Conventional deep learning-based bearing fault diagnosis methods typically assume that the training and testing datasets adhere to the same distribution. However, this assumption often fails in real-world industrial settings, leading to substantial degradation in diagnostic accuracy. A critical challenge in existing models is their lack of generalizability—while fault detection rates remain high under a single operational condition, these models struggle to maintain performance across varying real-world environments. To address this limitation, we propose a novel pre-training–fine-tuning framework tailored to bearing fault diagnosis, capable of accommodating diverse loading conditions. This framework integrates the pre-training–fine-tuning methodology with attention-embedded quadratic convolutional neural networks (QCNN), facilitating robust feature extraction and significantly enhancing fault diagnosis performance for one-dimensional time-series signals. Our approach not only accelerates model convergence but also mitigates the detrimental effects of load variation on diagnostic accuracy. Extensive experimental evaluations demonstrate that the proposed method surpasses existing models in fault classification tasks under diverse loading conditions, achieving superior accuracy, robustness, and fault diagnosis capabilities. This research provides a promising solution for real-world bearing fault diagnosis and offers substantial economic potential for large-scale industrial applications.