PR-Transformer: Long-term Prediction of Train Wheel Diameters Using Progressive Refinement Transformer

Massive amounts of wheel diameter data can be collected from the trackside wheel detection subsystem, providing a feasible way for developing a data-driven model for wheel diameter prediction. To address this issue, a novel wheel diameter prediction model, named progressive refinement transformer (PR-Transformer), which combines time embedding and data decomposition, is proposed. Time embedding encodes temporal information into the model, offering precise temporal context regarding wheel diameter changes and enabling the model to better understand the potential dependencies between time and wheel diameter wear. Data decomposition partitions the wheel diameter data into seasonal and trend components, simplifying each component and making them more predictable. The PR-Transformer, which adopts an encoder-decoder architecture, is designed to extract long-term dependencies using scaled dot-product attention (SDPA) and multihead attention (MHA) mechanisms. It progressively refines its predictions by integrating information from both the encoder and decoder. Results indicate that the PR-Transformer achieves minimal mean squared error (MSE) and mean absolute error (MAE) across various wheel diameter datasets, surpassing other state-of-the-art models and demonstrating its potential as a robust tool for wheel diameter prediction.