Intelligent recognition of defects in high‐speed railway slab track with limited dataset

Computer-Aided Civil and Infrastructure Engineering, Journal Year: 2023, Volume and Issue: 39(6), P. 911 - 928

Published: Oct. 12, 2023

Abstract During the regular service life of high‐speed railway (HSR), there might be serious defects in concrete slabs infrastructure systems, which may further significantly affect public transportation safety. To address these issues and fulfill functions HSR, traditional methods for engineers involve carrying out on‐site inspections manually or by semi‐automatic inspection vehicles, conducting timely corresponding repairing approaches maintenance, where are time‐consuming dangerous. In recent years, machine learning have been widely applied to intelligent automatic detection severe HSR. Currently, one most problems is lack sufficient high‐quality data model training, resulting low recognition accuracy HSR defects. solve this problem, paper proposed an based on a few‐shot model, that is, artificial intelligence limited size, recognizes three conditions HSR: cracks, track board gaps, unbroken state. Lightweight models specifically designed were proposed. Experiments conducted compare performances different lightweight‐designed models, including accuracy, parameter quantity, testing time. Results showed optimum can fast satisfactorily recognize with very size 10 samples each training category, satisfactory 73.9% test dataset 20 amounts 2.8 million, time 2.2 s per image. This study provides reference insufficient samples.

Language: Английский

---

Fine‐grained crack segmentation for high‐resolution images via a multiscale cascaded network

Computer-Aided Civil and Infrastructure Engineering, Journal Year: 2023, Volume and Issue: 39(4), P. 575 - 594

Published: Oct. 17, 2023

Abstract High‐resolution (HR) crack images offer more detailed information for assessing structural conditions compared to low‐resolution (LR) images. This wealth of detail proves indispensable in bolstering the safety unmanned aerial vehicle (UAV)‐based inspection procedures and elevating precision small segmentation. Nonetheless, achieving a balance between segmentation accuracy GPU memory consumption poses substantial challenge deep learning models when processing HR To overcome this challenge, novel “HR framework” (HRCSF) is proposed, specifically designed meticulously segment with resolutions exceeding 4K. First, multiscale feature extraction network (MsCFEN) was proposed embedment strip pooling operation enhance representation transverse longitudinal pixels from complex backgrounds. Subsequently, two cascaded operations were tailored MsCFEN, enabling comprehensive refinement process that incorporates both global local aspects. Furthermore, fully leverage potential each component process, complete architecture trained using loss function embedded boundary optimization. Conclusively, UAV‐based case study conducted on real bridge Changsha, demonstrating HRCSF's practicability segmenting The implementation HRCSF allows UAV perform effectively distance 3 m away girder, resulting significant 50% reduction time LR methods while maintaining high detection accuracy.

Language: Английский

---

Self‐training with Bayesian neural networks and spatial priors for unsupervised domain adaptation in crack segmentation

Computer-Aided Civil and Infrastructure Engineering, Journal Year: 2024, Volume and Issue: 39(17), P. 2642 - 2661

Published: July 29, 2024

Abstract This study proposes a novel self‐training framework for unsupervised domain adaptation in the segmentation of concrete wall cracks using accumulated crack data. The proposed method incorporates Bayesian neural networks uncertainty estimation pseudo‐labels, and spatial priors screening noisy labels. Experiments demonstrate that approach achieves significant improvements F1 score. Comparing scores, DeepLabv3+ U‐Net showed performance 0.0588 0.1501, respectively, after adaptation. Furthermore, integration Stable Diffusion few‐shot image generation enhances by 0.0332. enables high‐precision with as few 100 target images, which can be easily obtained at site, reducing cost model deployment infrastructure maintenance. also investigates optimal number iterations based on score, providing insights practical implementation. contributes to development efficient automated structural health monitoring AI.

Language: Английский

---

Evidential transformer for buried object detection in ground penetrating radar signals and interval‐based bounding box

Computer-Aided Civil and Infrastructure Engineering, Journal Year: 2025, Volume and Issue: unknown

Published: Jan. 7, 2025

Abstract Three‐dimensional (3D) buried object detection using ground penetrating radar (GPR) benefits from the powerful capacity of image‐wise deep neural networks. However, it still faces challenge information loss raw GPR signals to two‐ and three‐dimensional images, such as frequency‐domain when normalizing into gray‐scale images spatial stacked B‐ C‐scan replace inputs. To solve challenge, this study has proposed an ENNreg‐transformer model, directly 3D perform detection. In are first converted sequential voxelization obtain spatiotemporal features. The features then aggregated by intuition‐guided feature aggregation layer simulate expert behavior analyze data. Finally, evidential header outputs interval‐based bounding boxes for experiment on two road datasets demonstrates that model exceeds other state‐of‐the‐art models tasks thanks aggregation. addition, box represents bounding‐box uncertainty, which derives inherent limitations

Language: Английский

---

Semi-supervised semantic image segmentation by deep diffusion models and generative adversarial networks

International Journal of Neural Systems, Journal Year: 2024, Volume and Issue: 34(11)

