Employment of MQ gas sensors for the classification of Cistus ladanifer essential oils

Abstract

The chemical composition of essential oils (EOs) from has a huge variability throughout the year, impacting the oil quality. Nowadays, EO analytic chemistry techniques, which are expensive and destroy the sample, are utilized to measure the chemical composition. In the paper, we propose a combination of low-cost sensors and machine learning based system. As low-cost sensors, seven gas sensors are combined to obtain up to 36 features. Regarding machine learning, 31 multiclass classification algorithms are applied. Data from sensors were collected for 33 samples of EO from . The generated dataset was split into training and test datasets, with 75 % of the data for training. The datasets were created to ensure a homogeneous chemical composition distribution on both training and test datasets. There were three target chemical compounds: Alpha-pinene and Viridiflorol as individual compounds and Terpenic Hydrocarbons as a group of chemical compounds. The value of the percentage of each targeted compound is converted into a categoric variable with 5 possible values, 1 being the lowest concentration and 5 being the maximum one. The data of the MQ-sensors were included as the input for the models, and each one of the targeted chemical compounds was selected as an output for different models. The input features were ranged using different algorithms for the feature selection process. The results indicate that there is no valid classification model for Viridiflorol, and limited accuracy is achieved for Alpha-pinene. Meanwhile, for Terpenic Hydrocarbons, an accuracy of 91.6 % is achieved. It is important to highlight that these accuracies were attained when a reduced number of features were included, ranging the number of features from 11 to 13. This is the first case in which MQ-based gas sensors, or other metal oxide sensors, are used to correctly determine the concentration of a chemical compounds in a complex matrix formed by dozens of compounds. This system will provide a cheap method to determine the quality of EOs and confirm the benefits of combining low-cost sensors with machine learning.