Skip to Main content Skip to Navigation
Theses

Approches de classifications à partir de données fortement censurées pour l'analyse de fiabilité et la définition de stratégies de maintenance : application aux marquages routiers dans un contexte de véhicules autonomes

Abstract : The quality and reliability of road infrastructure and its equipment play a major role in road safety. This is especially true if we are interested in autonomous car traffic. Recent papers from VEDECOM Institut proves that a clear and reliable road marking is important in it decison making. Marking lanes are detected by camera. These markings need an accurate maintenance strategy to guarantee that the markings remain perceptible. This report proposes different solutions based on the reliabilty and maintenance theory. Today, the markings reliability is based on the retroreflective illuminance. A retroreflective marking reflects light from a vehicle headlight back in the direction of the driver. Marking retroreflectivity can be dynamically inspected using a retroreflectometer. The litterature of the last thirty years proposes degradation models for retroreflective marking based on a regression model. All of them have a common weakness: they are difficult to apply directly to a given road network. This report presents maintenance models who math with current maintenance actions. A marking lane is interpreted as multi-unit systeme. All unit are laid in parallel. The global maintenance strategy is based on four points. First, the whole inspection data is formalized into one monitoring base. If inspection data is missing or if the maintenance historic is unavailable else an estimation process based on the Agglomerative Hierarchical Clustering (AHC) is proposed. Second, to replace a whole markings lane is logistically difficult to work. Again, an AHC of the monitoring proposed several clusters. Each cluster presents it own degradation model. Clusters are geographically tracked and correlated to specific situation (interchange, urban area, bypass...). That's why a cluster is interpreted as a maintenance strategic area. Thirdly, a Weibull analysis of each cluster is done. Current retroreflectometers cannot detects the exact faillure moment. this information is statistically censored. Three cases are identified : left, right and interval censored. To parameter a Weibull model, an EM Algorithm is propoased as an alternative to the Maximum Likelihood Estimator. This algorithm is also an estimator to censored markings life time. Lastly, two classic preventive maintenance strategies are proposed : systematic according to the age and conditionned to the current degradation. Each one is credible according the current maintenance practice. The first prposed a passsive managament of the markings maintenance. The second ensures an advanced knowledge of the road network over the time. On a multi-unit system no-repairable and strongly censored, units which admit the same degradation model are identified by a clustering approach. Each cluster present it own Weibull analysis. Finally, an adapted maintenance strategy is done
Complete list of metadatas

https://tel.archives-ouvertes.fr/tel-02085842
Contributor : Abes Star :  Contact
Submitted on : Sunday, March 31, 2019 - 9:18:11 PM
Last modification on : Friday, July 17, 2020 - 5:08:34 PM
Long-term archiving on: : Monday, July 1, 2019 - 1:00:54 PM

File

TH2018PESC1118.pdf
Version validated by the jury (STAR)

Identifiers

  • HAL Id : tel-02085842, version 1

Collections

Citation

Maxime Redondin. Approches de classifications à partir de données fortement censurées pour l'analyse de fiabilité et la définition de stratégies de maintenance : application aux marquages routiers dans un contexte de véhicules autonomes. Infrastructures de transport. Université Paris-Est, 2018. Français. ⟨NNT : 2018PESC1118⟩. ⟨tel-02085842⟩

Share

Metrics

Record views

416

Files downloads

620