Deep learning-based power transformer condition monitoring and fault diagnosis algorithm

Power transformers are critical in smart grids, but accurate fault diagnosis remains challenging due to limitations in conventional dissolved gas analysis and data-intensive deep learning methods.This paper introduces an integrated deep learning framework combining convolutional neural networks with bidirectional long short-term memory via an adaptive multi-head attention mechanism.The model jointly processes dissolved gas analysis profiles and vibration signatures: convolutional layers extract spatial features from gas data, while bidirectional units capture temporal vibration dynamics.The attention mechanism dynamically highlights discriminative features, improving both interpretability and accuracy.Evaluated on the IEC dissolved gas analysis dataset and real vibration records, our method achieves 98.2% diagnostic accuracy, surpassing support vector machines (85.1%) and deep belief networks (90.8%).Additional tests confirm robustness in detecting incipient faults under varied conditions.Attention weights align with known failure mechanisms, offering a reliable, interpretable tool for predictive maintenance.