The acoustic signals generated by gas-insulated switchgear (GIS) operations are typically transient and impulsive in nature

and are often contaminated by noise. These characteristics can significantly degrade the performance of acoustic pattern recognition algorithms. To address this issue

a GIS action voiceprint recognition method based on learnable feature extraction and an improved Transformer is proposed. Firstly

Gabor convolution with learnable parameters

pooling layers

and single-channel compression are used instead of the traditional Mel-frequency Cepstral coefficient feature extraction method. By fine-tuning parameters for GIS action voiceprints

the feature representation capability of the spectrogram is enhanced. Secondly

the Transformer network introduces multi-head attention mechanisms and time-frequency attention mechanisms to learn audio features from multiple perspectives

capturing the characteristic information of instantaneous audio signals during time-frequency transitions

thereby enhancing its generalization capability and robustness to noise. Finally

GIS was used as the experimental subject to complete the collection of voiceprint signals during the opening and closing processes of various mechanical fault conditions. Experimental results demonstrate that

compared to traditional voiceprint recognition methods

the proposed algorithm achieves superior representation of the acoustic features in transient signals

leading to improved accuracy and enhanced robustness to noise. Under high signal-to-noise ratio conditions

the recognition accuracy remains stable at 96%

while at a signal-to-noise ratio of 10 dB

the recognition accuracy of the proposed algorithm improves by more than 9.4%

and the recognition performance degrades less under low signal-to-noise ratio conditions.