Transformer Fault Diagnosis Method Integrating Long Short-Term Memory Network and Enzyme Action Optimization

Abstract

Dissolved gas analysis (DGA)-based transformer fault diagnosis is often affected by class imbalance, limited feature representation, and sensitivity to model hyperparameters. To address these issues, this paper proposes a novel computational method that integrates a long short-term memory network with a bio-inspired numerical optimization strategy. K-means SMOTE is first applied to balance the training data. A multi-scale convolutional neural network extracts local fault features under different receptive fields, after which an LSTM models the ordered dependencies among the extracted representations. A multi-head attention mechanism is further introduced to capture global feature interactions and enhance discriminative information. Finally, an enzyme action optimization algorithm is applied as a global numerical optimization approach for adaptive hyperparameter tuning, improving convergence behavior and diagnostic accuracy. Experimental results demonstrate that the proposed framework achieves improved accuracy and stability compared with existing methods, highlighting its potential applicability to a broad class of engineering fault diagnosis and condition monitoring problems.