Vibration-Based Distribution Transformer Fault Prognosis Using Recurrent Neural Networks with Self-Attention

• MHSA integrated with BiGRU and BiLSTM for vibration-based prognosis • 48% RAE reduction in excitation voltage prediction • 28% RAE reduction in interturn fault current prediction • Attention enhances temporal feature extraction from vibration signals • Validated on controlled transformer fault experiments Distribution and power transformers are crucial equipment in AC energy systems, serving purposes such as voltage conversion, voltage regulation, load balancing under imbalanced conditions, and even harmonic elimination. Reliable monitoring of these equipment is essential for grid stability. In recent years, vibration analysis has emerged as an effective method for assessing transformer condition. Deep recurrent neural networks (RNNs), specifically bidirectional long short-term memory (BiLSTM) and bidirectional gated recurrent unit (BiGRU) architectures, have already shown promising results in analyzing vibration signals for transformer fault prognosis. In parallel with these achievements, another innovation, known as the multi-head self-attention (MHSA) mechanism, has gained interest due to its effectiveness in analyzing sequential data. However, it remains unclear whether MHSA can enhance RNN performance for transformer fault prognosis—an issue that this study addresses. As the outcome of this exploration, the MHSA-BiGRU model achieves a 48% reduction in relative absolute error (RAE) compared to BiGRU when predicting excitation voltage. Similarly, the MHSA-BiLSTM model reduces RAE by 28% compared to BiLSTM for interturn short-circuit current prediction. These results demonstrate that integrating MHSA improves RNN-based models for transformer fault prognosis using vibration signals, offering a promising direction for more accurate and robust condition monitoring in power systems.