Cross-condition rolling bearing fault diagnosis via channel-wise adaptive decomposition and gated time-frequency fusion.

Abstract

Cross-condition bearing diagnosis remains challenging because speed-load shifts and heterogeneous sensor responses can weaken fault signatures. This study proposes a measurement-driven framework combining channel-wise adaptive decomposition and dual-branch time-frequency fusion with gating for multi-channel bearing signals. Channel-wise VMD optimized by the Improved Whale Optimization Algorithm (IWOA-VMD) enhances fault-sensitive components, while temporal impact features and envelope-spectrum modulation features are jointly recalibrated for classification. The evaluation on measured data covers nine source-to-target tasks across discrete speed-load conditions. The proposed method reached 90.20% mean task-level accuracy, numerically higher than Raw 1D-CNN (87.12%) and WDCNN (83.02%); significance was observed for WDCNN only. Supplementary public-dataset results further support its applicability under different acquisition settings.