A Novel Rubber Composite Sleeper-Deformation-Prediction Model Based on Response Surface Method (RSM) and Machine Learning (ML) Techniques

Rubber composite sleepers can experience significant temperature variations in service, causing temperature-induced deformation. Real-time monitoring of this deformation is crucial for operational safety and maintenance; however, it is costly, time-consuming, and requires substantial resources and personnel. Developing temperature-dependent predictive models offers a cost-effective and efficient alternative, providing accurate insights into sleeper behaviour under different conditions while saving time, labour, and materials. This study attempts to develop a novel deformation model of rubber composite sleepers using response surface methodology (RSM) and machine learning (ML) techniques. Platinum temperature (Pt) sensors, embedded at various points on a full-scale rubber composite sleeper model, were used to measure both the sleeper temperature field and ambient temperature in real-time at 30-minute intervals over the period of a year. Simultaneously, lateral deformation was recorded using linear variable differential transducer (LVDT) displacement sensors. The temperature data were filtered to remove noise and normalized based on the Log-Pearson Type III outlier detection method and Box-Cox transformation, respectively, before being used to develop temperature-dependent models for sleeper deformation. To ensure accurate ML predictions, the dataset was split into 70% for training and 30% for testing. Model performance was evaluated using the correlation coefficient (R2), mean square error (MSE), root means square error (RMSE), and mean absolute error (MAE). The analysis revealed that the sleeper’s body temperature closely follows the changing trend of the ambient environment. Also, like any polymer material, the rubber composite sleeper expands when it absorbs heat from sunlight and contracts as it cools when sunlight intensity decreases, potentially reversing much of the deformation. The K-nearest neighbour algorithm outperformed the RSM and other ML techniques with R2, MSE, RMSE, and MAE values of 0.999, 0.000258, 0.016, and 0.000896, respectively. The developed model can serve as an important reference for monitoring lateral deformation for safety and maintenance purposes.