Voiceprint Recognition of Mechanical Faults in Transformers Based on Data Enhancement
In the field of power equipment voiceprint monitoring,the small scale of fault audio sample data is a major challenge.Therefore,a data enhanced method for transformer mechanical fault voiceprint recognition was proposed.Firstly,offline audio processing was used to enhance the audio samples,followed by the synthesis of new audio samples using wave generative adversarial network(WaveGAN).Finally,a recurrent neural network(RNN)was trained using the enhanced samples to achieve voiceprint recognition.The mechanical faults were simulated on transformers and audio samples were collected from transformers under different operating conditions for testing.Compared with other audio generation methods,the proposed method generates higher sample quality;when the duration of a single label is in the range of 30～60 s,it can generate twice the number of valid samples.After enhancement,the recognition accuracy increased by 2.95 percentage point.The experimental results show that the proposed method can effectively expand the voiceprint samples of power equipment and improve the accuracy of voiceprint recognition.
万方数据
维普
