Advancing Electronic Noses With Transformers: Real-Time Classification of Hazardous Odors and Food Freshness

Real-time detection of gas leaks and food spoilage poses significant challenges, particularly in settings where human olfactory perception is unreliable. This paper introduces Artificial Sensing, a novel AI-powered electronic nose (E-nose) system developed to accurately classify odors and evaluate food freshness in indoor environments. The system combines MQ-series sensors and the Grove Multichannel Gas Sensor v2, integrated with a Seeed Studio Wio Terminal for edge-level processing. To effectively handle the variability and complexity of time-series sensor data, we propose a Time-Series Transformer model enhanced with Multi-Scale Adaptive Convolution (MAC) and multi-head self-attention, enabling robust feature extraction across temporal scales. Additionally, a multimodal fusion strategy incorporating ambient temperature and humidity improves model resilience. A comprehensive dataset comprising 4040 gas leak samples and 4080 food spoilage samples (milk and chicken) was collected under varied conditions. The proposed model achieved F1-scores of 99.3% for gas leak detection and 94.8% for spoilage classification, surpassing traditional classifiers (SVM, Random Forest, ANN) and baseline ViT models by 7%-12%. Deployed on the Edge Impulse platform, the system demonstrated real-time responsiveness (120 ms latency) and low power consumption. These findings highlight the promise of Transformer-based architectures for edge-level environmental monitoring and food safety applications.