Applications of AI and ML in WR dynamic interaction
| No. | Purpose | Methodology | Input | Output | Findings | Limitations |
|---|---|---|---|---|---|---|
| 1 | Construction of surrogate models for nonlinear contact | Artificial neural network (ANN) (Sun et al., 2025a, b, c) Deep neural network (DNN) (Sahoo, Singh, & Racherla, 2025) Elliptical basis function (EBF) (Kwak & Lee, 2025) Physics-informed neural network (PINN) (Xia & Meng, 2024) | Normal force, creep coefficient, WR profiles, and Hertz/Kalker parameters | Tangential force, stress distribution, and adhesion coefficient | Reduced computational time compared with CONTACT; integration into MBD | Dependent on training data; limited extrapolation capability beyond the data domain |
| Convolutional neural network (CNN) (Mo, Wang, & Li, 2025) | Stress fields represented as images/contact maps | 2-dimensional stress distribution | Suitable for field-based data | Requires large datasets | ||
| 2 | Prediction of the hunting critical speed | ML (Folorunso et al., 2023) Random Forest (Baldyga, Barański, Belter, Kalinowski, & Weichbroth, 2024; Mao, He, & Jiang, 2024) | Suspension parameters, rail stiffness, friction coefficient, and dynamic simulation results | Critical speed value | Easy to implement; interpretable | Difficult to model strongly nonlinear relationships |
| You Only Look Once (YOLO) (Li, Ling, & Zhai, 2026) | Multi-parameter dynamic time-series data | Critical speed/vibration amplitude | Capability to capture strong nonlinear behavior | Requires large datasets | ||
| Two-dimensional convolutional neural network-gated recurrent unit (2DCNN-GRU) (Shuangxi, 2023) | Vibration equations + experimental data | Instability threshold | Preservation of physical constraints | Complex training architecture | ||
| 3 | Estimation of long-term wear evolution | Long short-term memory (LSTM) (Jin & Kang, 2024; Yang, Kaynardag, Sohn, & Salamone, 2026; Yang et al., 2024) Recurrent neural network (RNN) (Singh, Das, Singh, & Racherla, 2023) | Load sequence, micro-slip, operating cycles | Wear depth over time | Suitable for time-series data | Requires long-term datasets |
| Hybrid model (Archard + ML-based correction) (Liu, Song, Xu, & Yu, 2025a) | Experimental wear data | Wear correction coefficient | Physics-data integration | Requires periodic calibration | ||
| Fast simulated annealing algorithm-Gaussian process regression (FSA-GPR) (Yu, Su, Fan, & Dong, 2019) | Accumulated frictional energy | Predicted wear distribution | Prediction with uncertainty quantification | Computational cost increases with dataset size | ||
| 4 | Frequency spectrum analysis and early hunting detection | Fast Fourier transform (FFT) + ML classification (Cahya et al., 2025) | Acceleration signals and bogie vibrations | Stability indices, early warning | Real-time implementation | Sensitive to noise |
| CNN on the frequency spectrum (Li, Wan, Tsang, Hui, & Chung, 2023) | Time-frequency spectral map | Identification of vibration modes | Automatic feature extraction | Requires labelled data | ||
| 5 | Unsupervised anomaly detection | Autoencoder (AE) (Li et al., 2025d) | Normal vibration signals | Abnormal deviations | No labelled data required | Difficult to identify the root cause |
| Isolation forest (IF) (Zuo, Thiery, Chandran, Odelius, & Rantatalo, 2022) | Force data, vibration | Anomaly points | Fast computation | Assumes a random distribution | ||
| One-class support vector machines (OC-SVM) (Ghiasi, Khan, Sorrentino, Diaine, & Malekjafarian, 2024) | Baseline stable data | Deviation alert | Easy to implement | Sensitive to kernel parameters | ||
| 6 | Detection of WR geometry changes | CNN + image processing (Zhang, Wang, Yang, & Yao, 2025d; Haghbin et al., 2024) | 2-dimensional/3-dimensional profile scanning data | Profile deviation, wear zones | High accuracy | Requires geometric sensors |
| Graph convolutional network (GCN) (MajidiParast, Neamatian Monemi, & Gelareh, 2025) | Discrete geometric data | Prediction of geometric evolution | Suitable for complex structures | Not yet widely adopted in railway applications | ||
