AI fire-risk prediction
Classifies incoming readings as safe or fire risk using a trained Random Forest model.
University of Bahrain Senior Project
PyroSense
An AI-based smart fire detection and management system that predicts fire risk from sensor data, reduces false alarms, and guides emergency response across multiple building zones.
Accuracy
99.27%
Model
Random Forest
Inputs
6 sensors
Response
4 levels
Project introduction
Smarter decisions than a simple alarm
Traditional fire alarms depend on fixed thresholds, so harmless changes like steam, dust, or humidity can trigger unnecessary evacuations. PyroSense uses machine learning to analyze multiple environmental readings together and return a probability-based fire risk result. The system turns that result into clear operational actions for building managers.
Features
Built for multi-zone fire monitoring
Probabilistic confidence
Shows class probabilities and confidence instead of only producing a binary alarm.
Multi-zone dashboard
Lets building managers monitor floors, rooms, sensor IDs, exits, and affected zones from one interface.
Automated reports
Captures incident details, sensor readings, timestamps, predictions, and triggered actions for audit use.
How it works
From sensor readings to emergency response
Collect data
PyroSense receives temperature, humidity, pressure, eCO2, raw ethanol, and raw H2 readings.
Run prediction
The FastAPI backend sends the six features to the trained machine learning model.
Score risk
The model returns safe/fire probabilities, confidence, and a classification label.
Trigger action
The dashboard maps the result to Monitor, Warn, Act, or Contain decisions for each zone.
Decision hierarchy
Four response levels
Monitor
Track early abnormal readings while the model still classifies the zone as safe.
Warn
Notify supervisors, inspect the zone, and prepare evacuation if warning probability rises.
Act
Evacuate affected zones, activate suppression support, and notify emergency teams.
Contain
Escalate to a multi-zone incident when fire risk spreads across adjacent zones.
Dashboard
Designed for fast situational awareness
The interface highlights active floors, risk level, sensor readings, affected rooms, confidence scores, evacuation routes, and incident actions in a format that non-technical building managers can understand quickly during an emergency.
Lab
Corridor
Office
Server
Act
Fire confirmed
Evacuate Zone A and isolate HVAC on the affected floor.
Technology
System architecture
Frontend
TypeScript and Vite render the interactive monitoring dashboard, floor map, analytics, and reports.
Backend
Python FastAPI exposes simulation and prediction endpoints for the web dashboard.
Machine learning
Scikit-learn trains Random Forest, Decision Tree, and KNN models, then saves the best model with joblib.
Dataset
The model is trained on a Kaggle fire detection sensor dataset with labeled safe and fire readings.
Evaluation
High accuracy with practical response intelligence
PyroSense achieved 99.27% accuracy on the held-out test set, with near-perfect precision and recall for safe and fire classes. The project also benchmarks PyroSense against traditional threshold-based systems, SVM approaches, CNN fire detection, and commercial alarm platforms.
Model
Accuracy
Fire F1
Random Forest
99.27%
0.99
Decision Tree
99.10%
0.99
KNN
97.14%
0.97
Future work
Next steps toward deployment
Live IoT sensor integration
SMS, email, and mobile push alerts
Mobile monitoring application
Advanced temporal ML models