Abstract:In view of the characteristic that the converter failure of wind turbines is often accompanied by severe temperature changes, this paper proposes a fault detection method based on VMD-Informer-LOF algorithm. Considering the impact of wind turbine state variables on the converter, this paper utilizes Variational Mode Decomposition (VMD) to extract the steady-state components and high-frequency interference noise from electrical signals such as current and voltage. Additionally, Permutation Importance is employed to select variables highly sensitive to temperature. Moreover, leveraging the characteristic temperature extremes during faults (either too low or too high), the paper reorganizes the predicted temperature values from the Informer model and the actual values from the SCADA (Supervisory Control And Data Acquisition system) into a two-dimensional sequence. The Local Outlier Factor (LOF) algorithm is then applied to identify temperature patterns under different states of the converter, enabling efficient anomaly detection. Experimental validation demonstrates that the proposed VMD-Informer model achieves a temperature prediction coefficient of determination with 0.9841, outperforming time-series prediction methods such as LSTM and XGBoost. Furthermore, when combined with sliding window data partitioning, the LOF algorithm effectively identifies anomalous data within the window. Under a fault rate threshold of 0.2, it can detect faults approximately 14 hours in advance, significantly enhancing fault warning effectiveness.