| 7 | Supports predictive maintenance | LSTM + Remaining useful life (RUL) prediction (Huang, Gong, Zhang, Wang, & Yuan, 2024; Fernandes, Fernandes, & Borsato, 2025) | Historical operational data | Remaining time before reprofiling | Decision support | Dependent on data reliability |
| Reinforcement learning (RL) (Saleh, Remenyte-Prescott, Prescott, & Chiachío, 2024) Deep reinforcement learning (DRL) (Mohammadi & He, 2022) | Cost-risk data | Optimal maintenance strategy | LCC optimisation | Requires a reliable simulation environment |
| No. | Purpose | Methodology | Input | Output | Findings | Limitations |
|---|---|---|---|---|---|---|
| 1 | Construction of surrogate models for nonlinear contact | Artificial neural network (ANN) ( | Normal force, creep coefficient, WR profiles, and Hertz/Kalker parameters | Tangential force, stress distribution, and adhesion coefficient | Reduced computational time compared with CONTACT; integration into MBD | Dependent on training data; limited extrapolation capability beyond the data domain |
| Convolutional neural network (CNN) ( | Stress fields represented as images/contact maps | 2-dimensional stress distribution | Suitable for field-based data | Requires large datasets | ||
| 2 | Prediction of the hunting critical speed | ML ( | Suspension parameters, rail stiffness, friction coefficient, and dynamic simulation results | Critical speed value | Easy to implement; interpretable | Difficult to model strongly nonlinear relationships |
| You Only Look Once (YOLO) ( | Multi-parameter dynamic time-series data | Critical speed/vibration amplitude | Capability to capture strong nonlinear behavior | Requires large datasets | ||
| Two-dimensional convolutional neural network-gated recurrent unit (2DCNN-GRU) ( | Vibration equations + experimental data | Instability threshold | Preservation of physical constraints | Complex training architecture | ||
| 3 | Estimation of long-term wear evolution | Long short-term memory (LSTM) ( | Load sequence, micro-slip, operating cycles | Wear depth over time | Suitable for time-series data | Requires long-term datasets |
| Hybrid model (Archard + ML-based correction) ( | Experimental wear data | Wear correction coefficient | Physics-data integration | Requires periodic calibration | ||
| Fast simulated annealing algorithm-Gaussian process regression (FSA-GPR) ( | Accumulated frictional energy | Predicted wear distribution | Prediction with uncertainty quantification | Computational cost increases with dataset size | ||
| 4 | Frequency spectrum analysis and early hunting detection | Fast Fourier transform (FFT) + ML classification ( | Acceleration signals and bogie vibrations | Stability indices, early warning | Real-time implementation | Sensitive to noise |
| CNN on the frequency spectrum ( | Time-frequency spectral map | Identification of vibration modes | Automatic feature extraction | Requires labelled data | ||
| 5 | Unsupervised anomaly detection | Autoencoder (AE) ( | Normal vibration signals | Abnormal deviations | No labelled data required | Difficult to identify the root cause |
| Isolation forest (IF) ( | Force data, vibration | Anomaly points | Fast computation | Assumes a random distribution | ||
| One-class support vector machines (OC-SVM) ( | Baseline stable data | Deviation alert | Easy to implement | Sensitive to kernel parameters | ||
| 6 | Detection of WR geometry changes | CNN + image processing ( | 2-dimensional/3-dimensional profile scanning data | Profile deviation, wear zones | High accuracy | Requires geometric sensors |
| Graph convolutional network (GCN) ( | Discrete geometric data | Prediction of geometric evolution | Suitable for complex structures | Not yet widely adopted in railway applications | ||
| 7 | Supports predictive maintenance | LSTM + Remaining useful life (RUL) prediction ( | Historical operational data | Remaining time before reprofiling | Decision support | Dependent on data reliability |
| Reinforcement learning (RL) ( | Cost-risk data | Optimal maintenance strategy | LCC optimisation | Requires a reliable simulation environment |
Sharing content requires targeting cookies to be enabled. Please update your cookie preferences to use this feature.