BoltResvit: an enhanced residual vision transformer for robotic-assisted nondestructive railway bolt looseness monitoring

Railway bolt looseness threatens structural integrity and operational safety, necessitating robust automated monitoring methods. However, limited availability and severe class imbalance of bolt image data from practical railway environments greatly restrict the training effectiveness and generalisation capability of deep learning models. This paper proposes a robotic-assisted non-destructive monitoring approach termed BoltResViT, integrating generative adversarial network (GAN)-based data augmentation and a residual vision Transformer (ResNet-Transformer) model. StyleGAN2-ADA generates high-fidelity synthetic samples validated by Fréchet inception distance (FID) and expert assessment, constructing a balanced training dataset. The ResNet-Transformer model integrates local spatial feature extraction capabilities of ResNet18 with the global contextual modelling ability provided by Transformer-based multi-head self-attention. Furthermore, a channel attention module and auxiliary monitoring branch are incorporated to enhance feature discrimination and model robustness. A dual-supervision mechanism combining primary and auxiliary monitoring branches ensures accurate bolt looseness monitoring. Experimental results demonstrate that the proposed BoltResViT achieves monitoring accuracy of 99.28%, showing excellent generalisation and practical applicability in railway bolt monitoring scenarios characterised by limited data.