Identification method for wheel/rail tread defects based on integrated partial convolutional network

Wheelsets are critical components of the running gears in railway vehicles, making damage detection on wheel treads a key focus in train maintenance. However, the lack of standardized sampling equipment and the variability of sampling environments often result in detection datasets that contain underexposed images, which hinder the effective identification of minor tread damages. To address this challenge, an integrated partial convolutional network (I-PCNet) method was proposed for identifying wheel-rail tread defects. This method incorporated a dense underexposure self-correction network (D-SCNet) integrated into the initial layer of the identification model. The aim was to self-correct underexposed samples before feature extraction, thereby revealing more detailed features. Given the difficulties associated with accurately detecting minor wheelset damages, an enhanced adaptive spatial feature fusion (E-ASFF) detection approach was introduced. Furthermore, a partial convolutional network (P-Conv) technique was implemented for the lightweight purpose, resulting in the design of a lightweight backbone network termed FasterNet. Additionally, the design incorporated a novel loss function to improve the integration of the proposed strategies and enhance the model's focus on the samples, with its principles clearly explained. Experiments using actual wheel-rail tread datasets demonstrated that the proposed strategies outperformed traditional tread defect detection algorithms. The effectiveness and generalizability of the proposed method were further validated through comparative experiments, visualization analysis, and generalization tests.