Rolling Bearing Fault Diagnosis Based on CKEPM Encoding and Tri-CrossViT

Rolling bearings are indispensable components in mechanical equipment, and their condition directly affects the operational reliability and maintenance costs. To address the requirements of high-precision and robust bearing fault diagnosis, this paper proposes a dynamic-mapping method named CKEPM, based on Complete Ensemble Empirical Mode Decomposition (CEEMDAN), kurtosis-, energy-, and Pearson-correlation-based intrinsic mode function (IMF) selection, followed by Markov Transition Field (MTF) encoding. Furthermore, a triple-scale, cross-branch network Tri-CrossViT is constructed to hierarchically model fault-related features at different resolutions and enable cross-scale information interaction for bearing fault classification. Specifically, vibration signals are first decomposed using CEEMDAN and the fault-sensitive intrinsic mode functions are selected based on kurtosis, energy, and Pearson correlation criteria. Then, these components are encoded by the Markov Transition Field to form RGB time–frequency images. Subsequently, the proposed Tri-CrossViT processes the generated images through three parallel branches with different patch resolutions to capture fault-related information at triple scales, including localized impulsive transients, modulation-related envelope structures, and global spectral energy distributions, and a Cross-branch attention is introduced to fuse complementary fault information across scales, enabling joint learning of local impulses and global dynamic characteristics. The Tri-CrossViT achieves diagnostic accuracies of 98.87% and 99.51% on the Case Western Reserve University (CWRU) and Paderborn University (PU) datasets, respectively, realizing outstanding classification performance. The experimental results demonstrate the effectiveness of the integrated CKEPM-Tri-CrossViT framework for rolling bearing fault diagnosis under the evaluated benchmark conditions.