Abstract: Aiming at overcoming the problems of fuzzy information and low contrast at the defect in infrared images of substation equipment with complex backgrounds, which lead to low recognition accuracy and slow identification speed in mobile terminal deployment, this study developed an infrared defect identification algorithm (YOLOV8-GCAFM: YOLOv8 based on Ghostv2 and attention mechanism fusion). First, a lightweight GhostNetV2 BottleNeck was introduced in the backbone network to replace some CBS modules, which reduced the number of network parameters, enhanced the model expression ability, and increased the detection speed. Second, an attention fusion module (v3_CAFM) was developed, which was added to the neck network to capture more extensive feature information globally and locally, and improve network accuracy. Finally, the wise intersection over union loss function was used to optimize the gradient gain allocation strategy to enhance the adaptability of the detector to changes in multiscale targets. A test was conducted on a self-built dataset containing infrared defect images of various pieces of transformer equipment. The experimental results showed that, compared with the baseline algorithm, mAP@0.5, mAP@0.5:0.95, and P/% increased by 3.2%, 8.1%, and 3.8%, respectively. The model size was 6.1 MB, and the performance of the improved algorithm exceeded that of YOLOv8. The YOLOv8-GCAFM model could satisfy the requirements of the real-time identification of equipment defects and provide a reference for the subsequent fault diagnosis of substation equipment and applications on resource-limited devices such as mobile terminals.