This work proposes an adaptive fault diagnosis method for hybrid transformer mechanical failure in incubation period. An antenna-augmented radio frequency identification (RFID) sensor is used to obtain vibration signal as well as acquire identification information. The sparse stacked denoising autoencoder (SSDA) which has satisfactory performance in dealing with nonlinear and noise-polluted signal is adopted to extract features from high-dimensional raw signal on edge server. Since the hidden layer structure and learning rate of the SSDA model highly determines the performance of feature extraction, the chaotic quantum particle swarm optimization (CQPSO) is employed to optimize the above parameters. The performance of SSDA highly relies on abundant training data, which is hardly obtained from one single transformer and the data sharing among different transformers is impossible due to data privacy protection, so the Federated learning (FL) is employed to realize global diagnosis model training from several independent transformers collaboratively as well as protect data privacy. The data collected from different transformers are usually non-independent identically distributed (NIID), which would greatly decrease the performance of FL. Thus, an improved data sharing strategy is introduced to selectively deliver cloud shared data so as to reduce the influence of NIID data. The experiments validate that the augmented RFID sensor has reliable communication performance within the distance of 17.5 m. Moreover, the optimized FL achieves prominent results for transformer hybrid fault diagnosis in early stage.