Fine-grained visual recognition of residential air conditioner units for building energy management

Background Accurate identification of installed appliances is important for energy analytics, demand-side management, automated building inventories, and energy auditing. Residential air-conditioning systems are among the most energy-intensive household appliances, yet detailed information about installed units is often unavailable or difficult to collect at scale. This study investigates the recognition of residential air-conditioning indoor units directly from images as a fine-grained visual classification problem. Methods A curated dataset comprising 720 images from 216 air-conditioner manufacturer-model categories was constructed by combining field-acquired photographs captured through a custom mobile application with supplementary web-sourced images. Five representative deep learning architectures were evaluated under a unified experimental protocol. The best-performing baseline was subsequently refined through controlled ablation experiments examining activation functions, classification heads, learning-rate schedulers, and backbone architectures. Results Cross-layer feature aggregation provided the strongest baseline performance. Further improvements were obtained using GELU activation, CosFace cosine-margin supervision, OneCycleLR scheduling, and a DenseNet-161 backbone. The final CN-CNN configuration achieved a validation Top-1 accuracy of 74.03%, a Top-3 accuracy of 87.01%, and a macro F1-score of 0.7054, representing an improvement of 11.69 percentage points in Top-1 accuracy over the initial CN-CNN baseline. Conclusions The findings demonstrate that fine-grained visual recognition can support scalable and non-intrusive identification of residential air-conditioning units. The proposed framework provides a basis for integrating computer vision into automated building inventories, energy-auditing workflows, and smart energy-management applications.