Vibration Analysis and Fault Diagnosis of Dynamical System Under Strong Noise Background Based on Transformer-CNN

Dynamical system such as wind turbine gearboxes are subjected to heavy-load, and low-speed operational conditions, frequently generating vibrations with abnormal noise. To effectively diagnose gearbox faults under such conditions of strong noise, this paper proposes a novel fault diagnosis method integrating Transformer and Convolutional Neural Network (CNN) model. First, a Transformer encoder is utilized for the initial extraction of global-level features from raw data with strong noise. The output sequences from the Transformer are subsequently input into the CNN to further extract local features. Then the fault category probabilities are outputted based on the extracted features. Finally, the model was trained, validated, and tested on this dateset comprising 4800 samples from wind turbine gearboxes, categorized into four classes: healthy, bearing fault, high-speed stage fault, and low-speed stage fault. The results demonstrate an overall accuracy of 99.9%, significantly outperforming standalone CNN, Transformer, and ResNet models.