Published: July 5, 2024

Typically, deep learning models for image segmentation tasks are trained using large datasets of images annotated at the pixel level, which can be expensive and highly time-consuming. A way to reduce amount required training is adopt a semi-supervised approach. In this regard, generative models, concretely Generative Adversarial Networks (GANs), have been adapted tasks. This work proposes MaskGDM, architecture combining some ideas from EditGAN, GAN that jointly their segmentations, together with diffusion model. With careful integration, we find model improve EditGAN performance results in multiple datasets, both multi-class binary labels. According quantitative obtained, proposed improves when compared DatasetGAN respectively, by [Formula: see text] text]. Moreover, ISIC dataset, our proposal other up

Language: Английский

---

A generative adversarial network approach for removing motion blur in the automatic detection of pavement cracks

Computer-Aided Civil and Infrastructure Engineering, Journal Year: 2024, Volume and Issue: 39(22), P. 3412 - 3434

Published: May 20, 2024

Abstract Advancements in infrastructure management have significantly benefited from automatic pavement crack detection systems, relying on image processing enhanced by high‐resolution imaging and machine learning. However, motion blur substantially challenge the accuracy of analysis. Nevertheless, research mitigating remains sparse. This study introduces an effective system adept at deblurring segmentation, employing a generative adversarial network (GAN) with UNet as generator Wasserstein GAN Gradient Penalty (WGAN‐gp) loss function. approach performs exceptionally images improves segmentation accuracy. Models were trained sharp artificially blurred images, WGAN‐gp surpassing other functions effectiveness. innovatively suggests assessing quality through addition to peak signal‐to‐noise ratio (PSNR) structural similarity (SSIM), revealing that PSNR SSIM may not fully capture effectiveness for images. An extensive evaluation various generators, including UNet, lightweight TransUNet, DeblurGAN, DeblurGAN‐v2, MIMO‐UNet, identifies superior performance simulated blur. Validation actual motion‐blurred confirms proposed model. These findings demonstrate GAN‐based models great potential overcoming challenges marking notable advancement field.

Language: Английский

---

A response‐compatible ground motion generation method using physics‐guided neural networks

Computer-Aided Civil and Infrastructure Engineering, Journal Year: 2024, Volume and Issue: 39(15), P. 2350 - 2366

Published: March 31, 2024

Abstract Selecting or generating ground motions (GMs) that elicit seismic responses matching specific standards expected benchmarks for nonlinear time‐history analysis (NLTHA) is crucial ensuring the rationality of structural design and analysis. Typical GM inputs NLTHA, either natural artificial, are normally spectrum‐compatible, which may produce significant variations in results, even using multiple GMs. This paper introduces a response‐compatible motion generation (RCGMG) method GMs tailored to be response‐compatible. NLTHA results only few these artificial can closely approximate mean from large set spectrum‐compatible target responses. The RCGMG adopts response diagram time domain (RDTD) characterize nonstationary features their impacts on dynamic A physics‐guided conditional generative adversarial network developed RDTDs with These generated then mapped into through feedforward neural network. To verify effectiveness RCGMG, different structure models under various site conditions spectra conducted. Seismic RCGMG‐generated compared demonstrate closer responses, fewer robust generalization performance.

Language: Английский

---

Gyroscopic effects of the spinning rotor-blades assembly on dynamic response of offshore wind turbines

Journal of Wind Engineering and Industrial Aerodynamics, Journal Year: 2024, Volume and Issue: 247, P. 105698 - 105698

Published: March 18, 2024

An analytical solution for the gyroscopic effect of spinning rotor-blades assembly on dynamic response offshore wind turbines (OWT) is presented. A continuous coupled model rigorously developed to form partial differential equations fore-aft, side-side, and yaw motions. The moments caused by angular momentum are formulated handled into three boundary conditions at nacelle. procedure obtaining operational natural frequencies structure including these developed. Furthermore, a function each fore-aft motions obtained solving motion under wave load applied in only direction. Finally, calculated compared idling ones considered example OWT. different values assembly's velocity investigated. proposed this study unfolds revelational capturing turbine industries which also can be guideline developing floating OWT models.

Language: Английский

---

Deep learning‐based automatic classification of three‐level surface information in bridge inspection

Computer-Aided Civil and Infrastructure Engineering, Journal Year: 2023, Volume and Issue: 39(10), P. 1431 - 1451

Published: Nov. 6, 2023

Abstract Bridge inspection ensures that in‐service bridges are managed and maintained in conformity. To enhance the accuracy efficiency of bridge inspection, an automatic hierarchical model is proposed, which enables classification correlation surface images at three levels, namely, structure, component, defect type level. Thus, impact both types affected components on safety can be simultaneously considered. The proposed uses a group sub‐models instead common flat network to realize multiple tasks, advantageous accuracy, training simplicity, scalability. levels has reached 96%, 92%, 81%. Results demonstrate effectiveness method multi‐scale targets. This study may provide new strategy for developing systematic easily adaptable detection framework practical engineering.

Language: Английский

---

A novel classification method for GPR B-scan images based on weak-shot learning

Journal of Applied Geophysics, Journal Year: 2024, Volume and Issue: 221, P. 105287 - 105287

Published: Jan. 9, 2024

Language: Английский

---