Detecting a sign of severe fire events by image processing

Abstract

Fire whirls, also known as fire tornadoes, represent an extraordinarily severe fire phenomenon, exemplified by the catastrophic events shortly after the great Kanto earthquake in 1923. This research aims to propose an early-warning analysis employing image processing techniques for predicting before the fire-whirl formation. Scaled-down experiments were conducted using two speed-adjustable fans to control the wind movement and generate fire whirls. Through image processing, a set of input flame images is transformed into the evolution of flame height as output data, which serves as the basis of the signal processing. The difference between whether fire whirls occur or not can be detected from the flickering of the flame-height signal. Without delivering external winds, minor changes in noise components are observed at subsequent times, showing no signs of fire whirls. On the other hand, the noise component preceding the fire whirl occurrence highlights a significant increase in standard deviation and autocorrelation, attributing to a slower recovery rate from a perturbed state near a transition to a fire whirl. In this paper, a dynamical marker is constructed as a composite metric of smoothed flame height, standard deviation, and autocorrelation coefficient at lag 1 of the noise component, showing upward trends prior to fire whirl formation. The effectiveness of the dynamical marker as a warning sign for predicting fire-whirl occurrences is validated through experiments of three different wind speeds with the alarm threshold of +3σ to mitigate an unnecessary false detection.

Keywords

• fire whirl
• early-warning signal
• image processing
• noise characterization
• dynamical marker