Abstract
This paper presents a robust multi-stream hybrid convolutional neural network (CNN) and Transformer architecture for WEC gearbox fault diagnosis. To suppress severe background interference, an adaptive preprocessing method is developed by coupling Bayesian-optimized variational mode decomposition (VMD) with a windowed short-time Fourier transform to enable the reconstruction of task-relevant components and informative time-frequency representations. Within the diagnostic backbone, the CNN extracts localized morphological transients associated with gearbox defects, whereas a Transformer encoder captures global and long-range dependencies in the sequence representations. A channel-attention module is further employed to recalibrate discriminative feature maps, and a multi-stream fusion strategy integrates deep time-frequency features with statistical and envelope descriptors to enhance complementary fault evidence. A WEC gearbox testbed is built to acquire real-world vibration signals under multiple fault states. Experimental results demonstrate the superiority of the proposed framework, achieving a 99.14% diagnostic accuracy under nominal conditions. Comparative studies verify that the model maintains exceptional robustness, yielding an 89.66% accuracy even under severe interference, thereby providing a highly accurate and robust diagnostic solution for WECs deployed in complex marine environments.