The recognition of selected burning liquids by convolutional neural networks under laboratory conditions

Abstract

This paper deals with the recognition of selected burning liquids by convolutional neural networks (CNNs). Three CNNs (AlexNet, GoogLeNet and ResNet-50) were trained, validated and tested (in the MATLAB 2020b software) for the recognition of selected liquids (ethanol, propanol and pentane) using photographs of the flames they produce. For training, validation and test photographs of the liquids under investigation burning in a 106-mm-diameter vessel were used. The accuracy of all the CNNs under investigation during the tests was above 99%. In addition the trained CNNs were tested using photographs of the flames generated by the liquids under investigation burning in a vessel with a diameter of 75 mm. The accuracy of the trained CNNs in this additional test ranged from 37 to 42% (GoogLeNet) through 62-73% (ResNet-50) up to 51-80% (AlexNet) - the results varied dependent upon the relative size of the flame in the photograph under analysis (in most cases an increase in the relative size caused an increase in accuracy). The accuracy of the AlexNet can be improved from 80% to almost 96% using an algorithm. The principle of the algorithm is the analysis of 10 photographs of the same liquid in the same vessel (taken over a few seconds) followed by the recognition based on an identical classification for at least 6 out of 10 photographs. An accuracy of 96% is sufficient for the rapid recognition of burning liquids in practical applications.

Supplementary information: The online version contains supplementary material available at 10.1007/s10973-021-10903-2.

Keywords: Deep learning; Fire detection; Flame recognition; Flammable liquids; Identification of burning substances; Neural networks.