In this study, we attempted to construct a machine learning model to estimate the damage state of wooden houses using response acceleration time histories. For the reference model of wooden houses, a natural period of 0.5 s and a damping ratio of 5% were assumed. Using Optuna for hyperparameter tuning, we evaluated three models: LSTM networks, 1D CNNs, and a combined 1D CNN+LSTM model. The 1D CNN model demonstrated the highest accuracy, achieving over 83%. During the training process of the 1D CNN, there was a tendency towards overfitting. However, by applying regularization and retraining, we were able to suppress overfitting and increase the accuracy to 91.7%, confirming the effectiveness of regularization. Furthermore, to verify the generalization capability of the damage estimation model, we created datasets with different natural periods and damping constants and conducted tests. The accuracy significantly decreased across all datasets. Particularly, when the natural period changed, the accuracy dropped drastically, indicating that differences in the natural period greatly affect the local features in the response acceleration time histories. Hence, transfer learning was conducted on the model of wooden house with natural period of 0.4 s and damping ratio of 4%. Transfer learning could enhance the generalization capability of the damage estimation model. Based on these findings, it is considered effective to pre-determine the natural period and damping ratio of the building where the accelerometer is installed, and then create additional datasets accordingly and tune the model through retraining.