Pantograph Wear Classification via Dual-Backbone Feature-Fusion Ensemble Network

Vision-based pantograph wear recognition plays a critical role in the safety and reliability of railway power supply systems. Although recent studies report promising deep learning-based results, these models solely depend on the integrity of the dataset. Data integrity is a critical yet often overlooked factor in research and production, and neglecting it may lead to inconsistencies and compromised operational safety. In the proposed approach, we demonstrate that a widely used pantograph wear dataset contains severe redundancy and label inconsistencies, including duplicate images appearing within classes and across different wear categories. These issues undermine supervised learning, reduce model generalisation, compromise predictive reliability, and may weaken the safety of rail infrastructure systems. This work (i) preprocesses the dataset by employing MD5-based cryptographic hashing and manual verification, where 626 redundant samples were identified from a dataset of 909 images; subsequently, a manual relabelling procedure is used to correct inherited annotation errors and consistent class definitions. (ii) It devises a Dual-Backbone Feature Fusion Ensemble Network (DBFF-Net) for small and challenging datasets by integrating frozen ShuffleNetV2 and DeiT-tiny as the best individual performing classifiers using various fusion strategies, including concat, weighted sum, Bilinear, Cross-Attention, and Gated. Amongst the different fusion approaches, we obtain the best results with the Gated approach. We reproduced the comparatively improved pantograph wear classification results and conducted extensive experiments to demonstrate that dataset sanitization improves the stability and reproducibility of the model. Moreover, it has been shown that DBFF-Net outperforms individually employed pretrained CNNs and transformer models and achieves an accuracy of 96.46% even with limited but sanitised data.