Recognizing water leakage is a crucial task in the daily operation of shield tunnels. Conventional machine learning methods applied in tunnel maintenance rely heavily on labeled images, a process that demands significant time and manual labor. Self-supervised learning (SSL) offers a cost-effective solution by enabling training with a limited amount of labeled data. In this study, we propose a novel SSL model, namely, SSRecNet, to recognize water leakage using a small number of labeled images. Initially, unlabeled images are employed to pretrain the feature extraction process, which is achieved through the restoration of noisy images. Subsequently, a small number of labeled images are utilized to fine-tune the feature extraction and train the classifier. Finally, ablation experiments are conducted to assess the impact of pretraining on enhancing the accuracy of the proposed SSL model. The outcomes demonstrate that SSRecNet attains a commendable accuracy rate of 100% on the training set and 81.75% on the test set. Ablation experiments confirmed that pretraining the feature extraction by SSRecNet can significantly enhance the model’s performance.