Early Fire Detection: A New Indoor Laboratory Dataset and Data Distribution Analysis

Abstract

Fire alarm systems are typically equipped with various sensors such as heat, smoke, and gas detectors. These provide fire alerts and notifications of emergency exits when a fire has been detected. However, such systems do not give early warning in order to allow appropriate action to be taken when an alarm is first triggered, as the fire may have already caused severe damage. This paper analyzes a new dataset gathered from controlled realistic fire experiments conducted in an indoor laboratory environment. The experiments were conducted in a controlled manner by triggering the source of fire using electrical devices and charcoal on paperboard, cardboard or clothing. Important data such as humidity, temperature, MQ139, Total Volatile Organic Compounds (TVOC) and eCO2 were collected using sensor devices. These datasets will be extremely valuable to researchers in the machine learning and data science communities interested in pursuing novel advanced statistical and machine learning techniques and methods for developing early fire detection systems. The analysis of the collected data demonstrates the possibility of using eCO2 and TVOC reading levels for early detection of smoldering fires. The experimental setup was based on Low-Power Wireless Area Networks (LPWAN), which can be used to reliably deliver fire-related data over long ranges without depending on the status of a cellular or WiFi Network.