Detecting abnormal internal temperatures in inverters is a critical aspect of ensuring the safe operation of power equipment. To address the shortcomings of existing methods in detecting minute thermal targets, multi-scale feature fusion, and quantitative temperature estimation, we propose the Temp YOLO algorithm based on an enhanced YOLOv8. This algorithm innovatively integrates a hybrid backbone network combining EfficientViT and MobileNetV3, balancing global feature extraction with the need for a lightweight architecture, thereby effectively enhancing its representational capabilities; A temperature-aware bidirectional feature pyramid network (T-BiFPN) is designed, which significantly improves the localisation accuracy of minute thermal anomaly regions through cross-scale feature fusion and temperature-priority guidance; a multi-task dynamic detection head is developed to simultaneously achieve target classification, bounding box regression, and temperature level estimation, thereby overcoming the limitations of traditional single-task detection. In validation using a self-built inverter infrared image dataset, Temp YOLO achieved an mAP@0.5 of 9.28%, representing a 535.6% improvement over the baseline model YOLOv8n, whilst reducing computational load (GFLOPs) by 60.1% and model parameters to 2.7 million, with a frame rate of 65 frames per second, thereby meeting real-time monitoring requirements. Experiments demonstrate that this method outperforms state-of-the-art (SOTA) models such as YOLOv8s and PP-YOLO in terms of detection accuracy and robustness for small targets (e.g., overheating of IGBT pins) in complex scenarios. Ablation studies confirm significant synergistic effects among the modules, with the T-BiFPN contributing a 12.4% improvement in mAP for small targets. This method achieves significant improvements in detection accuracy and computational efficiency compared to baseline models, providing a viable technical solution for real-time thermal management of inverters and demonstrating strong application potential in specific scenarios.