Transformer fault diagnosis based on LIF technology and COA-GRU algorithm

Abstract Power transformers serve as a crucial link between generators and power grids, and their stable operation is paramount, necessitating regular inspections. Given the shortcomings of traditional DGA methods, such as high error rates and low sensitivity, this paper innovatively proposes a transformer fault diagnosis approach that integrates LIF spectroscopy technology with a COA-GRU algorithm. The experimental focus is on four types of insulating oil samples representing normal conditions, short-circuit faults, moisture-contaminated insulation, and thermal faults. These samples undergo spectral characteristic analysis using LIF technology.To optimize data quality, SNV and Z-score preprocessing techniques are employed to reduce noise. Furthermore, Linear LDA and T-SNE are utilized in parallel for dimensionality reduction, ensuring that the richness of spectral information is preserved while significantly reducing the data dimensions. Subsequently, three deep learning models - RNN, LSTM, and a specially optimized COA-GRU model - are constructed and trained on the dimensionality-reduced data.The results demonstrate that the COA-GRU model outperforms its counterparts across various metrics, emerging as the preferred solution. This effectively validates the model's efficiency and practicality in transformer fault diagnosis, offering a novel approach to safeguarding the stable operation of power systems.