交通基础设施中的大型模型：一个视角（英）Y-Du-2024

Large models in transportation infrastructure 3Table 2.Literature on large models in transportation infrastructure industryStageLiteraturePlanning and designChen et al.,2022;Shah et al.,2023;Roberts et al.,2023Project managementBashir et al.,2023:Prieto et al.,2023:Abbas et al..2023Operation and maintenanceRane,2023Safety controlChen et al.,2023;Chandra et al.,2023;Goecks and Waytowich,2023;Das et al.,2023Report documentEngineeringPlanning designDrawingDownloadedEnvironmentaldataBase large model合Project managementfrom https://acadPolicy document000百demic10000.oupIndustry standard14Operation maintenanceManagementDesign algorithmcom/iti/softw准reMaintenancearticlWorking criterionsystemTraining dataAlgorithm,software,ete.Safety controlle/doi/10Figure 2 The construction process of a large model for transportation infrastructurelarge models for project planning,operational management,expectation to create an AIGC (AI-Generated Content)platformdesign enhancement,and risk mitigation (Rane,2023)can fosterspecific to this domain,aiming to achieve artificial substitutioncollaboration,effective communication,and the integration ofin engineering processes.interdisciplinary expertise.Leveraging in-depth industry expertise and relying on AIGCtechnology,the planning and design of transportation infrastruc.3.1.4.Safe controlture is anticipated to achieve intelligent breakthroughs at theforefront.In particular,preliminary design,which relies heavilyIn the context of safe control for transportation infrastructure,on industry experience,requires the development of more sophis-research on the application of large models mainly involves train-ticated algorithmic models based on professional knowledge.Thising and fine-tuning the base models using accident images andprocess entails a longer development cycle and is challenging3textual reports.For instance,by integrating GPT-3 (Brown et al.,2020)with stable diffusion (Rombach et al.,2022),and incor-to realize in the short term.Conversely,construction drawingdesign is relatively procedural,but it encompasses numerousporating relevant knowledge about foreign object intrusions onspecialized sub-fields that need to be tackled one by one.In theuaryrailways,high-quality and highly realistic samples of railwayintrusion images can be generated,achieving the goal of pre-long term,as model algorithms continue to mature,large modelsare expected to become involved in various stages,indudingventing railway disasters (Chen et al.,2023).Alternatively,effec-conceptualization,preliminary design,three-dimensional modeltive action plans can be rapidly generated by induding disasterresponse and planning guidelines in the initial prompts(Goecksoutput,and drawing review.This integration aims to establishand Waytowich,2023).Similarly,large models can be utilized toa complete closed-loop system from 'demand input'to 'AIdesign','demand feedback',and 'design updates',therebyanalyze aviation safety (Chandra et al.,2023)by mining accidentsignifican