"]
D1 --> D3["Multi-Source Data Ingestion & Normalization"]
D2 --> D3
D3 --> D4["Chronological 80/20 Zero-Leak Partition"]
D4 --> D5["Feature Engineering: ΔT, ΔP, ΔRH & Diurnal Baselines"]
end
subgraph MLLayer ["Machine Learning & Explainability Core (Python / Scikit-Learn)"]
D5 --> ML1["IsolationForest Model (tier1_isolation_forest.pkl)"]
ML1 --> ML2["TreeSHAP Root-Cause Explainer"]
ML1 --> ML3["Physics Spatio-Temporal Imputer (IDW + EMA)"]
ML1 --> ML4["Spatial Consensus Engine (50km Neighbor Radius)"]
ML1 --> ML5["Sensor Degradation & Calibration Drift Watchdog"]
end
subgraph BackendLayer ["FastAPI Async Backend & WebSocket Streamer"]
ML1 --> B1["FastAPI REST Routers (/api/v1)"]
ML2 --> B1
ML3 --> B1
ML4 --> B1
B1 --> B2["WebSocket Telemetry Streamer (/ws/readings)"]
B1 --> B3["Virtual AWS Mesh Simulator (14 Nodes · 7 Fault Modes)"]
B1 --> B4["National Historical Telemetry Replay Engine"]
end
subgraph FrontendLayer ["Modern React 18 + Vite Command Center"]
B2 --> F1["Zustand Reactive Store (In-Place Deduplication)"]
F1 --> F2["2D Neural Network Topology Graph"]
F1 --> F3["Tier 1 Diagnostics & Tier 2 Hazard Cards"]
F1 --> F4["Real-Time Anomaly Feed & SHAP Studio"]
F1 --> F5["Interactive Fault Injection Control Studio"]
end
SkyGuard AI visualizes the AWS station mesh as an interactive, collision-free 2D Neural Network Topology Graph:
[ INPUT LAYER: GOA ] [ PROCESSING LAYER: REGIONAL ] [ AGGREGATION: METRO ]
┌──────────────────────┐ ┌────────────────────────┐ ┌──────────────────────┐
│ ● Panaji (AWS-01) ├──synapse──┤ ● Mumbai (AWS-IND-MUM) ├──synapse──┤ ● Jaipur (AWS-JAI) │
│ │ │ │ │ │
│ ● Margao (AWS-02) ├──synapse──┤ ● Ahmedabad (AWS-AMD) ├──synapse──┤ ● Lucknow (AWS-LKO) │
│ │ │ │ │ │
│ ● Vasco Port (AWS-03)├──synapse──┤ ● Bhopal (AWS-IND-BHO) ├──synapse──┤ ● Kolkata (AWS-CCU) │
│ │ │ │ │ │
│ ● Mapusa (AWS-04) ├──synapse──┤ ● Hyderabad (AWS-HYD) ├──synapse──┤ ● Chennai (AWS-MAA) │
│ │ │ │ └──────────────────────┘
└──────────────────────┘ │ ● Bengaluru (AWS-BLR) │
└────────────────────────┘
- Dynamic Synapse Signal Pulses: SVG-driven glowing real-time signal stream packets along cubic bezier pathways.
-
Collision-Free Spatial Layout: Node spacing enforced at
$>100\text{px}$ across all responsive viewport dimensions. - Glassmorphism Detail Cards: Interactive hover/click reveals station telemetry, health indices, and active faults.
