Development of a Neural Model of a Semiconductor Gas Sensor

Purpose of research: Development of a neural model of a semiconductor gas sensor in order to generate data for training an information-processing device of gas analyzers based on artificial neural networks (ANN). Search and optimization of cleaning data composition and volume. The neural model of the sensor should take into account the influence of those factors on the signal, the fluctuations of which make the maximum contribution to the measurement errors. Testing of the model based on semiconductor carbon monoxide and hydrogen sensors.Methods. Methods of computer modeling, numerical methods, theory of neural networks. To compare the simulation results and the responses of real sensors, the relative error and standard deviation were determined.Results. Studies of various structures of the neural model of a semiconductor sensor have been carried out, the structure of a multilayer neural network of direct propagation for two semiconductor carbon monoxide and hydrogen sensors has been selected, modeling errors have been estimated, recommendations have been given for choosing the optimal structure and the amount of training data.Conclusion. Neural models of semiconductor carbon monoxide and hydrogen sensors have been obtained, conclusions have been drawn about the possibility of using this ANN structure in solving typical problems. Based on the analysis of the errors obtained, the effectiveness of using neural models of sensors to generate training data has been shown. The maximum relative error of modeling the TGS2442 semiconductor carbon monoxide sensor did not exceed 5% for the main characteristic and 2% for additional ones. The maximum relative error of modeling of the TGS2442 semiconductor hydrogen sensor did not exceed 3% for the main characteristic and 1% for additional ones.