Environment Explosiveness Level Evaluation Improvement Using Machine Learning

Environment explosiveness level monitoring is the crucial method to prevent emergencies related to combustible gas leakages. In this paper, the results of advanced signal processing for environment integral explosiveness evaluation are presented. The signal processing is based on machine learning techniques application to take into account the information pre-sented in the multidimensional signal retrieved by the temper-ature modulated measurements. The signal measurements were performed for clean air, hydrogen (1 %vol.), methane (1 %vol.), ethylene (1 %vol.), propane (1 %vol.), butane (0.7 %vol.) and n-hexane (0.5 %vol.). To measure the signal, the node prototype with the catalytic gas sensor was used. The data from the prototype was transmitted by wireless channel to a personal computer to be stored and processed. Two models were trained: a linear regression and a neural network. The models were trained using stochastic gradient descent algorithm. The training was performed while the validation error value was decreased. The results for the models were compared with the known method of environment explosiveness level estimation. It was shown that the available method has a low performance for specific gases. The presented models are able to decrease the average error of explosiveness level estimation for the gases in research from 7.8 %LEL down to 0.12 %LEL. Thereat the error level stays almost the same for all gases.