Research on Transformer Fault Diagnosis Algorithm Based on MFCC Features and Deep Learning

This paper proposes a transformer fault diagnosis algorithm based on Mel-Frequency Cepstral Coefficients (MFCC) and deep learning to enhance the accuracy and robustness of power equipment monitoring. By analyzing transformer noise signals, MFCC features are extracted to capture fault-related acoustic characteristics in perceptually critical frequency bands. A convolutional neural network model is designed with three convolutional layers, max-pooling layers, and fully connected layers, optimized via Bayesian hyperparameter tuning. Experimental validation on a dataset demonstrates the model’s superior performance. The CNN architecture leverages local receptive fields and translation invariance to adapt to time-varying noise, enabling reliable non-intrusive fault diagnosis in complex environments. Challenges such as cross-regional generalization and unknown fault detection are identified for future research, with proposed extensions including transfer learning and multimodal data fusion.