13 / 2025-03-18 12:42:29

A Lightweight Federated Transfer Learning Method for Fault Diagnosis of Power Transformers Based on Voiceprint Signals

federated transfer learning,across-voltage,lightweight model,voiceprint

This paper presents a power transformer fault diagnosis method based on lightweight federated transfer learning, addressing the limitations of traditional voiceprint-based diagnostic methods in across-voltage levels and data privacy protection. The proposed approach integrates adversarial domain adaptation with a federated learning framework, effectively mitigating the challenges posed by differences in voiceprint feature distributions across power transformers of various voltage levels. It enables coordinated optimization of privacy protection and across-voltage knowledge transfer. For voiceprint feature extraction, a novel lightweight network is designed by combining residual structures with depthwise separable convolutional layers, enhancing feature extraction capabilities while reducing the number of parameters. Furthermore, a partial parameter aggregation strategy is employed, uploading only the parameters of the feature extraction network. This reduces the communication overhead and accelerates system response time. Experimental results demonstrate that the proposed method achieves high accuracy and robustness in across-voltage level power transformers fault diagnosis, significantly improving fault detection efficiency.