The lightweight segmentation model HMARU-net was proposed in order to address the issues of low detection accuracy, slow detection speed, and poor interference resistance in existing subway tunnel lining water leakage detection models. HC-MobileNetV3 combined with self-calibrating convolution was employed as the backbone feature extraction network, achieving lightweight modeling while enhancing multi-scale feature extraction capability. The partial attention convolutional aggregation network was designed to enhance global information modeling and complex detail feature extraction through attention mechanism and residual structure. A residual module (RAEPC Block) was incorporated into the decoder to reduce computational demand while improving segmentation accuracy and interference resistance. An Attention Gate introduced in the skip-connection layer effectively mitigated semantic discrepancy between encoder and decoder. The experimental results demonstrated that HMARU-net achieved the mean intersection over union of 86.0%, mean pixel accuracy of 93.07%, and accuracy of 98.33%. Model complexity was significantly reduced, with only 3.134×10⁶ parameters and 6.872×10⁹ computations, enabling image processing speed of 78.967 frames per second. HMARU-net significantly improved detection accuracy and efficiency while offering strong lightweight advantage compared with other mainstream semantic segmentation models.