Existing defect detection algorithms for power equipment are difficult to ensure both detection accuracy and speed, and the large scale of model parameter redundancy is a challenge for deployment in edge-side embedded devices. In this paper, we propose a power equipment defect detection technique based on federated gradient score correction(FedGSC) algorithm and real-time-detection Transformer(RT-DETR). First, the lightweight backbone network GhostNet is used to replace the original backbone network of the model in RT-DETR, and the model volume is further compressed using channel pruning, which significantly reduces the redundant parameters and improves the inference speed; on this basis, the federated learning architecture is introduced into the cloud server for the distributed training of the lightweight RT-DETR model at the edge end in order to solve the problem of non-independent homogeneous distribution(NID) in the federated learning training process. Non-independent and identically distributed(Non-IID) data exists, FedGSC is introduced to correct the gradient of each round of model update. Comparing the lightweight RT-DETR with traditional RT-DETR and YOLOv8, the algorithm model size is only 47 MB, and the mean average precision(mAP) is 90.46%, which can quickly and accurately identify the defects of power equipment; and the FedGSC algorithm reduces communication costs by approximately 40% and 20% compared to the FedAvg and FedFV algorithms, respectively.