Shuffle-fusion pyramid network for bearing fault diagnosis under noisy environments

Abstract Recent advancements in deep learning have driven the development of big data-driven fault diagnosis techniques. However, traditional models often face significant computational challenges, making them impractical for on-site deployment in rolling bearing fault diagnosis. To address this issue, we introduce the Shuffle-Fusion Pyramid Network (Shuffle-FPN), a novel lightweight fault diagnosis model with a pyramid architecture. Shuffle-FPN enhances fault diagnosis by integrating fault signals across various scales through its pyramid structure, expanding the network’s scope while reducing its depth. The use of depth-wise separable convolutions streamlines network parameters, and channel shuffling ensures comprehensive information fusion across convolutional channels. Additionally, a global representation module compensates for the loss of global context due to increased convolutional depth. These enhancements enable Shuffle-FPN to extract nuanced fault features amidst noise and operate efficiently on devices with limited memory, ensuring real-time fault diagnosis even in complex environments. Rigorous experiments on public dataset from the Paderborn University and our research group’s dataset demonstrate that Shuffle-FPN excels in fault identification under noisy environments and significantly reduces the memory footprint.