Crack Quantification Across Complex Scenarios For Ballastless Track Slabs Based On Digital Twin And Weakly Supervised Learning

The accurate identification of cracks in ballastless track slabs and the measurement of their geometric features are essential foundations for making maintenance decisions on ballastless tracks.While current machine vision solutions exhibit promising theoretical outcomes,they generally necessitate high-quality,human-annotated benchmarks and ample training samples to ensure the precision of the models.In contrast to typical concrete structures,ballastless track slab inspection images present complex structural backgrounds(including rails,fastener systems,miscellaneous ground markings,and distinct surface textures),poor inspection conditions(nighttime inspection,uneven lighting),and minuscule cracks with exceedingly low pixel ratios.These factors contribute to serious crack missed or crack misdetections in existing methodologies,and deviations from actual measurements.Based on digital twin modeling and weakly supervised deep learning,this study investigates the theories and methods for precise identification and geometric feature quantification of cracks in ballastless track slabs under complex image backgrounds.The research has been successfully applied to the detection of cracks in ballastless track slabs on the Beijing-Shanghai High-Speed Railway,yielding excellent practical results.The main research contents are as follows:(1)A cross-scenario ballastless track slab crack image synthesis dataset and crack status perception model have been established,addressing the challenges of scarce training samples and the visualization of crack status.Through parameterized BIM modeling of ballastless track structures,inverse rendering of actual crack topology features,and simulation of adverse inspection conditions,a dynamically updated digital twin model of ballastless track structures has been obtained,which simulates real inspection scenarios and outputs synthesized crack image datasets.The synthesized dataset expands the training samples required for crack identification algorithms,providing richer and more discriminative image features than traditional data augmentation methods,such as horizontal flipping and color jittering.Furthermore,the crack identification results can automatically update the digital twin model of the track structure,achieving intuitive perception and visualization of the ballastless track slab crack status.(2)A novel image preprocessing method that integrates the Deep Relighting Network(DRN)and Enhanced Super-Resolution Generative Adversarial Network(ESRGAN)has been proposed to address uneven illumination and blurring effects in real inspection images.Compared to traditional illumination enhancement methods,this integrated approach achieves a higher-quality brightness balancing effect while preserving the inherent structural properties.It eliminates blocky shadow regions and blurring effects,transforming unevenly illuminated,low-resolution original images into uniformly illuminated,high-resolution,high-quality images.The fusion method also enhances the contrast between cracks and their surrounding background,effectively improving the continuity of crack boundary features.(3)A three-stage coupled weakly supervised deep learning algorithm is proposed to address the challenge of fine-grained crack identification across complex scenarios.Through the coupled action and mutual supervision of the three stages: crack region extraction,coarse segmentation of crack pixel,and fine segmentation of crack boundary,the algorithm progressively processes,transfers,and refines crack features in inspection images until fine-grained crack boundaries are achieved.For crack region extraction,a novel YOLO-C network is designed by incorporating a coordinate attention module(CA),a Swin Transformer module,and an upsampling layer for low-level features,among other measures.This network extracts crack regions from the images to be inspected,achieving optimal average precision(m AP)and overcoming over-segmentation caused by complex structural backgrounds.For coarse segmentation of crack pixel,weak labels generated by the YOLO-C network are used to supervise the training of a Cycle GAN network,transferring the output style of the YOLO-C network to crack texture segmentation results consistent with the annotation style of a generic CFD dataset.This approach eliminates the "all-black image" generated by pixellevel supervised algorithms,requiring only 0.5% of the annotation time needed by the latter.For fine segmentation of crack boundary,the output from the style transfer is used to supervise the training of a boundaryenhanced Deep Labv3+ network,refining the rough crack texture features into fine-grained crack boundaries without the need for low-consistency manual pixel-wise annotations.For the first time,an improved crossentropy loss function based on a boundary weighting mechanism is proposed,making the Deep Labv3+ network more focused on crack boundary identification.(4)An improved capacitance model algorithm is proposed to calculate continuous widths pixel-by-pixel from crack boundary identification results obtained through deep learning algorithms.Compared to the original capacitance model and orthogonal projection algorithm,the improved algorithm is more adaptable to complex crack topology boundaries,allowing for the precise and continuous extraction of crack skeletons and the determination of crack width calculation direction and range.This approach resolves the issue of significant ambiguity in the maximum crack width.The standard deviation of the crack width calculation results is minimized,enhancing the measurement accuracy of average crack width.Based on digital twin modeling and weakly supervised deep learning,the solution proposed in this thesis overcomes over-segmentation caused by complex background interference,enhances the quality of inspection images,and completely eliminates the occurrence of "all-black image" recognition results.It achieves fine-grained crack recognition and geometric feature measurement while significantly reducing the time cost of data collection and manual annotation.The research results have been applied to the Beijing-Shanghai High-Speed Railway,promoting the transition of deep learning methods from ideal experimental scenarios to practical applications in complex conditions.