APPROXIMATION OF FIRE DYNAMIC CHARACTERISTICS BY A NEURAL NETWORK FOR THE DEVELOPMENT OF INTELLIGENT FIRE DETECTORS WITH SMOKE AND HEAT SENSORS

Abstract

Introduction. At the early stage of fire development, the phenomena and products formed while the combustion of different materials is different. However, there are also common features such as heat generation, smoke generation, radiation generation, etc. These combustion products are called fire signs, also known as fire dynamic characteristics. Therefore, the characteristic fire dynamic characteristics are temperature change, smoke concentration change and carbon
monoxide, and wavelength changes of infrared and ultraviolet radiation. Fire detection involves monitoring these fire parameters, which are random and uncertain and difficult to characterise due to statistical characteristics.
These dynamic characteristics are used for research and development of algorithms for the operation of fire detectors, which are built using theories of fuzzy logic, and neural networks on fuzzy neural networks.
Purpose. The purpose of the paper is to approximate the fire dynamic characteristics by a neural network, namely, the dependence curves of the average volumetric temperature in the room on time (temperature regimes of fire
development) and the dependence curves of smoke per unit length on time. These dependencies are necessary for the development and research of algorithms for the operation of intelligent multi-sensor fire detectors with heat and smoke
sensors based on fuzzy logic and neural networks.
Results. Today, from a practical point of view, multi-sensor fire detectors with heat and smoke sensors, which analyse temperature changes and smoke are most commonly used. Therefore, this paper approximates the fire dynamic
characteristics. With the help of computer simulation in the Fire Dynamics Simulator software, which works on the PyroSim interface platform, fire temperature regimes and the dependence of smoke per unit length on time are simulated.
Research conducted by scientists proves that the relative error between simulated data and experimental data does not exceed 28%. Therefore, the resulting curves can be used for further research. The resulting curves are approximated using
a neural network. The neural network model was built and trained in the Neural Network Start GUI package of the MATLAB 2020a software. After setting percentages, a neural network architecture is chosen to generate data for training,
validation, and testing. To achieve the best result of the approximation of the dependences of the curves in this study, the number of neurons of the hidden layer was determined during the training of the neural network. The use of non-linear
activation functions allows you to configure the neural network to implement non-linear connections between input and output. Three learning algorithms were used to train the neural network, namely: Levenberg-Marquardt, Bayesian
Regularization, and Scaled Conjugate Gradient.
Conclusions. In the Fіre Dynamіcs Sіmulator software, the fire dynamic characteristics in the office room, the administrative room and the plywood production premises are simulated. In the Neural Network Start GUI package of the MATLAB 2020a software, these dynamic characteristics are approximated using a neural network. In the process of training the developed neural network, studies have shown that a significant increase in the number of neurons of the hidden layer does not lead to improved results, but only increases the time of training the network. When the number of hidden layer neurons is 15, and 20, the mean square error and regression values are almost the same. To approximate the
fire dynamic characteristics, the best result of neural network training is provided by the Bayesian Regularization algorithm. Then the root means the square error is the smallest. Research shows that the neural network reproduces these
curves with sufficient accuracy. Thus, when approximating the curve of the dependence of the average volume temperature on time, the root means the square error of learning is 278.599, and the regression is 0.9673.
Thus, when approximating the curve of the dependence of the average volume temperature on time, the root means the square error of training is 278.599, and the regression is 0.9673. When approximating the smoke per unit length versus
the time curve, the root means the square error of training is 3.4714, and the regression is 0.9957. Approximated curves of the dynamic characteristics of a fire by a neural network can be used as input data in the development and research of
algorithms for the operation of fire detectors with heat and smoke sensors based on fuzzy logic or a neural network. With these approximated curves, the neural network of the fire detector can be trained to distinguish the signs of fire from
deceptive phenomena that are not related to fire.