Robust Odor Detection in Electronic Nose Using Transfer-Learning Powered Scentformer Model.

Area of Science:

• Artificial Intelligence
• Sensory Systems Engineering
• Computational Chemistry

Background:

• Electronic noses (E-noses) mimic human olfaction for odor detection.
• Current E-noses face limitations in detection range and generalizability.
• Developing advanced E-nose technology is crucial for diverse applications.

Purpose of the Study:

• To present a novel deep learning-based E-nose, Scentformer.
• To overcome limitations of existing E-nose systems.
• To enhance odor detection accuracy and adaptability.

Main Methods:

• Developed a novel E-nose system utilizing the Scentformer deep learning architecture.
• Implemented a self-adaptive data down-sampling method for efficient processing.
• Employed Shapley Additive exPlanations for model interpretability.
• Utilized transfer learning for rapid adaptation to new odors and gases.

Main Results:

• Achieved 99.94% classification accuracy for 55 natural odors.
• Demonstrated robust performance with minimal data (1‰) for new odor detection using transfer learning (99.14% accuracy).
• Provided quantitative interpretation of E-nose performance via Shapley Additive exPlanations.

Conclusions:

• Scentformer offers a significant advancement in E-nose technology.
• The system exhibits high accuracy, broad detection range, and excellent generalizability.
• Transfer learning significantly reduces data requirements for adapting to new odor detection tasks.