To address the scarcity of power meter detection data,significant distribution differences between generated and real data,and the weak generalization ability of feature extraction networks,an unsupervised object detection algorithm based on data generation is proposed.Firstly,a large number of annotated power meter images are generated using a text-to-image model.Then,a domain adaptation strategy based on sample mixing is proposed,where real images with high-confidence output are selected and spliced with generated images for mixed training,which can mitigate the negative impact of the distribution differences between generated and real data.Finally,a mask consistency module is added to enable the model to learn more universal feature representations and improve its generalization ability in unknown scenarios.The test results show that the algorithm improves the mean average precision (mAP) by 14.1% compared with networks trained only on generated images and outperforms the existing classic domain adaptation algorithm SWDA Faster R-CNN (strong-weak domain adaptation Faster R-CNN) by 10.9%.