Transformer voiceprint signal contains a lot of effective information reflecting the internal mechanical state. In order to realize uninterrupted detection of internal mechanical state of transformera voiceprint recognition method of transformer mechanical state based on feature screening and improved deep forest is proposed. Firstlythe intrinsic mode function(IFM) is obtained by decomposing the voiceprint signal with the complete ensemble empirical mode decomposition with adaptive noise (CEEMDAN)and the IMF component containing the fault information is obtained by filtering the IMF components through spectrum analysis and Pearson correlation coefficient. Secondlythe distribution of each IMF component in the frequency band is used to divide the high and low frequency bands. According to the difference of the IMF components in the high and low frequency bandsthe time-frequency energy of the IMF component in the high frequency band and the amplitude characteristic of the IMF component in the low frequency band are used as characteristic indicators to form a feature vectorwhich is input into the improved deep forest modeland the voiceprint recognition results of 10 mechanical loose states are obtained. Finallythe effectiveness of the method is verified by field experiments. The research results show that the average recognition accuracy of the proposed method is 99.2% for 10 mechanical loose states. Compared with the traditional transformer voiceprint featurethe proposed voiceprint feature has higher discrimination; Compared with the traditional recognition modelthe proposed improved deep forest recognition model has lower complexityfaster training speed and higher recognition accuracy.