An innovative transformer fault diagnosis method based on ACK-GRANDE

To address the limitations of existing DGA-based transformer fault diagnosis methods and further enhance fault diagnosis accuracy, this paper proposed a new transformer fault diagnosis method. First, during the pre-processing of DGA data, data cleaning was performed to remove anomalies and noise, and the Adaptive Synthetic Sampling Approach for Imbalanced Learning (ADASYN) algorithm was used to alleviate data imbalance, thereby improving data quality for subsequent diagnostics. Next, an improved model, referred to as ACK-GRANDE was proposed. The key optimization modules were divided into the following parts. Firstly, a refined feature weight allocation strategy was introduced, simulating a graph attention mechanism to adjust feature weights, avoid local optima, and enhance the ability to capture key information. Additionally, supplementary losses based on cosine similarity and Euclidean distance from K-Means clustering were incorporated, constructing a new loss function to enhance the model’s ability to recognize patterns in complex data. Finally, using 1200 DGA data samples with known fault types for case analysis, the results demonstrated that, compared to other traditional machine learning classification models, the proposed algorithm achieved higher accuracy in transformer fault diagnosis. The overall accuracy reached 93.75 %, exhibited superior long-term stability and broader applicability, and ablation experiments further validated the feasibility of each optimization module.