Accurate segmentation of surface defects on concrete bridges is crucial for bridge maintenance and safety assessment, yet it remains challenging due to multiple defect types and complex background interference. To solve this issue, this paper proposes a mamba attention UNet (MAUNet). MAUNet combines the innovative mamba attention encoding (MAE) and mamba attention upsampling decoder (MAUD) modules, which can significantly improve the accuracy and robustness of multi-defect segmentation. Specifically, the MAE module enhances the modeling ability of global context information and direction-sensitive features in the encoding stage, while the MAUD module integrates deep semantic information and shallow spatial detail information in the decoding stage. Through the experimental validation of the concrete bridge surface defect segmentation datasets, MAUNet improves 6.88% and 6.49% in mean Pixel Accuracy and mean Intersection over Union metrics, respectively, compared to the benchmark network UNet, which proves its superior performance in complex backgrounds and multi-defect scenarios.