To achieve live detection（without power-off） of a power transformer by voiceprint

the voiceprint recognition of transformer core based on Mel-spectrum and convolutional neural network（CNN） is proposed in this paper. Firstly

the vibration acoustic signals of the transformer core under various operating conditions are studied through the test platform. The collected vibration and acoustic data are analyzed and compared

verifying that acoustic signals with better acquisition approach could represent the core operating state information reflected by the vibration signal before. Then

the collected acoustic data are transformed to Mel-spectra for dimension-reduction preprocessing and used as the deep learning dataset. Finally

the Mel-spectrum-CNN model for transformer core voiceprint recognition is constructed. Through the hyperparametric adjustment and network structure optimization

the acoustic signals of different operating conditions are accurately identified. The results show that the Mel-spectra-CNN recognition model proposed in this paper achieves 99.71% recognition rate of acoustic signals in 3 different working conditions

which provides a valuable reference for deep data mining of power system.