Rolling bearings operating under coupled variable-load and time-varying-speed (VL-TV) conditions generate vibration signals with pronounced non-stationarity, speed-dependent impulse spacing, and strong noise contamination, which weaken both decomposition-based and purely data-driven diagnosis methods. To address this issue, this study proposes a physics-informed end-to-end diagnosis framework that integrates Multi-Constraint Improved Variational Mode Decomposition (MC-IVMD) with a Dual-Branch Cross-Layer Fusion Sparse Attention Network (CLSAM++). MC-IVMD introduces fault-mechanism-guided frequency priors, impact sparsity, modal orthogonality, and multi-scale residual feedback to determine the modal number adaptively and extract fault-sensitive intrinsic mode functions with reduced mode aliasing. CLSAM++ combines a local feature branch and a temporal dependency branch based on gated sparse-attention LSTM to jointly model impulsive local signatures and long-range temporal dependencies. The core novelty lies in coupling the physics-informed decomposition loss and bearing-fault physical consistency constraints with the classifier training objective, thereby enabling mutual optimization between signal decomposition and deep feature learning rather than a conventional serial pipeline. Experiments on the University of Ottawa VL-TV bearing dataset and a synthetic CWRU VL-TV dataset demonstrate that the proposed framework achieves 98.9 % accuracy on the 12-condition Ottawa task and 97.2 % mean accuracy on the synthetic CWRU VL-TV task, while remaining statistically superior to competing models in repeated trials ( p 0.05). These results indicate that the proposed framework is accurate, robust, and promising for real-time intelligent monitoring of rotating machinery.
Fan et al. (Sun,) studied this question.