An Intelligent Rail Defect Classification Framework Based on Transfer Learning

• Development of rail SDM system that generates vibration signals from different rail track samples. • Formation of field dataset consisting of vibration defect signals collected from the maintenance record of a southeast Asian railway company. • Designed a framework that integrates rail SDM with TrL-CNN, offering intelligent classification of rail track defects even when limited field data is available to train the model. • Comparison of classification accuracies between time and frequency domain vibration signals. Intelligent classification of railway track defects is necessary for ensuring safe and reliable rail operations. Supervised learning algorithms (SLAs) have shown strong potential for fully automated defect detection and classification. However, their application in railway systems remains constrained due to lack of labelled field data. Compared to purely physics-based models, scaled-down physical models (SDMs) can reproduce real-world rail conditions and generate large datasets that reflect the dynamic behavior of railway systems. In parallel, transfer learning (TL) has proven effective when the source and target domains are closely related. Therefore, this research proposes a rail track defect classification framework that integrates a rail SDM with a TL-based approach. A rail SDM coupled with a vibration measurement system was developed, and a database consisting of healthy and defective rail track signals was created. A convolutional neural network CNN was trained to classify the SDM signals in both time and frequency domains, with the frequency-domain model achieving 99.15% testing accuracy. However, its performance degraded significantly on field-acquired signals. To address this, the TL approach was implemented, where a pre-trained model on SDM was fine-tuned with limited field signals using few-shot learning. The resulting transfer learning-based CNN (TrL-CNN) achieved 98.31% accuracy on frequency domain signals. Comparative evaluation against baseline models, including deep neural networks (DNN) and residual networks (ResNet), shows that TrL-CNN delivers higher accuracy (98.31%), narrower confidence intervals (±0.3%), and shorter training time (∼16 mins) than ResNet (∼22 mins). These results confirm the proposed methodology’s optimal balance of accuracy, robustness, and computational efficiency.