Accurate and reliable fault diagnosis of wind turbine planetary gearboxes is essential for ensuring stable operation and reducing maintenance costs. However, vibration signals collected from planetary gear systems are usually characterized by strong non-stationarity, complex temporal dependencies, and multi-scale fault information, which makes effective feature extraction challenging. To address these issues, this paper proposes an efficient channel attention enhanced U-shaped temporal convolution network (ECA-U-TCN) for intelligent planetary gear fault diagnosis. The proposed model employs a fully connected projection layer to map raw vibration signals into a high-dimensional feature space, followed by stacked U-shaped temporal convolution blocks for hierarchical temporal feature extraction. Dilated convolutions are utilized to capture both local fault characteristics and long-range temporal dependencies, while the encoder-decoder structure and skip connections facilitate multi-scale feature fusion and preserve fine-grained fault information. Furthermore, an efficient channel attention mechanism is introduced to adaptively recalibrate the extracted feature responses and enhance fault-sensitive channels with limited additional computational complexity. Layer normalization is subsequently applied to stabilize feature representations, followed by a classifier for fault identification. Experimental results on a planetary gearbox dataset demonstrate that the proposed method outperforms conventional models including CNN, ResNet, TCN, U-Net, and LSTM, achieving an accuracy of 99.10%, with macro precision, recall, and F1-score all reaching 0.99. In addition, t-SNE visualization and the confusion matrix further demonstrate the strong feature discrimination and classification capability of the proposed model. These results indicate that ECA-U-TCN provides an effective framework for intelligent condition monitoring and fault diagnosis of wind turbine planetary gear systems.

Wind turbine, Fault diagnosis, Planetary gear system, Channel attention, Temporal convolution network