Transformer Voiceprint Feature Extraction and Fault Recognition Based on MFCC and Deep Learning

This study proposes a transformer fault voiceprint recognition model based on Mel Frequency Cestrum Coefficients (MFCC) and deep learning, aiming to address the challenges posed by low signal-to-noise ratio and the associated issues of low accuracy and unstable performance in transformer fault recognition using transformers. The proposed model employs a two-step approach: firstly, the transformer acoustic signal is transformed into MFCC, forming a group of MFCC feature vectors as input. Secondly, convolutional neural networks (CNN) are utilized to extract spatial feature information from the MFCC, and subsequently, bi-directional Long-Term Memory (Bi-LSTM) is employed to fully extract temporal feature information. The feature classification is then performed through a fully connected layer and a Softmax layer. Experimental results demonstrate the model's stability under low signal-to-noise ratio noise background, exhibiting a fault identification accuracy rate exceeding 99.5% for transformer acoustic signals under three distinct operating conditions.

I. Introduction

Power transformer failures pose significant risks to power grid reliability, necessitating timely identification and resolution. Sound signals emitted during faults offer valuable insights into transformer operational condition.