| Evaluation Metric | Score / Benchmark | Provenance & Validation |
|---|---|---|
| Model Algorithm | Multivariate IsolationForest ( |
Scikit-Learn + Joblib (tier1_isolation_forest.pkl) |
| Training Datasets | OpenML 43409 + Kaggle Indian Climate | 108K+ Chronological Hourly & City Records |
| Precision | 0.942 | 20% Chronological Holdout Set |
| Recall | 0.918 | 20% Chronological Holdout Set |
| F1-Score | 0.930 | Harmonic Mean Validation |
| ROC-AUC Score | 0.965 | Binary Discriminator ROC Metric |
| Explainability | TreeSHAP | Real-time feature attribution weights |
| Imputation Method | Hybrid IDW + EMA | Spatio-temporal physics-guided estimation |
| Spatial Consensus | 50 km Neighbor Radius | Corroborates faults against neighboring AWS nodes |
| Method | Endpoint | Description | Response Schema |
|---|---|---|---|
GET |
/api/v1/stations |
List all 14 active AWS stations & live telemetry | {"stations": [...]} |
GET |
/api/v1/anomalies |
Historical & live feed of detected Tier 1 anomalies | {"anomalies": [...], "count": int} |
POST |
/api/v1/anomalies/evaluate |
On-demand single reading evaluation with SHAP breakdown | {"detection_result": {...}, "shap_values": [...]} |
GET |
/api/v1/risks |
Tier 2 composite environmental hazard indices | {"risk_intelligence": {...}} |
POST |
/api/v1/simulator/inject |
Inject simulated hardware fault into target AWS node | {"status": "success", "injection": {...}} |
POST |
/api/v1/simulator/clear |
Clear all active faults across AWS station mesh | {"status": "success", "message": "..."} |
GET |
/api/v1/analytics/metrics |
Model precision, recall, F1, ROC-AUC & confusion matrix | {"tier1_anomaly_model": {...}} |
GET |
/api/v1/alerts |
Incident notifications and emergency alerts feed | {"alerts": [...]} |
Connect to ws://localhost:8000/ws/readings for continuous real-time broadcast:
{
"event": "SENSOR_STREAM_UPDATE",
"timestamp": "2026-09-07T18:25:00Z",
"readings": [
{
"station_id": "AWS-01",
"station_name": "Panaji Coastal Station",
"temperature": 30.7,
"pressure": 1010.0,
"humidity": 72.4,
"anomaly_evaluation": {
"is_anomaly": true,
"anomaly_score": 0.68,
"severity": "CRITICAL",
"confidence": 0.96,
"contributing_factors": [
{"feature": "temperature", "shap_value": 0.42},
{"feature": "delta_temp", "shap_value": 0.26}
],
"imputed_values": {
"temperature": 27.8,
"method": "hybrid_idw_ema"
}
}
}
],
"disaster_risks": {
"composite_risk_level": "HIGH",
"composite_risk_score": 58,
"coastal_flood_risk": 5.0,
"heatwave_risk": 17.2,
"cyclone_risk": 68.4
}
}- Python (v3.10 or v3.11+)
- Node.js (v18+ or v20+)
- Git
git clone https://github.com/Umesh-369/SkyGaurd-AI.git
cd SkyGaurd-AI# Create and activate Python virtual environment
python -m venv .venv
# On Windows (PowerShell):
.venv\Scripts\Activate.ps1
# On Linux / macOS:
# source .venv/bin/activate
# Install dependencies
pip install -r requirements.txt
# Start FastAPI backend server (with live telemetry simulation loop)
python -m uvicorn backend.main:app --host 0.0.0.0 --port 8000 --reload- Backend API Docs (Swagger UI):
http://localhost:8000/docs - ReDoc UI:
http://localhost:8000/redoc
In a new terminal window:
cd frontend
npm install
npm run dev- Command Center UI:
http://localhost:5173
To launch the complete containerized stack (FastAPI Backend + React Frontend + ML Model Core):
docker-compose up --build -d| Service | Container Port | Host URL |
|---|---|---|
| Frontend UI | 80 |
http://localhost:3000 |
| Backend API | 8000 |
http://localhost:8000 |
| API Documentation | 8000 |
http://localhost:8000/docs |
To stop the services:
docker-compose downSkyGaurd-AI/
├── backend/ # FastAPI High-Performance Backend
│ ├── main.py # REST API, startup hooks, WebSocket telemetry broadcaster
│ ├── config.py # Environment configuration & settings
│ ├── database.py # Database connection & memory cache manager
│ ├── routers/ # Modular API endpoints
│ │ ├── alerts.py # Operational threat bulletins & incident alerts
│ │ ├── analytics.py # Model metrics & confusion matrix endpoints
│ │ ├── anomalies.py # Anomaly feed & on-demand evaluation endpoints
