Sample-adaptive dual-representation fusion with statistical alignment for cross-radial-load rolling bearing fault diagnosis

Abstract

Cross-radial-load rolling bearing fault diagnosis is affected by load-induced changes in impact morphology, spectral energy distribution and class boundaries. The relative diagnostic contributions of raw time-domain responses and CWT-based time-frequency energy representations vary across samples and fault states. To address this problem, a sample-adaptive dual-representation fusion network, termed ATFF-Net, is proposed. For each vibration segment, the raw sequence retains impact responses and periodic modulation, while the continuous wavelet transform (CWT) image describes the local time-frequency energy distribution. The two representations are extracted through parallel branches and projected into a common feature space. A contribution fusion module estimates their weights from branch responses, inter-representation discrepancy and feature interaction. Statistical alignment is imposed on the fused feature space to reduce source-target distribution discrepancy across radial loads. Experiments were conducted on the Paderborn University bearing dataset, the Politecnico di Torino DIRG bearing dataset and an independently built bearing fault test rig. On the PU cross-radial-load task, ATFF-Net achieved 97.98 ± 0.15% accuracy and 97.99 ± 0.15% macro-F1 over five independent runs, yielding the highest mean performance among the compared methods. On the two DIRG transfer tasks, the highest mean accuracies were 98.91 ± 0.21% and 98.71 ± 0.14%, respectively. Independent-batch validation yielded 98.00% accuracy and 98.00% macro-F1. Ablation and feature analyses showed that sample-adaptive fusion adjusts representation contributions across samples, while fused-space statistical alignment reduces cross-load feature discrepancy. ATFF-Net provides an effective approach to cross-radial-load rolling bearing fault diagnosis.