This study proposes a two-step approach for detecting damaged tethers in submerged floating tunnels. The proposed method employs two different artificial neural network algorithms. First, the long short-term memory (LSTM) autoencoder model trained using response datasets under intact conditions was used to reconstruct the measured acceleration data of the target structure. Further, the data reconstruction error was used as the input for the deep neural network algorithm trained in advance using the reconstruction error pattern in various tether damage cases. The proposed method was verified by conducting a well-validated simulation based on hydrodynamics. The damage-detection accuracy of the proposed method was directly compared with that of a conventional supervised learning algorithm-based approach. In addition, the case study results confirmed that the proposed approach was applicable to other submerged floating tunnel (SFT) structures by retraining the LSTM autoencoder and deep neural network algorithms with intact datasets only. Thus, this approach does not require a large amount of training data or simulation model updates for other SFT structures.