Real-Augmented Mixed Training for Wheel-Flat Geometry Classification under Data Scarcity on a Laboratory-Scale Test Rig

Abstract

Accurate defect classification requires sufficient data; however, early-stage studies often face data scarcity when acquiring data through field measurements. This paper proposes a real-augmented mixed training strategy on a 3D-printed, laboratory-scale railway test rig for wheel-flat geometry classification. Limited real axle-box vibration data are expanded with augmented data generated by a long short-term memory (LSTM) model trained in a reconstruction-based manner. This produces effective variations that preserve impulsive events relevant to wheel-flat responses. A convolutional neural network (CNN) classifier is evaluated under three data compositions: real-only, mixed real+augmented, and augmented-only training and validation with real testing. We found that mixed training improves test accuracy and reduces inter-defect confusion, reaching 0.979 ± 0.006 with 200 real and 1,200 augmented segments per class, whereas augmentedonly training shows degraded generalization to real test despite high validation accuracy. The proposed strategy reduces the required experiments for robust classification under data scarcity.