An intelligent identification method of railway wheel defect based on the improved deep convolutional generative adversarial network

An intelligent wheel defect identification method based on an improved Deep Convolutional Generative Adversarial Network (DCGAN) is presented. A generator is developed to integrate wheel-rail strain responses with axle box vibrations, producing physically consistent augmented data. The architecture is refined via nearest‑neighbour interpolation and channel attention to eliminate artefacts and enhance feature sensitivity. A Wasserstein distance with a gradient penalty is implemented to ensure stable adversarial training. Experimental verification is conducted through coupled dynamics simulations and a 1:5 scaled test rig. High-fidelity datum is generated to address data scarcity in railway maintenance. The obtained results demonstrate that a superior identification accuracy of 97.2% is achieved. The proposed approach is found to be in good agreement with the predictions of the mathematical model and experimental results. The proposed method achieves significantly better wheel defect recognition performance than traditional models in a small-sample environment.