Identification of gas mixtures via sensor array combining with neural networks.

• A sensor array comprising four gas sensors has been utilized to detect 11 types of mixtures of NO 2 and CO, whose concentration varies from 0 to 50ppm. The range of mixed gas concentration is wide enough, and the theoretical LOD could reach dozens of ppb levels. • Features were extracted by principal component analysis (PCA), and then C-means clustering and back propagation neural network (BPNN) were employed to identify gases. Also, we have verified the optimized effect of genetic algorithm (GA) for BPNN. Through transforming response curves into gray image, CNN could be used to directly recognize various gases. • To investigate which feature of input sample is most influential to BPNN model, a random variable substitution method has been proposed. • Experiments for sensing 6 kinds of CO and NO 2 mixtures under 4 different relatively humidity (25%, 33%, 50%, 75%) have been implemented to evaluate the effect of environmental humidity on our sensor array. The results demonstrated our method could eliminate the effects of humidity. • Even though this work stressed on the identification of mixtures of NO 2 and CO, the approach presented here is generic, and could in principle, be extended to other relevant gas species. In this work, a sensor array comprised four sensors has been employed to detect 11 types of mixtures of nitrogen dioxide (NO 2) and carbon monoxide (CO), with concentration varying from 0 to 50 ppm. To reduce the effect of sensor noise and ensure high recognition accuracy, average resistance over a period of time was introduced. Then, 12 features including response value, response time and recovery time were extracted from each sample. After that, C-means clustering and back propagation neural network (BPNN) were performed to identify various gases, with classification accuracy of 94.55 % and 100 %, respectively. Genetic algorithm (GA) was also employed to further improve BPNN's performance. Moreover, a random variable substitution method has been introduced to study which feature of the input sample influence the BPNN model most. Through gray processing, dynamic curves have been transformed into gray images, from which convolutional neural network (CNN) was introduced to automatically extract high-level features, and an identification accuracy of 100 % has been realized. Finally, experiments for sensing gas mixtures of CO and NO 2 under various humidity levels have been done to test the impact of humidity on the sensor array. The results demonstrated the proposed method could eliminate the effects of humidity. [ABSTRACT FROM AUTHOR]