Abstract
As power grids transition into "prosumer" networks with high levels of distributed energy and electric vehicle integration, the operational stress on transformers has increased significantly. Intelligent inspection using deep learning is a promising solution, but its effectiveness is often limited by the scarcity of high-quality fault images. This paper proposes a comprehensive framework for transformer appearance fault diagnosis. The approach integrates a seamless cloning algorithm for realistic data generation, an XGBoost-based quality screening process to refine the dataset, and a heterogeneous model ensemble using weighted voting to maximize diagnostic accuracy. Experimental results show that this optimized pipeline achieves a 96% accuracy rate, significantly outperforming individual deep learning models.
Introduction
Monitoring the health of power transformers is critical for grid stability. While automated image recognition can streamline inspections, collecting sufficient data on specific faults like surface corrosion (rust) is challenging because equipment is typically well-maintained. Traditional data augmentation techniques like flipping or rotating do not provide sufficient variety, and simple "copy-paste" methods for fault synthesis often result in unrealistic images that degrade model performance. This study addresses these challenges by optimizing the dataset through synthetic generation and rigorous cleaning, followed by a robust ensemble-based diagnostic method.
Methodology
Data Generation via Seamless Cloning
To resolve the data scarcity issue, the authors employ a seamless cloning algorithm based on the Poisson equation. Unlike standard image overlays, this method operates in the gradient domain, matching the texture and lighting of a source fault (e.g., rust) to the target transformer surface. This ensures that the synthetic faults appear naturally integrated into the metallic structure, providing more effective training samples for deep learning models.
Data Refinement: XGBoost Quality Screening
Synthetic generation can occasionally produce unrealistic artifacts, such as rust appearing on non-metallic surfaces or in the background sky. To maintain dataset integrity, an XGBoost-based quality assessment model is introduced. This model evaluates the realism of generated images and automatically discards low-quality samples. In this study, approximately 30% of the initial synthetic dataset was removed during this cleaning phase to ensure the model learns from high-fidelity data.
Heterogeneous Model Ensemble
Rather than relying on a single architecture, the framework utilizes a heterogeneous ensemble of models, including GoogLeNet, LeNet-5, BP Neural Networks, K-Nearest Neighbors (KNN), and Support Vector Machines (SVM). A weighted voting strategy is used to fuse the outputs, where the weight of each model's vote is determined by its individual F1 score and accuracy. This approach mitigates the risk of individual model failures and improves overall diagnostic reliability.
Results and Insights
The study found that a 1:3 ratio of rusted to normal images in the training set yielded the best results. The inclusion of synthetic data boosted GoogLeNet's accuracy from 85% to 91%, and the addition of the XGBoost cleaning step further increased it to 94%. The final heterogeneous ensemble achieved a peak accuracy of 96% and an F1 score of 96%, outperforming standalone models like BP (53%) and KNN (91%). A key insight was that diversity in model types (e.g., combining CNNs with traditional classifiers) is more resilient than relying on a single state-of-the-art architecture.
Conclusion
By combining gradient-based data generation, machine learning-driven data cleaning, and an ensemble of diverse diagnostic models, this framework provides a robust solution for transformer fault detection. Future work will focus on automating mask selection for fault synthesis and expanding the system to detect other faults, such as oil leakage or foreign body intrusion.
Summary updated