In the field of voiceprint online monitoring of high-voltage shunt reactors

the acoustic signals exhibit long time-series characteristics with inherent high complexity

large data dimensions

and energy dispersion

thus problems such as the low information utilization efficiency of voiceprint recognition models

along with inadequate robustness and recognition accuracy may arise. To address these issues

an improved ConvNeXt-T network voiceprint recognition model based on multi-level feature map is proposed. Firstly

the acoustic signal is converted into time domain and frequency domain feature maps by symmetrized dot pattern and Gram-like matrix graphical refinement spectrum. Based on the characteristics of reactor voiceprint

the 50 Hz Gammatone filter banks are proposed to generate energy feature maps. Then

the lightweight CA(Coordinate Attention) attention mechanism is introduced as the feature map adaptive fusion module to improve the input side of the ConvNeXt-T network. Finally

the superiority of the model is verified by the measured data. The results show that the average recognition accuracy of the proposed model on the test set is 97.82%

which is 3.14% higher than that of the single-domain map

and 6.51% higher than that of the FCN

RsNet

ApR-IDRSN and other comparison models. At the same time

the model shows the best anti-noise performance in Gaussian white noise

human voice and bird sound environments. The model combines high-dimensional multi-domain feature extraction method and graphical dimension reduction representation method

which can significantly improve feature richness and information utilization.