ARAM-Swin: a hybrid architecture with an adaptive residual attention module and Swin Transformer for rolling bearing fault diagnosis

Abstract Existing deep learning methods struggle to simultaneously capture global temporal dependencies and local transient impulses of rolling bearing fault signals, while static attention mechanisms often fail to adapt to complex industrial noise. To address these challenges, this paper proposes ARAM-Swin, a dual-branch hybrid diagnostic architecture. Firstly, a composite Gramian angular field encoding strategy is employed to fuse dynamic and static signal features into multi-view images. Secondly, a parallel extraction framework is constructed: a Swin Transformer branch models long-range dependencies, while a convolutional neural network branch embedded with an adaptive residual attention module (ARAM) extracts fine-grained fault textures. Notably, ARAM utilizes a learnable residual scaling factor to dynamically balance feature enhancement and information preservation. Experimental results on the Case Western Reserve University and XJTU-SY datasets demonstrate that ARAM-Swin delivers superior diagnostic efficacy under multi-condition, cross-domain, limited-sample, and strong-noise scenarios, achieving fault classification accuracies of 99.98% and 99.89%, respectively.