Although data-driven methods are widely used in fault diagnosis, traditional one-dimensional feature analysis strategies still have bottlenecks in feature utilization and representation capabilities. Although previous studies have attempted to utilize convolutional neural networks by converting one-dimensional signals into two-dimensional images, such conversion methods are often accompanied by new problems such as information loss, complex computation, and poor feature separability. To break through the above limitations, this paper innovatively proposes a Multi-dimensional Fused Image and Temporal Dual-branch Fusion Network model (ITDFNet). This model adopts a dual-path collaborative architecture, aiming to capture and integrate complementary information of faults from different dimensions respectively: The image branch fuses the original signal matrix, the temporal gram and the structural gram signal matrix into a multi-dimensional image, and focuses on the key spatial texture features by using a dedicated dual fusion attention block. The temporal branch models the original sequence enhanced by Gram denoising to retain the precise time-varying law. Under the CWRU and UCONN datasets, this method achieved a 100% fault identification accuracy rate, outperforming existing methods. Meanwhile, compared with Swin-FFRN, under strong noise interference, its accuracy rate is increased by 8.44% and the number of parameters is reduced by 29M, highlighting the enhancing effect and application prospects of the dual-branch fusion design on the robustness of the model.
Cai et al. (Tue,) studied this question.