Electronic Nose-Based a Time-Sensor Feature Classification Network for Adzuki Bean Origin Traceability.

Area of Science:

• Agricultural Science
• Analytical Chemistry
• Machine Learning

Background:

• Origin fraud in agricultural products like adzuki beans poses a significant threat to regional brands and consumer trust.
• Rapid and non-destructive methods are needed to verify the geographical origin of adzuki beans.
• Existing methods may lack the speed, accuracy, or non-destructive nature required for practical application.

Purpose of the Study:

• To develop an efficient and reliable technique for identifying the geographical origin of adzuki beans.
• To integrate an electronic nose (e-nose) with a specialized deep learning network for gas pattern recognition.
• To combat adzuki bean origin fraud and protect regional brand integrity.

Main Methods:

• Collected gas data from adzuki bean samples from six geographical origins using a PEN3 e-nose system.
• Developed a time-sensor feature attention module (TSFAM) to extract temporal dynamics and sensor cross-sensitivity.
• Designed a lightweight time-sensor feature classification network (TSFC-Net) based on TSFAM for origin classification.

Main Results:

• The TSFC-Net achieved high accuracy (98.52%) and F1-score (98.33%) in classifying adzuki bean origins.
• The network demonstrated lightweight characteristics with low parameter count (0.0152 M) and computational complexity (3.2859 M).
• Ablation studies confirmed the effectiveness and rationality of the TSFC-Net design, outperforming other gas classification methods.

Conclusions:

• The proposed e-nose and TSFC-Net system offers an efficient and reliable method for tracing the geographical origin of adzuki beans.
• This technique provides a robust solution for combating origin fraud and safeguarding regional agricultural product brands.
• The study highlights the potential of integrating advanced sensor technology with deep learning for agricultural authentication.