Rapid forecasting of hydrogen concentration based on a multilayer CNN-LSTM network

Abstract Gas sensors with rapid response are desirable in many safety applications. Reducing the response time of gas sensors is a challenging task. Computing a part of the initial temporal signals of gas sensors based on neural networks is an effective and powerful method for forecasting sensors’ output. To rapidly and robust forecasting hydrogen concentration, a sensor array is composed of a temperature and humidity sensor, and two hydrogen sensors. A neural network combined with convolutional neural networks and long-short-term memory networks is proposed to fuse temporal signals of the sensor array to forecast hydrogen concentrations. The structure of the neural network is optimized by increasing its depth. For the optimal neural network, the lowest mean absolute percent error is about 12.8% by computing initial 30 s of transient signals within 300–400 s response curves, the predicted mean absolute error is 1158 ppm in the testing range of 18 000 ppm. When the time span of initial transient signals of the sensor array increase to 150 s for the computing, the mean absolute percent error decreases to 5.7%. This study verifies the potential and effectiveness of the neural network for concentration forecasting by computing the temporal signals of the sensors.