A Novel Causal Federated Transfer Learning Method for Power Transformer Fault Diagnosis Based on Voiceprint Recognition

Fault diagnosis of power transformers based on voiceprint analysis has developed rapidly due to its non-intrusive advantages in recent years. However, it faces challenges in generalization across different voltage levels and collaborative training difficulties under distributed data barriers. Existing federated transfer learning methods rely on statistical correlations, which are easily affected by noise and hinder better fault diagnosis performance. Therefore, this paper proposes a novel causal federated transfer learning method for power transformer fault diagnosis based on voiceprint signals. First, a causal federated transfer learning framework is proposed by integrating a causal graph autoencoder into federated transfer learning to capture nonlinear causal features between voiceprint features and faults. Second, a graph autoencoder with a wavelet convolutional encoder layer and a sub-pixel convolutional decoder layer is constructed to extract domain-invariant causal features from key fault related frequency bands. Third, a strategy is designed to aggregate encoder layer information using adversarial-loss-sensitive weighting, which effectively evaluates the contribution of each client while reducing communication overhead. Experimental results show that the proposed method can quickly identify fault types in cross-voltage-level power transformer fault diagnosis scenarios and outperforms existing models in all three scenarios. Even under a high noise level of –5 dB in the third scenario, the accuracy still exceeds 94%.