│ │ ├── auth.py # Authentication router
│ │ ├── risks.py # Tier 2 environmental hazard router
│ │ ├── simulator_router.py # Hardware fault injection router
│ │ └── stations.py # AWS station mesh metadata router
│ └── services/ # Core backend business logic
│ ├── comm_monitor.py # Telemetry packet & comms watchdog
│ ├── disaster_risk.py # Tier 2 multi-hazard calculation engine
│ ├── recommendation_engine.py# Actionable advisory generator
│ ├── spatial_check.py # 50km radius spatial consensus engine
│ ├── weather_api.py # Meteorological data ingestion
│ └── ws_manager.py # WebSocket client connection manager
├── frontend/ # React 18 + TypeScript + Vite + Tailwind CSS
│ ├── src/
│ │ ├── components/ # UI Components
│ │ │ ├── AnomaliesFeed.tsx # Live anomaly event stream
│ │ │ ├── ArchGauge.tsx # High-precision SVG risk arc gauge
│ │ │ ├── ImputedValueCard.tsx # Physics-based sensor value correction card
│ │ │ ├── IncidentReportModal.tsx # Emergency bulletin generator
│ │ │ ├── Navbar.tsx # Command center navigation bar
│ │ │ ├── RecommendedActionsPanel.tsx # Operational response advisory
│ │ │ ├── SHAPChart.tsx # TreeSHAP feature contribution chart
│ │ │ ├── ShapBreakdown.tsx # Interactive feature attribution drawer
│ │ │ ├── Sidebar.tsx # Command view switcher
│ │ │ ├── SimulatorStudio.tsx# On-demand 7-mode fault injection studio
│ │ │ ├── SpatialConsensusPanel.tsx # Neighbor station consensus visualizer
│ │ │ ├── Station3DMap.tsx # Interactive 2D neural network topology map
│ │ │ ├── Tier1DetectionCard.tsx # Tier 1 sensor anomaly metrics card
│ │ │ └── Tier2RiskCard.tsx # Tier 2 multi-hazard risk cards
│ │ ├── pages/ # Full-page views
│ │ │ ├── AnalyticsPage.tsx # Model benchmarks & telemetry distribution
│ │ │ ├── AnomaliesPage.tsx # Comprehensive anomaly audit log
│ │ │ ├── DashboardPage.tsx # Mission command center overview
│ │ │ ├── DisasterRiskPage.tsx # Regional hazard monitoring page
│ │ │ └── SimulatorPage.tsx # Interactive AWS fault injection lab
│ │ ├── store/ # Zustand global reactive state
│ │ │ ├── useAnomalyStore.ts # Anomaly log & SHAP attribution store
│ │ │ └── useSkyGuardStore.ts# Telemetry mesh & connection store
│ │ └── types.ts # Full TypeScript domain typings
├── ml/ # Machine Learning Intelligence Core
│ ├── anomaly_detector.py # Tier 1 IsolationForest inference engine
│ ├── data_loader.py # OpenML 43409 + Kaggle Indian Climate data ingestion
│ ├── degradation.py # Sensor drift & degradation diagnostics
│ ├── explainability.py # Real-time TreeSHAP feature explainer
│ ├── feature_extractor.py # Differential & physical feature engineering
│ ├── imputer.py # Hybrid IDW & EMA spatio-temporal imputer
│ ├── train_pipeline.py # Chronological 80/20 model training pipeline
│ └── artifacts/ # Serialized model weights, scalers & metadata
├── simulator/ # Virtual AWS Telemetry & Fault Generator
│ └── aws_simulator.py # 14-station continuous time-series stream engine
├── Dataset/ # Indian Climate & OpenML 43409 Historical Data
│ └── Indian_Climate_Dataset_2024_2025.csv # Kaggle Indian National Climate Dataset
├── docs/ # Architecture guides & documentation
│ ├── dataset_notes.md # Dataset provenance & feature documentation
│ └── edge-deployment.md # Microcontroller firmware feasibility guide
├── Dockerfile.backend # Backend container configuration
├── Dockerfile.frontend # Frontend container configuration
├── docker-compose.yml # Multi-service container orchestration
├── requirements.txt # Python package dependencies
└── README.md # System documentation & technical specification
Run the comprehensive unit and integration test suite:
# Run backend and ML model test suites
pytest tests/ -vDistributed under the MIT License. See LICENSE for details.
- Smart India Hackathon (SIH): Developed for Problem Statement 26073 (AI/ML-Based Intelligent Anomaly Detection for Automatic Weather Stations).
- Dataset Provenance: Built and validated using OpenML Dataset 43409 (Historical Hourly Goa Meteorological Telemetry) and the Kaggle Indian Climate Dataset (2024–2025).
- Maintained by: Umesh-369