Deep Learning-Based Approach for Frost Quantification in Domestic Refrigerators

Frost formation in refrigerators disrupts heat ex-change, lowering system efficiency and causing operational is-sues, making defrosting essential for maintaining performance. However, modern defrost systems rely on sensors for frost detection. The sensory approach is reliable but it could degrade over time in changing environmental conditions. To address this limitation, the paper proposes a real-time frost quantification method using digital image processing and deep learning. A custom Convolutional Neural Network (CNN) model is designed to classify frost thickness in RGB images captured inside a refrigerator. We have used a custom dataset, collected from a real refrigeration environment and made a comparison with other methods previously used for frost thickness estimation. The dataset is collected from several experiments to ensure its diversity such as illumination and environmental changes. This method enables precise frost detection, targeted for optimal defrost control for improved efficiency. Results indicate that the CNN model achieves a high test accuracy of 98.10%. The model's performance surpasses traditional methods, demonstrating its robustness in distinguishing varying frost levels from 0 to 4 (mm). This method enables precise frost thickness estimation and is a suitable choice for defrost control to improve the energy efficiency of the refrigerator.