使用 AI 為 AIXM 5.1.1 采集數據.pdf

《使用 AI 為 AIXM 5.1.1 采集數據.pdf》由會員分享，可在線閱讀，更多相關《使用 AI 為 AIXM 5.1.1 采集數據.pdf（11頁珍藏版）》請在三個皮匠報告上搜索。

1、Data acquisition using AI for AIXM 5.1.1April 23,2025Maxime Jumelle-Blent.aiJean-Michel Varon-DSNAFLY AI Forum1-Context of the study1.1-FROM AIS to AIMAeronautical Information Service(AIS)was historically based on provision of HTML or PDF documents(AIP and Charts)and conventionalNOTAM.Documents esta

2、blished with help of database.Complete enough to generate docs but not enough to provide added value services based ondata.Provision of added value services to stakeholders requires to enlarge the scope and volume of data.CP1 regulation gives a regulatory frame andEuropean AIS Database a recipient1-

3、Context of the study1.2-FRENCH AIS DATA GAPPresent French database exported to EAD embedded in AIXM(Aeronautical Information Exchange Model)4.5 contains a volume of approximately 22 000 data.Goal of data completeness to comply CP1 requirements would raise this volume to 1 000 000 data(multiplied by

4、50).Time and HR to reach this goal by the end of 2025 enormous.May AI help us?2-Our methodology for data acquisition2.1-Overview of our methodologyIn order to best support the SIA in this project,the solution we are developing aims to build a set of software scripts that can recognize theinformation

5、 contained in different data sources and encode it in the AIXM 5.1 format.Set of AI modelsWeb pages(eAIP)PDF filesWord filesStructured data(XML,CSV)Data sourcesAIXM 5.1.1 databaseSchematic operation of our solutionThe main benefit of this methodology is its reusability,allowing encoding to be trigge

6、red again when the raw data is updated.2-Our methodology for data acquisition2.2-Identification of information through Artificial IntelligenceOne of the most important aspects of implementing the functional solution is the recognition and identification of information contained inunstructured data(p

7、articularly text)through AI.This implementation will enable the recognition of information for which business rules may be difficult or even impossible to define.In order to represent all AIXM objects,the set of scripts will include two engines for recognizing and identifying information:expert syst

8、emrules and AI models(LLMs).Both engines produce logs and deliver metrics,providing explainability for the results obtained.Set of AI modelsData sourcesExpert system rulesRecognizable information following a pattern,or sufficiently explicit.Modles dIA(LLM)Information that cannot be detected by a pat

9、tern(contained in text)or unpredictable phrasing.Representation of AIXM objectsAIXM objects are represented as software classes,allowing the creation of relationships.3-Focusing on airspace activations3.1-Focusing on airspace activationsWhen certain information cannot be identified by expert system

10、rules,we will use an AI model,specifically a large language model(LLM),inorder to easily extract information contained in unstructured data sources.UTC WORK_DAY 08:00 16:00 YES UTC SAT 08:00 13:00 YES UTC SUN 08:00 13:00 YES UTC HOL YES MON-FRI 0800-1600,SAT-SUN 0800-1300(SUM:-1HR).Except on holiday

11、s.The language model recognizes,based on numerous examples provided to itbeforehand,the different Timesheets mentioned in the raw text,extractingthe associated attributes(day,activation hours,events,etc.)as structureddata.A script is then used to convert this structured data into XML format,adhering

12、to the schema imposed by AIXM 5.1.1(example on the right).The language model used allows for extracting the different availabilityhours(activation)of an airspace,as defined from raw text(below)intostructured data.Example of using a language model for identifying service hours3-Focusing on airspace a

13、ctivations3.2-Realtime encoding of airspace activations3-Focusing on airspace activations3.3-Extracting annotations with HOR encoded separately from remarks DESCRIPTION SR-1000(sauf DIM et JF).DESCRIPTION SR-1000(except SUN and public HOL).REMARK RIMAP-P:TEL(00 689)40 46 31 78.IFR/VFR:lattention des

14、 navigateurs ariens est attire sur le caractre particulirement dangereux pour la vie humaine des activits sy droulant.REMARK RIMAP-P:TEL(00 689)40 46 31 78.IFR/VFR:pilots attention is drawn to particularly dangerous activity in this area for human life.HOR are encoded in a separate column from the r

15、emarks.WecanthenaddHORcontentasannotationtotheAirspaceActivation object,while the airspaces remarks areappended to the Airspace annotations.3-Focusing on airspace activations3.4-Running at scaleIn order to maximize the computation speed for extracting the activation times of airspace sectors,we have

16、 built an LLM inference architecturebased on vLLM to support as many concurrent queries as possible.Inference ServerCPU:24 vCPURAM:256 GBGPU:NVIDIA H100 SXM 80 GBLoad Balancer(Round-Robin)Inference ServerCPU:24 vCPURAM:256 GBGPU:NVIDIA H100 SXM 80 GBLocal ClientHTTP/2 StreamingUsing this architectur

17、e,we are able to deliver a throughput of 70 tokens per second,with a concurrency of 500 requests.4.1-Extending on others use cases4-Future worksOur work on encoding the activation schedules of airspaces allows us to apply this same approach to other use cases.Operational hours of air traffic control(ATS)services and ground services.Activation hours of routes and communication means.Activation hours of navigation aids(lighting,navaids).Thank you for your attention!April 23,2025