Objective As the core component of mechanical equipment rolling bearings suffer from wear and deformation during long-term operation. Addressing the difficulty of existing methods in distinguishing bearing faults a fault diagnosis method based on Manhattan local-global discriminative space learning is proposed. Methods The method has a richer underlying manifold structure and uses the Manhattan distance to reconstruct the original space graph structure. By constructing local intra-class and inter-class graphs it extracts potential discriminative and local information. Additionally global intra-class and inter-class graphs are introduced based on the original global structure enhancing inter-class separability and intra-class cohesion. Firstly feature extraction is performed on the original fault signals to obtain feature testing and training sets. Then these feature training sets are inputted into the Manhattan local-globaldiscriminative space learning model to extract local information global structure and category information from the original space. Next by solving this model an analytical solution for spatial projection can be obtained. Finally the obtained analytical solution for spatial projection is input into a support vector machine along with the feature testing set for fault classification. Results Experimental results demonstrate that the proposed method exhibits excellent performance on the constructed bearing fault platform achieving a final fault recognition rate of 94. 23%. Conclusion The method proposed in this paper shows high recognition accuracy in bearing fault diagnosis marking significant progress in this field and carrying profound implications.