Abstract: Addressing the issue of low model accuracy caused by small-sample, weak-patterned data in dam safety monitoring, this paper proposes a deformation monitoring model (LSTM-TL) that integrates Transfer Learning (TL) with Long Short-Term Memory (LSTM) networks. The model first constructs source-domain and target-domain models using finite element simulation data and measured data, respectively. It then extracts superior parameters from the source-domain model and transfers them to initialize the target-domain model. Finally, it fine-tunes these parameters using measured data to generate the final monitoring model. Engineering validation demonstrates that LSTM-TL effectively suppresses noise interference and overfitting in small-sample data, substantially enhancing model accuracy. Accuracy analysis reveals that pattern similarity (rather than numerical similarity) between source and target domains is the core influencing factor. When the sample size is below 150, LSTM-TL significantly outperforms traditional LSTM; performance converges when samples exceed 150. This research provides a technical support framework for real-time dam safety monitoring and early warning.