Agricultural Gas Classification Using Gas Sensor Arrays and PCA-Based Dimensionality Reduction

This project focuses on the development of a gas classification system using a gas sensor array integrated with an Arduino Mega 8260. The sensor array consists of various gas sensors, including Df-NH3, MQ-136, MQ-135, MQ-8, MQ-4, and MQ-2. These sensors capture feature vectors for ten distinct odor (gas composition) classes: air, onion, garlic, lemon, tomato, petrol, gasoline, coffee, orange, and colonia perfume. Each sensor records data points such as temperature, humidity, gas concentration levels, and corresponding voltage outputs over 500 time samples per class. To enhance classification accuracy, various machine learning algorithms, including linear classifiers, ensemble methods, random forest, and decision trees, were initially tested. However, we observed that not all sensors contributed equally to the classification task, likely due to different gas emissions per class. To optimize the sensor array’s performance, Principal Component Analysis (PCA) was employed to reduce the dimensionality of the feature vectors, revealing the most relevant features for classification. Remarkably, ∼94% classification accuracy was achieved by retaining only three key features after dimensionality reduction and ∼100% with two more, demonstrating the efficacy of PCA in identifying the most significant sensors for gas detection. This optimized sensor array could pave the way for more efficient and cost-effective gas classification systems, particularly in agriculture where multiple gas compositions and volatile organic compounds (VOCs) are present, and where air quality and safety are critical.