A Counterfactual Diagnosis Method for Power Transformers Based on a Differentiable Structural Causal Model

Lichuan Lei

Abstract: With the increasing intelligence of power systems, multimodal monitoring data of transformers have become increasingly abundant. However, existing diagnosis methods mostly rely on statistical correlations, lack explicit causal reasoning capability, and struggle to handle unseen fault combinations. To address these issues, this paper proposes a counterfactual diagnosis method for power transformers based on a differentiable structural causal model (DSCM-CFD). The method explicitly constructs a causal generative mechanism from vibration to temperature to audio, and introduces path consistency loss and counterfactual consistency loss, jointly optimized within a variational autoencoder framework. The model achieves high classification accuracy, accurate root cause localization, and effective zero-shot diagnosis. Experimental results demonstrate that the proposed method outperforms existing methods in diagnostic accuracy, zero-shot generalization, and causal interpretability, providing a new technical pathway for intelligent operation and maintenance of power equipment.

Keywords: causal reasoning, differentiable structural causal model, multimodal fusion, power transformer fault diagnosis