A Dual-Branch Deep Learning Framework for Macroscopic Area and Microscopic Contour Prediction of EMU Wheelset Wear

Abstract

Accurately predicting the nonlinear degradation trajectories of electric multiple-unit (EMU) wheelset treads is crucial for ensuring dynamic safety and optimizing predictive maintenance. However, existing data-driven methods struggle to simultaneously quantify macroscopic material loss and reconstruct microscopic worn geometry under complex operating conditions. To address this limitation, this paper proposes a dual-task prediction framework based on multi-source state features and a parallel dual-branch neural network. First, Chebyshev polynomials are used to fit tread profiles accurately and extract the macroscopic equivalent cross-sectional wear area. Subsequently, a 47-dimensional feature space is constructed exclusively from information available before the predicted re-turning cycle, including operating exposure, geometric baseline and measurement-quality indicators, individual historical wear areas, individual historical wear rates, homologous-wheelset historical wear areas, homologous-wheelset historical wear rates, and polygonization-state descriptors. Based on this feature system, the Area Prediction Network (APN) branch combines a Transformer and a one-dimensional convolutional neural network (1D-CNN) to capture global dependencies and local feature combinations for scalar area estimation, whereas the Tread Contour Prediction Network (TCPN) branch uses an asymmetric autoencoder-style multilayer perceptron to map the same 47-dimensional state features to high-dimensional geometric coordinates. Validation using actual re-turning records collected from 2019 to 2025 for powered and trailer wheelsets demonstrates competitive and robust prediction performance. The APN achieves a coefficient of determination (<inline-formula> <tex-math notation="LaTeX">$R^{2}$ </tex-math></inline-formula>) as high as 0.950 and a mean absolute error (MAE) as low as <inline-formula> <tex-math notation="LaTeX">$3.128{\,}\mathrm {mm}^{2}$ </tex-math></inline-formula>. Concurrently, the TCPN reconstructs tread profiles with an <inline-formula> <tex-math notation="LaTeX">$\mathrm {MAE}_{\mathrm {contour}}$ </tex-math></inline-formula> below <inline-formula> <tex-math notation="LaTeX">$0.15{\,}\mathrm {mm}$ </tex-math></inline-formula> in most operating scenarios and reduces geometric deviations in critical high-curvature regions, such as the flange root. Furthermore, a feature-ablation study shows that historical wear-rate features provide important degradation-gradient information that improves prediction stability for the top 15% of severe-wear samples. This study links macroscopic wear quantification with microscopic morphological reconstruction and provides a data-driven basis for EMU wheelset maintenance decision-making.