Abstract: In view of the easy availability of sound data of industrial equipment, the sound data is collected for experiments such as load rejection during the start-up test of a hydraulic unit, and the collected sound data is analyzed by RMS, spectrum and spectrogram. Based on the waveform, spectrum and spectrogram, a neural network is selected as an auxiliary means, and the spectrogram is used as a training sample to enter the input layer of the neural network to obtain voiceprint features to realize the classification and scoring of test samples. The results show that different load conditions and tailgate leakage accidents can be correctly identified in the test. This research will help to establish a machine voiceprint feature map library for the whole and important key components of electromechanical equipment in hydropower stations.