Research on prediction model of mixed gas concentration based on CNN-LSTM network

Abstract

Rapid prediction of concentration in mixed gas is a challenging task in the field of gas sensing. In view of the large error of mixed gas concentration prediction due to the nonlinear response characteristics of sensor array to gas, a prediction model of mixed gas concentration based on Convolutional Neural Network and Long-Short Term Memory is proposed, which has good time series processing ability. The sensor data of carbon monoxide and ethylene are used as the input of this model, RMSE and R2 are used as evaluation indicators. Experimental results show that the accuracy R2 of mixture concentration prediction can reach 0.99 in a short response time of 20 seconds. In addition, RMSE of carbon monoxide and ethylene is 11.4 ppm and 1.6 ppm, respectively. Relative to their maximum presented concentrations, the error ratio is 2.1% and 8%, respectively. Compared with the conventional machine learning algorithms including reservoir-computing and support vector regression (SVR), this method has certain advantages in concentration prediction accuracy and detection time, effectively solves the cross-sensitivity characteristics of MOX sensors, and reduces the measurement delay.