Diffusion-TS Guided Small Sample Data Augmentation for Power Transformer Fault Diagnosis

This is the authors' abstract. We don't add key points for this paper.

The shortage of fault samples remains a significant challenge in artificial intelligence-based fault diagnosis for power transformers. To address this challenge, here we introduce and improve the Diffusion-TS method for generating reliable Dissolved Gas Analysis (DGA) training data for transformer fault diagnosis. Initially, a low-pass filter is incorporated into the deep decomposer to construct a trend-extraction layer, thereby improving the accuracy of trend extraction. Then, a dynamically decaying Exponential Moving Average (EMA) is employed during training to optimize model parameters, mitigating issues of unstable training and limited generalization. Using DGA data collected from power transformers, we conducted three experiments to evaluate the model performance in terms of generation similarity, correlation, and utility. Experimental results demonstrate that the generated data exhibits high similarity and correlation with real-world data, making it highly effective for training classification models for transformer fault diagnosis.