Autoencoder-based compression and reconstruction of railway tunnel vibration data: a comparative study

Abstract This paper provides a proof of concept comparative study of four autoencoder architectures for compressing acceleration signals collected from a real railway tunnel under realistic and noisy conditions. The evaluated models include a Single-layer Sparse Autoencoder (SAE), a Fully Connected Autoencoder (FC), a Convolutional Autoencoder (CAE), and a hybrid architecture that combines convolutional (CNN) and Long Short-Term Memory (LSTM) layers. All models are trained and tested on segmented time-series acceleration data acquired under real operating conditions, including measurement noise and operational variability typical of real-world tunnel structural health monitoring (SHM) signals. The performance of the models is evaluated is evaluated using multiple quantitative metrics, including Mean Absolute Error (MAE), Coefficient of Determination (R $$^{2}$$ ), Peak Signal-to-Noise Ratio (PSNR), and Cepstral Distance. A composite objective based on these metrics is employed to guide Bayesian hyperparameter optimization to select the optimal configurations. The best-performing model was a CNN-LSTM configuration. On average, CNN-LSTM models excelled at preserving temporal features (mean: MAE $$=$$ 0.38, R $$^{2} = $$ 0.73, PSNR $$=$$ 16.33 dB), while the SAE better preserved spectral characteristics (mean CD $$=$$ 0.72). These results demonstrate the feasibility of autoencoder-based compression for tunnel vibration data and highlight the importance of architecture selection according to SHM objectives.