Multi-Dimensional Prediction of Deformation in Track-Steel Composite Structures Under Multi-Excitation Noise Sources

The deformation monitoring data of track-steel composite structures， collected in complex construction environments， are often contaminated by strong noise from various excitation sources such as equipment vibration， soil disturbance， and load variation. This noise obscures the underlying trends in the deformation time-series data， leading to low prediction accuracy with traditional models. To address this issue， this study proposes a novel deformation prediction method based on SVMD-BiLSTM-BO. First， Sequential Variational Mode Decomposition （SVMD） is applied to decompose the data into modal components. The Pearson correlation coefficient is then used to identify trend changes in the historical data， effectively suppressing the impact of noise. Since the traditional Bidirectional Long Short-Term Memory （BiLSTM） network is prone to falling into local optima due to its hyperparameter sensitivity， Bayesian Optimization （BO） is introduced to optimize these hyperparameters， enabling the model to accurately capture the deformation trends of the track. Experimental results showed that the proposed SVMD-BiLSTM-BO model outperformed the benchmark models in prediction accuracy. The absolute errors of deformation in the horizontal， vertical， and settlement directions were 0.0567 mm， 0.0899 mm， and 0.0445 mm， respectively. Furthermore， the model demonstrated a certain level of generalization capability in cross-point prediction tasks， confirming its accuracy and effectiveness in predicting railway track deformation.