Abstract: To address the non-stationarity of rolling bearing vibration signals and the limitations of traditional diagnostic methods in extracting weak fault features, this study proposes a hybrid diagnostic model that integrates variational mode decomposition (VMD) with a Transformer-BiLSTM-CrossAttention network. First, VMD is employed to adaptively decompose the raw vibration signals into multi-scale intrinsic mode function (IMF). A multi-domain feature set is then extracted, combining time-domain statistical features, frequency-domain envelope spectrum energy, and time-frequency wavelet entropy. Subsequently, a deep learning framework is designed to fuse global temporal modeling with local feature interaction: a Transformer encoder captures long-range dependencies, a bidirectional long short-term memory (BiLSTM) network extracts bidirectional local temporal features, and a cross-attention mechanism dynamically integrates their outputs to enhance the model’s focus on fault-sensitive characteristics. The model is validated experimentally using the Case Western Reserve University bearing dataset. Results demonstrate that under varying load conditions (0~3 HP), the proposed method achieves an average diagnostic accuracy of 99.61% in classifying normal states, inner race faults, outer race faults, and roller element faults. This approach provides a high-precision and generalizable solution for bearing fault diagnosis under complex operating conditions.