A Multi-Granularity Voiceprint Feature Adaptive Extraction Method for Rotating Machinery Based on Multi-Spectrogram Fusion

Aiming at the difficulty of extracting low-frequency features and incomplete voiceprint features in the strong noise environment in rotating machinery fault recognition, this paper proposes a multi-granularity voiceprint feature adaptive extraction (MG-FAE) method for rotating machinery based on multi-spectrogram fusion. Firstly, the space-spectrogram joint feature set based on the nonlinear mapping advantages of Mel spectrum and mel frequency cepstrum coefficients (MFCCs) is constructed to improves the voiceprint feature representation capability of low-frequency signals. Then, the multi-granularity representation method of multi-spectrogram based on Gaussian Difference Pyramid (GDP) is proposed to fully characterize the geometric and semantic information in spectrogram, the Multi-Head Attention Mechanism (MHA) is introduced to adaptively extract multiple voiceprint features to achieve comprehensive representation. Finally, a MG-FAE method is proposed to representat the voiceprint features of rotating machinery. The effectiveness of the method is verified by iFL YTEK rotating machinery challenge dataset. The recognition rate is 97.64% in multi noise state, which is 37.36% higher than the original audio signal. Moreover, the t-SNE feature visualization technology verifies that the MG-FAE method proposed in this paper has clear boundary of voiceprint feature classification and good verification effect.