Frequency-aware transformer fusion for intelligent fault diagnosis of rolling bearings under variable operating conditions

Abstract Sensor based time-series vibrational signals are difficult to classify under variable operating conditions, particularly during frequency overlapping RPM shifts. The current preprocessing techniques struggle to detect frequency overlapped faults, as their shifting frequency patterns fail to align with the fixed scales. The proposed study designs a frequency-aware transformer fusion, an adaptive hybrid architecture that replace absolute frequency analysis with adaptive pattern recognition on a unified transient energy map. The utilization of continuous wavelet transform and Volterra series form a transient energy map and provides a short-term memory, which replaces the absolute frequency analysis to the pattern recognition. After that, the encoded energy map is processed via a dual block, comprising two concurrent branches: a Kronecker convolutional feature pyramid for repetitive transients and a channel attention MLP block that highlights semantic structures within the transient energy maps. The outputs of these parallel branches are fused through a Swin Transformer module, which uses frequency-aware shifting windows to confidently diagnose frequency overlapped faults. This modular fusion strategy enables the model to learn robust representations across both local and global temporal scales. Experimental validation on multi-class sensor position datasets obtained from Shandong University, Case Western Reserve University and PU demonstrates superiority over state-of-the-art methods. Ablation experiments demonstrate the significance of the dual-branch and dual-attention mechanisms, which enhance diagnostic consistency under frequency overlapped conditions.