An Attention-Guided Semi-Supervised Model for Power Transformer Fault Diagnosis via Vibration-Acoustic Data Fusion

Abstract

Reliable fault diagnosis of power transformers is vital for ensuring the safe and continuous operation of power systems. Although deep learning methods have shown success with single-sensor data, their diagnostic performance remains limited due to the inability to capture complex, multimodal fault characteristics. To address this, we propose an attention-guided semi-supervised vibration-acoustic fusion (AG-SVAF) model, which combines vibration and acoustic signals to enhance diagnostic robustness under limited labelled data conditions. The model integrates time-frequency representations derived via short-time Fourier transform (STFT) with a multilevel attention mechanism—including channel, spatial and self-attention—to highlight fault-relevant features and model cross-modal dependencies. A novel attention-weighted consistency loss further improves the utilisation of unlabelled data during training. Validated on practical transformer datasets, AG-SVAF achieves superior performance in terms of diagnostic accuracy and stability, particularly under challenging scenarios involving class imbalance and label scarcity. This approach provides a promising and scalable solution for intelligent condition monitoring in real-world power system applications.