Multi-branch global Transformer‐assisted network for fault diagnosis

Abstract

Fault Diagnosis (FD) is critical in smart manufacturing, enabling predictive maintenance, reducing operational costs, and enhancing system reliability. To deal with this task more accurately, this paper proposes a generative, effective, and novel framework, a multi-branch global Transformer-assisted network (MBGTNet), for accurate FD. First, a multi-branch global-wide one-dimension convolution operation (MBG-WideConv1D) is proposed to obtain global features in different views. Meanwhile, a Transformer assist scheme (TAS) is designed to leverage the Transformer's global feature extraction capacity. The features extracted by the Transformer are fused with those extracted with MBG-WideConv1D by minimizing their pairwise correlation alignment (CORAL) distances. Benefiting from the well-designed MBG-WideConv1D and TAS, the global features hidden in the raw signals are fully extracted from multiple viewpoints. Each branch of global features is then fed into a one-dimension convolutional neural network (1DCNN) to extract local features in a multi-supervised scheme (MSS) that helps each branch learn thoroughly. Furthermore, the proposed method employs a local feature correlation enhancement scheme (LFCS) to reduce distribution differences and increase robustness among the local features of each branch. As a result, the final features used for FD are a fusion of multi-view global and local features with strong robustness, enabling accurate FD in noisy environments. Comparative experiments on four datasets, including CWRU, MFPT, SU Bearing, and SU Gear, validate the proposed method's effectiveness, achieving over 99.6 % accuracy across four datasets. Moreover, the TAS and LFCS's generalities have been demonstrated on two 1DCNNs and hybrid CNN-LSTMs with four subsets. Also, the effectiveness of each component in the proposed framework has been thoroughly analyzed.