Frost-EffNet: a deep learning model for frost detection in refrigeration systems

Abstract

Frost accumulation on the heat exchange units in refrigeration systems significantly impairs performance by restricting airflow, leading to increased power consumption. Despite its impact, the accurate and timely detection of frost buildup remains a challenging task, hindering the efficient initiation and control of defrost mechanisms. This study introduces Frost-EffNet, an innovative model to predict frost accumulation on refrigeration unit evaporator coils. Frost-EffNet is a modified version of the EfficientNet architecture. The model was validated using five distinct EfficientNet variants, with performance evaluated on a dataset of evaporator coil images. Comparative analysis was conducted by varying key hyperparameters to identify the optimal configuration for the best-performing model. Additionally, Grad-CAM (Gradient-weighted Class Activation Mapping) analysis was employed to highlight the areas of the evaporator coil that the model focused on for frost estimation, providing valuable insights into the model's decision-making process. The proposed Frost-EffNet model demonstrated high predictive accuracy, achieving 92.47 % accuracy on the test dataset, with a mean absolute error (MAE) of 0.11 mm. A comparative performance analysis revealed that Frost-EffNet outperforms other convolutional neural network (CNN) models in terms of accuracy, while also being computationally more efficient.