Volume 11 Issue 3 ( September 2024)

Conventional exhaust emission testing of motor vehicles requires cost, time, and a certified place to obtain reliable results. Meanwhile, artificial intelligence provides an alternative for faster and cheaper exhaust gas detection. This research combines several metal oxide semiconductor (MOS) gas sensors to identify carbon monoxide (CO) compounds and other compounds, such as hydrocarbons (HO) and mononitrogen oxides (NOx), which are commonly found in vehicles. Several classifiers are used to classify the presence of these compounds, such as random forest (RF), decision tree (DT), artificial neural network (ANN), K- nearest neighbors (KNN), support vector machine (SVM), and multiple kernel learning (MKL). In addition, this study also selected the optimal sensors using feature selection and simplified the data dimension using PCA during the classification process. Data samples were obtained from five different motorcycles with different production years. Meanwhile, binary data labels are determined based on emission tests, which refer to regulations concerning the implementation of motor vehicle emission quality standards in Indonesia. The test results indicate that the average classification accuracy of MKL for the two classes was 99.96%, followed by SVM and KNN (99.88%), ANN (99.87%), RF (99.86), and DT (99.46%), which is quite convincing for the early detection of emission testing.

Keywords: Classification; Emission testing; Gas sensors array; Machine learning; Multiple Kernel Learning