Traditional graph neural networks often encounter limitations in fault diagnosis due to insufficient feature extraction at a single scale, particularly in complex operational scenarios. To overcome these challenges, we introduce an innovative multi-scale graph Transformer framework for rolling bearing fault diagnosis. This framework incorporates a distinctive multi-scale feature aggregation mechanism, along with centrality and spatial encoding for graph nodes, to enhance structural insights. Leveraging multi-head self-attention, our approach efficiently extracts and learns fault features, thereby significantly improving fault identification. Extensive experiments on the designed bearing dataset, as well as a customized rolling bearing apparatus, validate the efficacy of our method. Our model achieves a peak diagnostic precision of 99.5% and maintains an average accuracy exceeding 97.9%, underscoring its robustness and adaptability across diverse operational scenarios.