Research on Gas Recognition Based on Gas Sensor Array and Feature Analysis

In this work, a sensor array consisting of four sensors is designed. Detection of any mixture concentration in the range of 0–450 ppm hydrogen, 0–450 ppm ethylene, and 0–450 ppm carbon monoxide were achieved. Filter processing reduces the influence of sensor noise and ensures high recognition accuracy. Then, a total of 48 features such as response value, time feature, and first- and second-order features of sensor resistance were extracted from four sensors in each sample. In this article, the first- and second-order features are creatively proposed, which have a significant impact on the accuracy of the later model. It is found that gradient boosting decision tree (GBDT) has the highest classification accuracy, reaching 100%. Compared to models trained using only the first 180 s of data, gas prediction can be achieved at the initial sensor response stage, and the GBDT model has a preparation rate of up to 99.64%. To study the influence of different features on the sensor, the feature contribution of selected features in GBDT method is described. For the classification tasks, response time, recovery time, and second-order features contribute the most. Finally, the regression designed by GBDT method has a root-mean-square error of 23.52 ppm and an accuracy of 94.77%. This method did not require a lot of calculation and was suitable for running in small-embedded devices.