LNFformer: an enhanced transformer-based network for intelligent fault diagnosis

Accurate fault diagnosis of rolling bearings under strong noise conditions remains challenging, as noise can significantly degrade feature quality. Although data-driven methods such as convolutional neural networks and Transformer-based models have shown strong feature extraction capabilities, their performance often deteriorates in low signal-to-noise ratio environments due to insufficient modeling of global correlations and loss of subtle fault-related information. To address these limitations, this paper proposes a dual-branch fusion diagnostic model, termed LNFformer. The model integrates a local normalization mapping (LNM) branch to preserve structural information and a Gramian-based correlation-enhanced representation (LNR) branch. A patch-based Transformer is employed to capture long-range dependencies, while an invertible neural network is introduced to retain detailed information. In addition, a bridging fusion module is designed to effectively integrate complementary features. Experiments on the CWRU and PU datasets demonstrate that the proposed method achieves superior accuracy and robustness compared with existing methods, particularly under strong noise conditions.