Abstract
CO2 and SO2 gases are utilized in various industrial applications and are subjects of environmental research. However, these gases are considered toxic and pose dangers at certain concentrations. Therefore, it is crucial to monitor and control the exposure to these gases in the environment to prevent reaching hazardous levels that could endanger both humans and the environment. A non-contact detection and monitoring system is essential to minimize the adverse effects of direct gas exposure. In this research, a non-contact detection system for CO2, SO2, and mixed gases was developed using optical imaging analysis generated by infrared cameras. The images were captured using the FLIR Vue Pro-R infrared camera, with infrared absorbing gas sourced from a 50-watt tungsten lamp. Visual identification of these gases through optical imaging is challenging; however, this study successfully identified these gases using a Convolutional Neural Network (CNN). The CNN architecture used in this study is DenseNet (Densely Connected Convolutional Networks), comprising 169 convolution layers. The CNN model was trained and tested on experimental optical imaging data, categorized into three classes: CO2, SO2, and a mixture of gases. A total of 1030 optical imaging data points were utilized for training. Training was conducted using the AdamW optimization function over 28 epochs. The evaluation of results yielded accuracy, precision, recall, and F1-score metrics. The novelty of this study lies in the successful identification of CO2, SO2, and their mixture by the CNN model with an accuracy of 85%. Precision, recall, and F1-score values are all 0.85. These results indicate that the CNN model effectively distinguishes optical imaging of each gas (CO2, SO2, and their mixture) consistently and accurately. Consequently, it can be concluded that the CNN model performs well in distinguishing between these gases in optical imaging analysis.