Cross-Condition Vibration-Based Fault Diagnosis of Rolling Bearings Using Dual-Branch Feature Fusion and MK-MMD Domain Alignment

A cross-condition fault diagnosis method is developed for multi-channel vibration measurements. For every 0.5-s segment, Short-Time Fourier Transform (STFT) spectrograms and Fast Fourier Transform (FFT) spectral sequences are generated from the same four-channel observation. A convolutional neural network and a temporal convolutional network encode local time–frequency textures and global spectral dependencies, while a sample-adaptive gate fuses the two representations. Multiple Kernel Maximum Mean Discrepancy (MK-MMD) is then applied only in the common fused space to reduce load-induced distribution shift without target labels, pseudo-labels, or adversarial optimization. Experiments cover all three many-to-one transfer directions among 0, 2, and 4 Nm. The proposed model obtains accuracies of 87.62 %, 85.71 %, and 70.59 %, and macro-F1 scores of 89.24 %, 76.00 %, and 72.31 % for T1–T3, respectively. Against four external domain-adaptation baselines, it achieves the highest task-averaged accuracy (81.31 %) and macro-F1 (79.18 %). Five controlled ablations and class-level visualizations further show that fused-space alignment is the most consistent contributor, while the normal–unbalance ambiguity remains the main limitation under the more difficult transfer directions.