Abstract：In order to improve the accuracy of intelligent diagnosis of reactor mechanical fault，according to

the correlation characteristics between reactor vibration signal and mechanical state，a vibration diagnosis method

of reactor mechanical fault based on stacked auto-encoder（SAE）was proposed. Firstly，the original vibration

signal of reactor was decomposed by wavelet packet decomposition algorithm，and the time-frequency energy

matrix of the signal was extracted. Then，the diagnosis model of reactor mechanical fault based on SAE was built，

the deep feature mining of the time-frequency energy matrix was completed through unsupervised self-learning，

and the identification of reactor mechanical fault was realized through supervised fine-tuning. Finally，vibration

data of 10 kV oil immersed reactor under different mechanical states was used to train the fault identification model

and optimize the super parameters. The numerical results show that the proposed method can identify reactor

mechanical fault better than the traditional vibration signal identification method，and the accuracy can reach 98%.

      Key words：reactor；mechanical failure；vibration signal；wavelet packet decomposition；stacked auto-encoder（SAE）