Prediction of Potato Rot Level by Using Electronic Nose Based on Data Augmentation and Channel Attention Conditional Convolutional Neural Networks

• Published in: Chemosensors 2024-12, Vol.12 (12), p.275
• Main Authors: Mai, Jiayu, Lin, Haonan, Hong, Xuezhen, Wei, Zhenbo
• Format: Article
• Language: English
• Subjects: Accuracy; Artificial neural networks; channel attention modules; Classification; Computer vision; Data augmentation; Datasets; Decay; Deep learning; electronic nose system; Electronic noses; feature-optimized deep learning; Gases; gaussian mixture embedded GAN (GMEGAN); Generative adversarial networks; Learning algorithms; Machine learning; Neural networks; Odors; potato decay prediction; Potatoes; Predictions; Quality assessment; Quality control; Robustness; Sensors; Signal processing; Sulfide compounds; Synthetic data
• ISSN: 2227-9040; 2227-9040
• DOI: 10.3390/chemosensors12120275

This study introduces a novel approach for predicting the decay levels of potato by integrating an electronic nose system combined with feature-optimized deep learning models. The electronic nose system was utilized to collect volatile gas data from potatoes at different decay stages, offering a non-invasive method to classify decay levels. To mitigate data scarcity and improve training model robustness, a Gaussian Mixture Embedded Generative Adversarial Network (GMEGAN) was used to generate synthetic data, resulting in augmented datasets that increased diversity and improved model performance. Several machine learning and deep learning models, including traditional classifiers (SVM, LR, RF, ANN) and advanced neural networks (CNN, ECA-CNN, CAM-CNN, Conditional CNN), were trained and evaluated. Models incorporating feature-optimized channel attention modules (f-CAM, f-ECA) achieved a classification accuracy of up to 90.28%, significantly outperforming traditional machine learning models (72–77%) and standard CNN models (83.33%). The inclusion of GMEGAN-generated datasets further enhanced classification performance, especially for feature-optimized Conditional CNN models, with an observed increase in accuracy of up to 5.55%. A comprehensive evaluation of the GMEGAN-generated data, including feature mapping consistency, data distribution similarity, and quality metrics, demonstrated that the generated data closely resembled real data, thereby effectively enhancing dataset diversity. The proposed approach shows significant potential in improving classification accuracy and robustness for agricultural quality assessment, particularly in predicting the decay levels of potatoes.