Abstract: In order to reduce the false alarm rate of voiceprint recognition algorithms in environment of low signal-noise ratios, this paper proposed a transformer mechanical fault voiceprint recognition method based on fast incremental support vector data description (FISVDD) and gate recurrent unit (GRU). The transformer was taken as the experimental object, and the sound signals of the transformer under normal working conditions, loose iron core state and loose coil state were obtained, respectively. Mel frequency cepstrum coefficients were used for feature extraction. As the first-level algorithm, FISVDD was used to separate strange classes and learn new samples through incremental learning. As the second-level classification algorithm, GRU was used to identify the samples passing the first-level algorithm. The experimental results show that FISVDD requires less training time than traditional closed-set recognition algorithms. Compared with traditional machine learning algorithms, GRU has higher recognition accuracy and anti-noise recognition ability in the transformer audio recognition task. Compared with the one-level algorithm, the proposed method is more effective. When identifying samples with a signal-noise ratio higher than 10 dB, the recall rate decreases by not more than 1%. When identifying samples with a signal-noise ratio not more than 0 dB, the false alarm rate does not exceed 10%.