Novel bearing fault diagnosis model for wind turbine: PSR-CNN-DLSTM combined with transfer learning

Wind turbine bearings operate in harsh environments, which can complicate fault diagnosis due to the strong nonlinearity of vibration signals and imbalance data. To tackle these challenges, we propose a novel fault diagnosis model that integrates phase space reconstruction (PSR), convolutional neural networks (CNN), deep long short-term memory (DLSTM), and transfer learning (TL). Firstly, PSR is used innovatively to transform the non-linear vibration signals into chaotic phase diagram, facilitating more accurate feature extraction for the fault diagnosis model. Secondly, to further improve the generalization ability of LSTM, a DLSTM is designed and combined with CNN to form a fault diagnosis model. Subsequently, TL is employed to generalize the model across different working conditions, effectively mitigating the data imbalance issue. The proposed model enhances fault diagnosis accuracy for wind turbine bearings. Comparative experimental results demonstrate the proposed model's superior accuracy, achieving a significant improvement over existing methods in various working conditions.