Enhancing Building Safety Through Machine Learning And Deep Learning Based Smoke Detection

Abstract

Traditional fire and smoke detection systems installed in buildings rely on single-sensor threshold-based
triggers, which produce high false-positive rates from cooking fumes, steam, construction dust, and
humidity fluctuations. These systems also lack spatial awareness of fire origin and fail to detect slowburning
smouldering fires at early stages. This paper proposes a dual-module intelligent detection
system. Module 1 trains and evaluates seven classical ML classifiers—Random Forest, Gradient
Boosting, AdaBoost, Logistic Regression, SVM, Decision Tree, and KNN—on 62,630 real-world IoT
sensor readings comprising 13 environmental sensor features. Module 2 employs a MobileNetV2 CNN
for binary fire/no-fire image classification achieving 96.98% validation accuracy through a two-phase
transfer learning strategy, and YOLOv8 for spatial bounding-box localisation of fire and smoke regions.
Both modules are integrated into a Django 5.2 web application with role-based access. All seven ML
classifiers achieved AUC-ROC scores above 0.999. The full system is live at https://building-safetysmoke-
detection-production.up.railway.app.