Stochastic flame locating method hybridizing Kalman filter and deep neural network for rapid fire response at construction sites

Rapid flame locating (FL) is a crucial issue to timely response to fire at construction sites. Existing FL methods are not sufficiently accurate to locate flame source. This study presents a stochastic flame locating (SFL) method, hybridizing Kalman filter (KF) and deep neural network (DNN), which identifies the flame location in real-time by minimizing the uncertainty associated with ultraviolet (UV) radiation signals attributed to the momentary, transitory, and/or random nature of flame. SFL complements the limitations existing FL methods that use existing deep learning, which results in high variance, by updating continuously the Kalman gain. Hence, minimizing the mean and covariance of the posterior state at the final updating iteration and improving the convergence and accuracy of the output. It facilitates fine calibration of the coordinates of the estimated flame location by updating the statistics of UV signals using the sliding window method. Furthermore, SFL enables rigorous experiments that measure the sensitivity of performance to the variability of flame magnitude and the existence of obstacles on the path of radiation.