WIND TURBINE ROLLING BEARING FAILURE PREDICTION BASED ON PSO-TRANSFORMER-BILSTM MODELING

• Published in: Scientific Bulletin. Series C, Electrical Engineering and Computer Science no. 3; p. 535
• Main Authors: Zhao, Zhenqiang, Pu, Houxu, Dong, Jianhua, Chen, Hong
• Format: Journal Article
• Language: English
• Published: Bucharest University Polytechnica of Bucharest 01.01.2025
• Subjects: Algorithms; Bearings; Datasets; Failure; Fault diagnosis; Feature extraction; Gearboxes; Particle swarm optimization; Prediction models; Roller bearings; Root-mean-square errors; Vibration; Vibration meters; Wind power; Wind turbines
• ISSN: 2286-3540

The failure of wind turbine gearbox bearings can negatively impact gridconnection efficiency and may even result in severe accidents, causing substantial financial losses for wind farm operators. This study introduces a PSO-TransformerBiLSTM model, which integrates the Transformer model, Bidirectional Long ShortTerm Memory (BiLSTM) model, and Particle Swarm Optimization (PSO) algorithm to solve wind gearboxes bearing failure problem. In this model, the Transformer component is responsible for extracting signal features, while the BiLSTM model is applied for both signal feature extraction and signal timing prediction. The PSO algorithm is utilized to fine-tune the model's hyperparameters for optimal performance. The vibration fault dataset for the rolling bearings was obtained through a wind power fault simulation platform using components such as inner-ringwear rolling bearings, acceleration sensors, and cloud vibration meters. Experimental results demonstrate that the PSO-Transformer-BiLSTM model achieves Root Mean Square Error (RMSE), R² values, and Mean Absolute Error (MAE) of 13.26, 0.93, and 10.37, respectively, on the faulty bearing vibration dataset. These results indicate an improvement in performance compared to individual Transformer and BiLSTM models, with the R² value increasing by 0.01 and 0.03, respectively. Additionally, when compared with other phase prediction models, such as CNNLSTM-XGB, TCN-LSTM, and CNN-LSTM-SE, the R² coefficients of the PSOTransformer-BiLSTM model surpass them by 0.12, 0.02, and 0.01, respectively. These findings confirm that the PSO-Transformer-BiLSTM model delivers reliable performance on faulty bearing datasets and offers valuable insights for diagnosing vibration faults in wind turbines.