Converter Transformer fault diagnosis method based on double-weighted loss function and DT-Lightgbm-Gcforest

A converter transformer fault diagnosis method based on double-weighted loss function and DT-Lightgbm-Gcforest is proposed to improve the effect of converter transformer fault diagnosis under the unbalanced sample data. Aiming at unbalanced fault samples and the difficulty of diagnosing complex categories, the category sample distribution and the interpretation degree of the binary Logistic model are employed to calculate the category balance weight and interpretive quantitative weights for different fault categories, and the multi-category cross-entropy loss function is constructed. Then, since the ensemble learning parameters have a large dimension and are difficult to optimize, the tree-structured parzen estimator (TPE) algorithm is employed to optimize parameters of the Light Gradient Boosting Machine (Lightgbm) and Multi-Grained Cascade Forest composed of multiple Lightgbm, dual-stage TPE optimized Lightgbm-Gcforest is established as the final fault diagnosis model. Finally, the double-weighted cross-entropy is adopted as a loss function to perform the model’s training and fault diagnosis. Simulation results indicate that the overall accuracy of the proposed method is 94.94%, which is 5.05% higher than that without weighting the cross-entropy loss function and 2.81% higher than that before using the TPE optimization algorithm, demonstrating the effectiveness of the proposed method.