A Residual U-Net-Based Model for Intelligent Identification of Dust Accumulation on Mesh-Grid Air Filter of Split Air Conditioner

Split air conditioners are widely used in residential homes where a central air conditioning system is not available. Dust accumulation on the mesh-grid air filter of the device can significantly increase airflow resistance, leading to reduced energy efficiency of air conditioners. Current filter maintenance relies primarily on visual inspection results or a fixed cleaning schedule, both of which are ineffective. Based on a dataset of 4900 images of mesh-grid filters collected from split air conditioners, this study proposes a Residual U-Net model with residual connections to simultaneously perform pixel-level dust area segmentation and three-level dust thickness categorization of the filters (mild, moderate, and severe). To bridge visual outputs with physical performance, a dust accumulation index (DAI) was established to quantitatively evaluate the overall dust accumulation degree by integrating the predicted dust area and thickness levels. Model performance on the 500 independent field test samples achieved an average intersection over union of 80%, a Dice coefficient of 86%, and a thickness classification accuracy of 93%. These numbers deviated only marginally (approximately 1%) from the validation performance, confirming the model’s reliable generalization capability under complex real-world operating conditions. This machine learning model provides a solid technical foundation for the image-based assessment of dust accumulation from photographs of the filter to support effective cleaning schedules.