Imbalanced Sample Learning Method for Transformer Fault Diagnosis Based on GHM Loss and Stacking Frames

Aiming to address the significant deviations in transformer fault identification when uneven fault data are processed, a hybrid integration model based on the improved light gradient boosting machine (GCLightGBM) is proposed in this study for effective transformer fault identification. This model consists of two key components. On the one hand, an improved LightGBM model called GCLightGBM is introduced. On the other hand, a hybrid integration model based on GCLightGBM is developed to resolve the problem that the recognition ability of the model is affected by parameter specificity values in GCLightGBM. GCLightGBM combines a gradient harmonized loss function (GHM Loss) with a cross-entropy loss function, which makes the model better focus on a few critical samples in the dataset. By doing so, GCLightGBM improves the performance of the model in identifying faults accurately. The hybrid integration model aims to further improve accuracy while ensuring that the model maintains a high performance on realistic variable unbalanced datasets. This approach improves the accuracy of fault identification while still rendering the dataset imbalanced. The experimental results demonstrate that GCLightGBM is effective in solving the low accuracy for certain classes of samples. In addition, the validity of the fault identification method based on the GCLightGBM hybrid integral model is validated on a publicly available dataset, making the model suitable for transformer fault identification.

I. Introduction

Power transformers are the crucial equipment required for power transmission and conversion so that whether they function well has a considerable impact on the stability of the entire power system [1]. At present, a large majority of the transformers in operation are oil-immersed transformers. When a transformer fails, it produces various gases that dissolve in the transformer oil, including hydrogen (H₂), methane (CH₄), ethane (C₂H₆), ethylene (C₂H₄), acetylene (C₂H₂), and many more. There is a close correlation between the gas content in the oil and the type of transformer failure [2], [3]. For this reason, the gas concentration in oil can be detected by means of dissolved gas analysis (DGA) to determine the types of fault [4]. With the advancement of artificial intelligence, the integration of deep learning, machine learning, neural networks, and other intelligent algorithms has been widely practiced for the detection of faults in oil-immersed transformers, which has been shown to significantly improve the accuracy of fault identification [5], [6]. To ensure the high accuracy of prediction results, sufficient and balanced training data are required for the process of AI model fitting. However, this imbalance impairs the generalization performance of the training network significantly, thus leading to a significant deviation in the training model [7].