Multiple data analysis-based approaches have proven highly effective across diverse domains and transfer tasks in identifying bearing failures. However, current deep learning-centric feature extraction techniques often neglect the exploration and incorporation of prior knowledge. In light of this, this paper introduces a fault diagnosis method for bearings under variable operating conditions, leveraging deep learning and feature fusion. This method combines data-driven features with time–frequency domain features, seamlessly embedding them within the deep learning network framework to enhance the model’s generalization capabilities. The extracted features are then fed into the joint distribution alignment mechanism, which simultaneously adjusts the edge distribution and conditional distribution in the source and target domains. In order to verify the effectiveness of the proposed method, we validate the fault data under different operating conditions in the Case Western Reserve University (CWRU) bearing fault dataset, and the experimental results show that the generalization and diagnostic performance of the proposed method in this paper have achieved better results.