Intelligent fault diagnosis of yaw damper in graph convolutional neural networks based on convolution and attention mechanisms

Abstract

Traditional fault diagnosis methods for yaw damper based on deep learning focus solely on the features of each individual sensor signals, neglecting the interrelationships of signal features under the spatial layout of the sensors. To address this, the sensors are treated as nodes in a graph neural network. To effectively capture the topological spatial information between these nodes, a new fault diagnosis method for yaw damper is proposed, combining Variational Mode Decomposition (VMD) optimized using the Black Kite Algorithm (BKA) with Graph Convolutional Networks (GCN) incorporating Convolutional Neural Networks (CNN) and Attention Mechanisms (ATT). The effectiveness of the model was validated using multi-sensor lateral vibration acceleration datasets obtained from rolling vibration tests at different speed levels under six fault conditions of yaw dampers. The results show that compared to CNN, CNN-ATT, and GCN, the CNN-ATT-GCN model consistently achieves a diagnostic accuracy of 99.9% across all speed levels, demonstrating higher accuracy and stability in diagnosis. These results provide a high-precision and widely applicable fault diagnosis method for the development of intelligent maintenance technology for yaw dampers.