Rolling element bearings (REBs) are critical components in rotating machinery, where reliable operation depends on accurate and timely fault diagnosis. This paper introduces a deep transfer learning framework designed to achieve robust cross-domain fault diagnosis across both laboratory and industrial environments. The framework integrates a Stacked Autoencoder (SAE) for hierarchical feature extraction with a Deep Neural Network (DNN) classifier, while leveraging Adaptive Batch Normalization (AdaBN) and selective fine-tuning of the output layer to effectively address domain shifts. The main contribution lies in the combined use of SAE-based feature learning, AdaBN-driven distribution alignment, and limited-sample fine-tuning using small sets of labeled industrial and fixed-condition laboratory data, enabling high diagnostic reliability under diverse operating conditions. To assess the contribution of each component, a baseline version of the model without the fine-tuning stage was also evaluated. The substantial performance degradation observed when testing on unseen target domains confirms the essential role of fine-tuning for achieving robust generalization. The proposed method was validated using laboratory datasets collected under variable and fixed operating conditions, as well as an industrial dataset consisting of four states, including healthy (H), inner race fault (IRF), outer race fault (ORF), and rolling element fault (REF). Experimental results show that the complete framework provides stable and accurate fault classification, achieving 92.65% accuracy on industrial data and 90.91% on fixed-condition laboratory data, consistently outperforming the baseline and conventional deep learning approaches in cross-domain scenarios.