The respirator is a critical auxiliary component in the operation of power transformers, and its abnormal condition may lead to reduced transformer performance or even failures. To achieve efficient and accurate detection of respirator anomalies, this paper proposes an end-to-end respirator anomaly detection method based on an improved RT-DETR model. The method first employs an enhanced Backbone for multi-scale feature extraction, introducing SE channel attention at the output of each scale to strengthen the features. Next, the method employs an efficient hybrid encoder to enable feature interaction within scales and fusion across different scales. Finally, the decoder refines the outputs iteratively to produce bounding boxes and their corresponding confidence values. Experiments on a self-built dataset reveal that this method delivers outstanding detection results, significantly boosting the overall detection accuracy while reducing the computational cost. Specifically, the proposed approach enhances the mAP@0.5 and mAP@0.5:0.95 metrics by 8.1% and 13.8%, respectively, compared to the baseline. Additionally, it reduces model parameters and computational load (FLOPs) by 17.5% and 14.5%, respectively. This provides an efficient and reliable solution for the intelligent monitoring of respirators.